Research On Operation State Recognition Of Distribution Equipment Based On Machine Vision

The use of robots to replace humans in the tedious and harsh environment of daily inspection and maintenance tasks in power distribution rooms can significantly improve inspection efficiency,safety,and stability.This thesis researched the recognition of the operating status of power distribution equipment.Given the multiple types of different instruments,multiple color indicators,differences in shooting spacing and viewing angle,instrument size and tilt angle,and illumination changes at different times and places in the images of the power distribution cabinet panel collected by inspection robots,a combination of traditional machine vision and deep learning was adopted.The running state of three kinds of power instruments,such as a pointer instrument,digital instrument,and indicator light installed on the panel of the distribution cabinet,is automatically recognized and interpreted.The main research elements are as follows:This thesis uses an improved YOLOv5s-based model for the identification and location of power meters in distribution cabinets.The YOLOv5 s network structure was improved by replacing the YOLOv5 s backbone network CSPDark Net with a lightweight network Mobile Net V3 with an improved ECA attention mechanism.Model training and ablation experiments were also conducted on a self-constructed dataset.The results show that the improved model has a size of only 9.5M,a volume compression of 34%,an average detection accuracy mean m AP of 97.0%,a detection speed of 28.3FPS on the server(RTX2080Ti),with an FPS increase of 9.2 compared with that before the improvement,and a detection speed The performance is excellent and confirms the effectiveness of the model improvement method.This thesis designs an algorithm for reading the displayed value of pointer-type meters.Firstly,an improved PSPNet semantic segmentation algorithm is used to implement the segmentation of the pointer and scale area.The main improvements to the PSPNet model in this thesis include two aspects,one is to construct a lightweight model using the deeply separable convolutional neural network Mobile Net V2,and the other is to change the activation function Re Lu to Swish.Secondly,the angle of deflection of the pointer and the center of rotation of the pointer are obtained in turn using the Hough transform linear detection,least squares method.Then,the contour finding method and the Google OCR tool Pytesseract were used to achieve the character positioning and character recognition of the meter.Finally,the angle of pointer deflection,the upper and lower adjacent main scale values and the angle of deflection between the upper and lower adjacent main scale and the centre of rotation are substituted into the angle formula to obtain the pointer meter reading.The experiments show that the segmentation accuracy of the lightened PSPNet-M2 S model is 97.07%,and the segmentation performance is good.For different types,different ranges of pointer meter representation value interpretation have high accuracy,in the error allowed range of pointer meter reading error of up to 3.27%,to meet the needs of practical applications.The method of reading the displayed value of the digital display meter is studied.The experimental comparison of the traditional algorithm and the single-stage PGNet algorithm for character recognition is carried out for the digital display meter.The experimental results show that a better PGNet algorithm model is obtained by adjusting the training optimization strategy,and the accuracy of reading recognition under the normal display of the digital display meter is as high as 99.1%.The state recognition of indicator lights is realized.Firstly,a circle detection method based on the Hough gradient is used to segment the circular indicators;then,the state of each type of indicator is discriminated according to the information of each HSV channel.The experimental results show that the algorithm in this thesis can identify the open and closed status of different colored indicators very well,and the recognition accuracy is 99.5% at the highest.