| Railroads are our country’s transportation backbone,playing a vital role in the healthy development of the national economy.With the high-speed development of the railroad and the change in working mode,the rail,axle,fasteners,and other important structures to bear the load and extrusion impact are more likely to appear defective.These problems directly affect the safe operation of the railroad,so a more efficient,intelligent,and accurate detection of rail injury and damage is the future development trend of rail damage detection technology.In such a context,this paper establishes a deep learning-based method for rail damage identification based on the analysis of rail vibration signals.1.In this paper,we first investigate the traditional damage detection theoretical methods and obtain the corresponding features by selecting several time-frequency analysis methods that perform well in the task of rail damage identification and calculating the original signal.The original signal and the features obtained by the time-frequency analysis method are combined by a feature combination method,and the combined one-dimensional data are transformed into two-dimensional grayscale maps using a signal-image conversion method as the input to the convolutional neural network.The designed convolutional neural network is then used to extract key information for image classification of the two-dimensional grayscale map,thus indirectly achieving the classification of the original data,the presence or absence of damage to the rails.This method of transforming the signal analysis problem into an image processing problem realizes the combination of a traditional theoretical detection method and a deep learning algorithm.The experimental results prove that this traditional theoretical detection method combined with a deep learning algorithm can accurately and efficiently identify whether there is damage to the rail.2.Algorithm improvement: Based on convolutional neural networks for damage recognition,we analyze the problem of implied weights in the joint approach using multiple features.We propose a pre-convolution method to reduce the impact of different feature sizes.It has been experimentally demonstrated that the pre-convolution approach can effectively solve the problem and improve the classification accuracy.In addition,this paper introduces the residual structure to further optimize the network and analyzes and compares several models in the paper using the model evaluation method.The results prove that the convolutional neural network with residual structure and pre-convolutional processing has a better recognition effect on the existence of damage to rails. |
