Instrument Recognition Method Of Substation Based On Deep Learning And Fusion Attention Mechanism

At present,intelligent robots are widely used in substation inspection work,and the recognition of the number of pointer instrument representations is an important task of robots.Traditional image recognition technology has the problem of low detection accuracy and large reading error in the identification of pointer instrument representations.In recent years,deep learning algorithms and attention mechanisms have made important breakthroughs in the field of image recognition.Therefore,according to the characteristics of pointer instruments,this paper integrates attention mechanisms on the basis of deep learning,and focuses on the identification method of substation pointer instrument representations.The main contents and progress are as follows:(1)The CenterNet object detection algorithm based on attention mechanism is used to complete the instrument dial positioning.This method first replaces ResNetl8 in the original CenterNet model with lightweight feature extraction network MobileNetV3,and on this basis,replaces the SE attention mechanism with a lightweight ECA attention mechanism without dimension reduction;then draws on the FPN feature fusion idea to upsample the highdimensional feature map and fuse it with the low-dimensional feature map of the same size;and finally introduces the CA spatial attention mechanism in the detection head module.Have the model also focus on learning spatial features when learning classified features.(2)The DeepLabV3+semantic segmentation algorithm based on the attention mechanism is used to separate the pointer from the background.First of all,in the ASPP module of DeepLabV3+,the DAMM dual attention mechanism is connected in parallel,and the DAMM can not only strengthen the learning of the important spatial characteristics of the pointer area,but also strengthen the learning of the important channel characteristics of the pointer,improve the classification accuracy of the pixels of the pointer area,and achieve the effect of smoothing the edges of the pointer;then replace the cross-entropy loss function with the Focal Loss loss function to solve the problem of imbalance between the classification samples with more pixels on the dial and the less pixels in the pointer.Guide the model through mining learning of fine pointer difficult samples by setting different hyperparameters.(3)In the identification of the instrument representation number,first correct the elliptical distortion instrument dial according to the scale disc segmentation result,and obtain the upright circular instrument dial;then obtain the tilt angle of the pointer according to the rotation angle of the smallest outer rectangle of the pointer outline;and finally calculate the instrument representation value according to the category,range and angle method of the instrument.According to the above method,a large number of simulations and field tests were done,and the experimental results showed that in the instrument detection stage,the mAP of the proposed model was increased by 7.51%,the FPS was increased by 15f/s,and the model size was reduced by 56MB.In the identification stage of instrument representation,the maximum error between the predicted value of the instrument representation and the true value of the instrument before correction is 6.7%,and the average error is 4.2%,and the maximum error between the predicted value of the instrument representation and the true value of the instrument is 2.2%and the average error is 1.3%,thus verifying the effectiveness of the proposed method.