Research On Intelligent Recognition Of Common Defects Of Pins In Power Fitting Based On Deep Learning

In recent years,UAV intelligent inspection technology has been widely used in the power industry due to its excellent performance.Although the application of UAV inspection technology improves the convenience of inspection to a certain extent,a large amount of visible light image data is generated during the inspection process,and these image data still need to be analyzed manually.Manual labeling is time-consuming and labor-intensive.In view of this problem,this paper takes the common defects of split pins on power fittings as a research object,and proposes a smart recognition method for common defects of split pins based on deep learning algorithm Retina Net.Here,a deep convolutional neural network is used to automatically extract image features,and a feature pyramid with strong semantics on multiple scales is constructed through the feature pyramid network.Finally,the position and classification of the defects of the split pins is marked with the help of the classification subnetwork and the frame regression subnetwork.It provides theoretical support for the realization of future intelligent inspection technology.Considering that the key parameter learning rate and the optimizer have an important impact on the performance of the model.Therefore,the effects of different learning rates and optimizers on the recognition results are compared and analyzed.The experimental results show that choosing a suitable learning rate and optimizer can not only enhance the generalization ability of the model,but also accelerate the model's convergence rate.In view of the fact that the number of defective data in the split pin is far less than the number of normal samples in the actual situation,first of all,the impact of the lack of defective data samples on the recognition results is analyzed.Secondly,this paper proposes a training strategy based on the class-balanced sampling method,that is,by expanding a small number of data samples to ensure that each class of data has the same opportunity to be selected during the training process.Finally,considering the small amount of loose split pins data in the actual situation,a method for collecting auxiliary split pins sample data is proposed.With reference to the power fittings connection method,we construct a simplephysical model to obtain auxiliary data under the premise of ensuring the consistency of the main foreground object(power fittings).Then,the auxiliary data is added to reduce the negative impact of insufficient sample data on recognition results,and a quantitative analysis is performed on the auxiliary data samples.Based on the analysis results,we proposed a dynamic adjustment strategy for auxiliary data samples.The experimental results show that when there are insufficient samples of defective data,the trained model will pay more attention to the normal categories,resulting in a large number of defective data being misidentified as normal categories.The training strategy based on the class balance sampling method ensures that the opportunities for each class to participate in the training tend to be balanced,which helps the model to obtain better recognition results in the test set.Adding appropriate amount of auxiliary data can effectively alleviate the adverse effects caused by category imbalances,but when the auxiliary data exceeds a certain amount,the difference between the data collected at the inspection site and the auxiliary data will cause the generalization ability of the detection model to decrease.Aiming at the difference between the auxiliary data and the target data,the adjustment strategy of the auxiliary data proposed in this paper effectively reduce the adverse effects caused by this difference and further improve the recognition rate of the minority class.Considering that it is difficult to train deep convolutional neural networks with small-scale data,the Image Net dataset is used to pre-train the convolutional neural network in combination with transfer learning strategies,and then the network weights are fine-tuned with the help of small-scale target data.The experimental results show that the transfer learning method can promote the model to learn useful common features in different fields.In order to improve the recognition accuracy of fuzzy pictures,a generative adversarial network is constructed at the data level to improve the quality of such pictures.The experimental results show that the generative adversarial network not only improve the clarity of the blurred image to a certain extent,but also enhance the local texture of the object to be detected in the image,which is helpful for the Retina Net model to extract richer features,thereby further improving the recognition rate of blurred images.