Research On Anti-collision Early Warning Of Offshore Wind Power Facilities Based On Machine Vision

In the context of clean and low-carbon energy,China’s offshore wind power development and construction has entered the fast lane of development,large-scale offshore wind power projects successfully completed and placed into operation.Offshore wind power continues to develop to the deep sea at the same time,but also to the offshore wind farms and the surrounding waters bring greater risk of safety.Ships as the main body of marine transportation,with the large-scale development of offshore wind turbine installed capacity,will certainly increase the possibility of passing ships colliding with offshore wind turbines,damage to submarine cables.Therefore,it is important to carry out research on the detection and early warning of offshore wind power facilities to reduce the occurrence of ship collision wind power accidents.Nowadays,offshore wind power monitoring mostly relies on manual fieldwork,which has problems such as low detection accuracy,time-consuming and laborious,and untimely detection.Therefore,for the demand of offshore wind power safety monitoring,this paper takes offshore wind turbines and ships as research objects,and establishes an improved anti-collision early warning system for offshore wind power facilities combined with YOLOv5 s based on machine vision fusion deep learning algorithm.The main research contents are summarized as follows:(1)An offshore wind power dataset is constructed.The collected wind power videos were intercepted one image every 25 frames,and the images containing wind turbines and ships were selected after filtering,and the classes and locations of the objects in the images were labeled.Finally,the intercepted wind power images were expanded by rotating,cropping and scaling the data with the Open CV tool.(2)A lightweight YOLOv5 s network model is proposed to address the problems of redundancy in the number of parameters and high complexity of existing maritime target detection models.The backbone network of the original YOLOv5 s is improved to Ghost Net,which can guarantee certain detection accuracy and detection speed while reducing the number of network parameters.The experimental results show that the improved YOLOv5s-Ghost Net model is half the size of the original network,and the detection speed is improved by 24 FPS compared with the original YOLOv5 s,which effectively reduces the complexity of the original network model and meets the real-time requirements of the detection task.(3)Since the sea objects are mostly in complex environments,the original network has low detection accuracy,missed detection and false detection in the detection process.To improve the detection performance,the SENet(Squeeze-and-Excitation Networks)attention mechanism is added to the backbone extraction end and feature fusion layer of YOLOv5 s,respectively,to deepen the focus on important features.In addition,the Neck layer of the original network is improved into a Bidirectional Feature Pyramid Network(Bi FPN)structure to fuse features of different scales in order to improve the small target detection accuracy.The experimental results show that the introduction of both SENet and Bi FPN modules improves the average accuracy by 2.3% compared to the original YOLOv5 s,and the detection accuracy of small targets of fishing vessels improves by 1.1%.(4)An anti-collision early warning software system for offshore wind power facilities was constructed.According to the actual needs of offshore wind power safety monitoring,the overall architecture and functional model design of the software system were first carried out,and the system was divided into two functional modules: object detection and distance warning,the former being responsible for detecting the wind turbines and ships in the video based on the trained model file,and the latter based on the two-dimensional camera retaining perspective transformation method to determine whether the passing ships in the video are close to the wind farm waters.Then the development is based on Python’s graphical user interface Py Qt and module interface functions.Finally,the model files and program instructions are integrated into the corresponding modules,and the final detection results can be visualized in the software system interface.