Design And Implementation Of Visual Inspection System For Electric Power Based On The Ascend 310 Processor

The safety production of electric power industry is related to social stability and economic development,so electric power inspection is of great significance to ensure the safe production of electric power business.The shortcommings of the traditional electric power inspection method are large work load,low efficiency and inaccuracy,which is no longer suitable for the current inspection requirements.Deep learning algorithms can deeply mine data features and perform well in the field of target detection,meanwhile,the rapid development of AI processor technology improves the arithmetic power of hardware,so the combination of AI processor technology and Deep learning algorithms provides new ideas for power inspection.We deploy deep learning algorithms in the embedded platform,and achieve the visual power inspection system based on Ascend 310 AI processor,which can detect and analyze common power targets.To meet the requirements of the application scenario of power inspection,we select the Ascend 310 AI processor to build the hardware environment of the power visual inspection system from multiple perspectives such as performance,arithmetic power and accuracy,and complete the design and achievement of the system software on this basis.The system is developed based on Linux operating system and mainly contains streaming media processing,algorithm analysis,system management and other functions.In order to achieve the streaming media processing function,we complete the system middleware design at the software level,by using the hardware driving capability of the Ascend 310 processor.To achieve detection and analysis of common power targets,we use the neural network acceleration core of Ascend 310 to deploy acceleration for deep learning algorithms.The results show that the system can complete common power inspection tasks and meet the real-time and stability requirements of the inspection system,with great practical applications.For the deployment of deep learning algorithm under embedded platform,we take YOLO v4 algorithm as an example,deploy the deep learning algorithm on the Ascend 310 processor,and carry out research on the optimization of algorithm deployment based on the characteristics of processor architecture.According to the arithmetic constraints and data constraints of the platform,the algorithm deployment task is completed in terms of both model conversion and custom arithmetic.Based on the Linux operating system,we use multi-core and multi-threading technologies to improve the system performance.At the same time,we improve the accuracy and performance of the embedded algorithm deployment by using the architectural characteristics of the Ascend 310 processor,such as access memory reading/writing optimization and arithmetic optimization.The results show that these optimization methods can further improve the accuracy of the algorithm and the overall performance of the system.