Optimization method of parallel pooling layer for optimizing surface abrasion detection model of pantograph carbon contact strip

Abstract

The invention discloses an optimization method of a parallel pooling layer for optimizing the surface abrasion detection model of a pantograph carbon contact strip. The method comprises the followingsteps: 1) collecting a defect image on the surface of the pantograph carbon contact strip, and carrying out image preprocessing to obtain a data set; wherein the data set comprises training data and test data; the training data comprise two types of data, namely label data and label-free data; 2) building a semi-supervised convolution neural network under a CAFFE framework, and training the modelby utilizing the label-free data; 3) replacing an original pooling layer by a parallel pooling layer on the basis of a random pooling principle, subjecting the label data and the label-free data to differential sampling, and optimizing the surface abrasion detection model of the pantograph carbon slide plate on the basis of the semi-supervised convolution neural network. According to the invention, differential sampling is carried out on different attribute data by adopting the parallel pooling layer. The utilization efficiency of the semi-supervised convolution neural network on the label-free data is improved. The optimization effect of the surface abrasion detection model of the pantograph carbon contact strip is improved.

Description

technical field [0001] The invention relates to the field of rail transit vehicle equipment fault diagnosis. More specifically, it relates to an optimization method for the detection model of the surface wear of the pantograph carbon sliding plate with parallel pooling layers. Background technique [0002] With the rapid development of computer computing power and related technologies, deep learning theory is being used more and more widely. Its automation and intelligence in the field of image recognition makes it possible to apply it to the judgment of the wear type of the pantograph slide surface based on image recognition. Compared with traditional image detection methods, deep learning only needs to build an appropriate network model and perform simple preprocessing on the original image, then it can use the network for autonomous learning and feature extraction to achieve fully automatic image recognition; the model Once trained, it can be used directly for image rec...

AI Technical Summary

Problems solved by technology

However, for the traditional convolutional neural network structure

Images