Electric energy metering box, a thing that sounds strange, but you must feel familiar when you see it.
As a key component of the power system, the electric energy metering box is widely installed in various production and living areas to protect the safety of metering devices such as electric energy meters and transformers inside it, and is an important facility to ensure the stable operation of the electric power system. As the power industry actively promotes digital and intelligent upgrades, the digital filing of electric energy metering boxes is also in full swing in the power industry. A comprehensive and accurate metering box operation file can reflect information such as the electricity consumption in the area, equipment failure and damage, and the geographical location of the equipment.
Energy Metering Box
Hezhong Weiqi, a company that focuses on providing comprehensive solutions for informatization and digital construction for electric energy and other industries, uses the ultra-lightweight real-time target detection model PP-PicoDet in the paddle detection end-to-end development kit PaddleDetection to develop electric energy The digital app for metering boxes improves the quality and efficiency of data entry by on-site operators, helps State Grid reconstruct the files of electric energy metering boxes, and provides strong support for digital management of power grids and asset security.
It is time-consuming and labor-intensive to enter the file information, and the accuracy of the file is facing challenges
The filing of the metering box involves very complicated information collection. Taking the filing requirements of a provincial power grid metering box as an example, inspectors need to enter the appearance of the metering box and the defect information of the electric energy meter inside the box on site, including whether the door of the metering box, the viewing window, the lock, the switch cover, and the box are Whether there is damage or missing, whether the electric energy meter switch cover, seal, screen display, etc. are damaged or missing, this undoubtedly poses a great challenge to the meticulousness of the information entry personnel.
At the same time, metering box equipment is often installed in a complex environment, with various specifications, and a large number. Whether it is a residential area, an industrial site, or an outdoor field, etc., inspection personnel are required to check a large number of metering boxes at different heights and collect information one by one, resulting in manual mining. The entry is difficult, omissions and mistakes occur from time to time, and the entry efficiency and accuracy are low, which seriously affects the quality of digital file building of the metering box.
Electric energy metering box has various specifications and complex installation environment
Based on the keen insight into the problem of manual entry, Hezhong Weiqi decided to apply artificial intelligence technology to the digital filing of metering boxes to solve the problem of actual scenarios based on years of in-depth experience in the power industry.
Deploy the model on the mobile phone, respond quickly even without a network, and the efficiency of AI assistance is increased by 20%
Replacing the traditional human eye judgment with artificial intelligence vision technology is the direction of Hezhong Weiqi's technology implementation. By taking photos of the appearance and interior of the metering box, with the help of the target detection model, the key information in the photos can be quickly detected, so as to achieve the goal of information entry.
However, in the actual implementation of the "last mile", the team faced a lot of challenges.
"The photos must be taken by the on-site workers using mobile phones. At the same time, because the file entry has immediate requirements, it cannot be uploaded to a certain place for unified identification after the shooting is completed, so the speed of the model must pass the test." Hezhong Weiqi Data Service Department Hao Zengcai, the person in charge of the technology, said, "However, the on-site operation environment is complicated, and the network conditions near the metering box are not guaranteed in many cases, such as underground garages, field places, etc., which make it impossible to call the server-side model. Therefore, after the model development is completed, we will Model capabilities must be migrated to mobile terminals at the same time, that is, deployed on mobile phones, to ensure that it can be used in a non-network environment, and it is necessary to ensure that the accuracy and speed of the model meet business needs. ”
After investigation and comparison, the team finally chose to use the ultra-lightweight real-time target detection model PP-PicoDet in the paddle detection end-to-end development kit PaddleDetection to realize model development and mobile phone deployment. PP-PicoDet is a lightweight detection model customized for mobile devices/low computing power devices. It provides xs, s, m, l and other series of models to meet more refined accuracy and speed selection requirements. At the same time, the model maintains a high ease of use in terms of deployment, supports a rich inference library, and provides an out-of-the-box quantization and compression scheme.
Overall technical solution design
Specifically, the technical steps of Hezhong Weiqi Model R&D are as follows:
- Data collection: Based on the inspection work in the last year, the image data collection of the metering box is carried out through mobile phones or handheld devices;
- Data preprocessing: Clean and label the collected data according to the labeling specification, and use image enhancement, cropping, zooming, rotation and other operations to expand and enhance the data, so as to eliminate the influence of factors such as illumination and angle, and improve the applicability of the recognition model;
- Model training: use the ultra-lightweight real-time target detection model PP-PicoDet in PaddleDetection to train the model, and finally generate the metering box material and defect recognition model, the electric energy meter recognition model, and evaluate the model performance;
- Model verification and optimization: evaluate the performance of the model through on-site trial operation, iteratively optimize the model, and improve the recognition accuracy and generalization ability;
- Model integration: Integrate the optimized model into the server and the mobile terminal. The mobile terminal model will reduce the model size and improve the inference speed through pruning, compression, quantization, distillation and other technologies to meet the application requirements of the mobile terminal.
In the end, the accuracy rate of the metering box material identification model in the electric energy metering box digital app reached 99.2%, the accuracy rate of the metering box defect identification model reached 94%, and the accuracy rate of the electric energy meter identification model reached 99.6% . Liaoning, Jilin, Heilongjiang, Henan and other provincial power grids have been promoted and used. One province has completed the collection of tens of millions of pictures of more than one million metering boxes, and the inspection efficiency has increased by 20%, which greatly reduces manpower investment and improves digital files. information accuracy.
Inspectors use mobile phones to take photos and enter files in underground places
"The implementation of the project is inseparable from the strong support of the leaders of the company's technology center, and Feidao's open source ecological community has also provided a lot of help." Hao Zengcai said, "During the development process, we have learned a lot from the community. A case of model deployment on the mobile side. Flying Paddle also organized a special guidance support for us by PPDE (Flying Paddle Developer Technical Expert), which solved some difficulties we encountered during the project. Me and my other My colleagues have finally obtained PPDE certification, and I hope to promote technological innovation with more outstanding developers and contribute to the industry."
Hezhong Weiqi has officially joined the Flying Paddle Technology Partner Program, and will continue to explore the infinite possibilities of AI+power and solve the pain points of the power industry. For example, apply target detection technology to more power operation scenarios, such as line inspection, electric personnel safety operation inspection, etc., to improve inspection efficiency and personnel operation safety; or build an AI decision-making model to monitor the power generation and load of the grid Comprehensive judgment and forecasting assist in the realization of optimal energy allocation; another example is the combination of the Wenxin large-scale model to develop the internal knowledge base of electric power companies, realize document analysis and intelligent question and answer of internal data in the enterprise, and continue to use AI to empower power grids to improve quality and efficiency. The large model of Wenxin and Flying Paddle will also continue to work with partners from thousands of industries to help the implementation of artificial intelligence in the industry and promote the intelligent upgrading of the industry.
Baidu's AI technology ecological support plan to open up the last mile of enterprise applications
Baidu's AI technology ecosystem gathers Paddle Developer Technical Experts (PPDE) with rich experience in industrial implementation to provide enterprise users with special technical support for application implementation based on Wenxin large model and Paddle deep learning platform.
Provide 4 times of technical guidance for the key points of enterprise projects to help enterprises overcome technical difficulties, reduce enterprise development costs, and achieve rapid project implementation.
expert guidance
Paddle Developer Technical Experts (PPDE) provide one-on-one support for enterprise project implementation.
Free for a limited time
Users who apply in August 2023 and pass the review will receive free guidance rights for a limited time.
For the Baidu AI technology ecological support plan, please add "Flying Paddle Assistant" (WeChat ID: baidupaddle) on WeChat to apply.