Self-adjusting Residual Network Recognition Method For Substation Equipment Status Based On Inspection Image

The safe operation of substation equipment is the key to ensure the reliable power supply of power grid and the guarantee for the steady development of national economy.In order to meet the requirements of safe operation of modern substations,inspection robots are gradually applied in the construction of intelligent substations.However,the image recognition technology currently used in robot inspections basically relies on traditional image segmentation and feature extraction methods,it’s difficult to accurately identify the operating status of some devices with similar appearances and low discrimination in various defect status.Therefore,on the basis of constructing a multi-class data set of substation inspection images,this paper analyzed the imaging characteristics of equipment’s operating status,used the Faster Rcnn algorithm to implement the preprocessing of target region extraction,and combined three optimization measures,such as convolution kernel decomposition,multi-scale feature fusion,and hyperparameter self-adjustment of improved Bayesian optimization,etc.Then a self-adjusting residual network state recognition method for substation inspection images is proposed.The main works of the study include:(1)The types of operation status and the imaging characteristics reflected in the substation inspection images were analyzed,which were used to construct a multi classification image data set.And a target region extraction model based on Faster Rcnn was established to realize the preprocessing of substation inspection images.On this basis,a device state recognition model based on CNN was constructed to verify the applicability of the pretreatment technology.The example analysis has shown that the classic networks in the field of deep learning such as Alexnet and Resnet can identify the device status in the inspection image and the preprocessing technology can effectively improve the recognition accuracy.(2)Aiming to solve the problem that classic networks such as Alexnet and Resnet are not suitable for equipment status recognition of the substation inspection image,three optimization measures were proposed for improvement.Firstly,the residual network infrastructure is optimized by convolution kernel decomposition technology,which can reduce the number of model parameters while improving generalization performance.Then the multi-scale convolution feature fusion method is used to fuse the judgment features generated in the shallow and deep layers of the network,which can improve the recognition accuracy of defect states.Finally,an improved Bayesian optimization algorithm based on coupling constraints was proposed.Under the constraints of accuracy and network volume,self-tuning of hyperparameters such as the number of convolution kernels and network depth can be achieved to obtain a lightweight diagnostic with optimal performance model.(3)Characteristic parameters such as state recognition accuracy,test speed,and model memory footprint were used to analyze the applicability of the self-adjusting residual network to three types of inspection images,including infrared images,visible images,and online monitoring vibration signal images.It is proved that the model has good applicability.At the same time,a visual analysis of the activation degree,network characteristics and classification results of the convolution kernel was performed,which explains why the self-adjusting residual network has high performance.