Research On Automatic Recognition And Reading Of Pointer Instrument Nased On Deep Learning

With the adjustment of energy consumption structure and the proposal of "5G+smart grid" and other related concepts,electric energy,as the most widely used clean energy at present,is the basis of people’s daily life,and also holds the lifeline of the national economy.As the hub of energy production and transmission,it is very important to build a fully intelligent substation.Pointer instruments in substations often rely on manual reading,which is easy to cause problems such as false detection and missed detection,and labor consumption.In view of the above problems,this paper uses deep learning and machine vision technology to realize an automatic identification and reading method of pointer instruments based on patrol robots,the main contents of which are as follows:Aiming at the problem that the traditional machine vision algorithm produces distortion and blurring of the instrument during the shooting process of the inspection robot,resulting in inaccurate spatial position,complex automatic correction,and poor effect,a R2D2 image automatic correction method based on the two-channel residual shrinkage network attention mechanism is proposed.CBAM residual shrinkage network is introduced into the backbone network of R2D2 feature point matching algorithm,and the salient feature descriptor of the image to be registered is judged twice to determine whether the key point has strong repeatability and substitutability,and unreliable points are eliminated,so as to find secondary advantages and obtain more accurate matching results.In the instrument registration experiment,it can be seen that the algorithm in this paper performs well.In the image to be registered that lacks nearly 80% features,obvious local features can still be found,and the correction is successful.An improved Blend Mask model is proposed to solve the contour segmentation problem between the dial and the pointer caused by uneven illumination and blurred instrument panel.This method uses the idea of closed-loop control to introduce "feedback link",replaces the original feature pyramid model with iterative feature pyramid RFPN,analyzes the performance of common backbone network and RFPN,and finally integrates with Res Net101 backbone network.A special loss function is proposed for the improved algorithm.In comparison experiments,the improved method in this paper has the smallest loss.In the case segmentation experiment of the instrument with overexposure,underexposure and compound noise,the segmentation accuracy of the improved algorithm in this paper has reached more than 92%,all the evaluation indicators have been improved,the data has reached the ideal requirements,the edge of the dial and pointer segmentation graphics is smoother,and the segmentation area is accurate.Aiming at the problem that the traditional reading method is computationally heavy and the straight line fitting based on statistics and graphics is inaccurate,a pointer direction discrimination method of density vector is proposed,which simplifies the reading model of angle method,and finally constructs the instrument detection and interpretation system.Based on the patrol robot as the hardware,the pointer instrument reading experiment was carried out,and a set of real-time information display software was developed.Due to the small range of the instrument,the allowable error range is 0.001.The experimental results show that the reading error of the automatic reading system of the pointer instrument based on deep learning in this paper is within the allowable range.Compared with the manual reading,it will be more accurate,with strong adaptability,in line with the patrol inspection standard under the actual working conditions,and meet the requirements of stability and reliability.