| China is now leading the world in high-speed rail construction and operating miles,and progress is rising.As an important component of turnout,one of the weak links of rail lines,the slide plate is used in a harsh environment and its surface is easily corroded,which not just influences the service life of switch slide baseplate,but also has a direct correlation with the safety of high-speed train operation,so how to predict the life of switch slide baseplate under corrosive environment to provide theoretical basis for train maintenance is the key to ensure the safe operation of railroads.In this study,in view of the problem of few research results of slide plate and the existence of certain errors in the traditional single model to predict the life of the slide plate,the sequence parameters containing rich life information such as corrosion images,quality and depth of the slide plate,combined with image quality and numerical prediction model to explore the intrinsic correlation mechanism between defects and life,and establish fusion models VGG-16+improved ResNet50 network and LSTM network respectively.To realize the corrosion damage detection of switch slide baseplate,so as to develop comprehensive evaluation method for service life of switch slide baseplate in corrosive environment.The main work is as follows:(1)Prepare slide plate specimens that meet the experimental standards,conduct neutral salt spray corrosion experiments on them,and collect the slide plate corrosion images and numerical characteristics parameters in time to make data sets.(2)Aiming at the corrosion image data set of slide plate,this study proposes an image-based algorithm for the detection of corrosion damage on slide plates and the effect on their service life.The algorithm has five major improvements,including increasing the input image size to ensure that the convolutional layer can extract enough features;using small convolutional kernels instead of large convolutional kernels to effectively reduce the number of parameters and training time;optimizing the residual structure to improve the prediction effect of the model;adding the SENet feature extraction module to the network to distinguish the importance of slide plate corrosion features;utilizing a model fusion method to increase the model’s generalizability and precision.The results show that the fusion model VGG-16+improved ResNet50 has a recognition accuracy of 99%under the same number of iterations,which is greater than that of the improved ResNet50 and Inception+improved ResNet50 network models.It indicates that it is easier to identify the slide plate corrosion defects,and its GPU training time is 3.44h,which is nearly 39 times shorter than the CPU training time.(3)In order to improve the accuracy of model prediction and fully exploit the intrinsic relationship between the slide plate specimen data,this work suggests an LSTM-based numerical prediction model.The model mainly predicts the life of the slide plate by the collected slide plate specimen characteristic parameters,and the input data mainly includes the four characteristic parameters of quality,depth,average corrosion rate and corrosion area.The findings demonstrate that the LSTM model is more effective than the GM since its projected value is more in line with the actual value.In addition,the loss functions MSE,MAE,RMSE and MAPE are close to zero when LSTM is used to predict the four types of feature values.(4)This study combines the image and numerical prediction model results to develop a comprehensive evaluation method for the service life of switch slide baseplate in corrosive environment,which not only can quickly and accurately evaluate the damage degree of switch slide baseplate from multiple angles,but also can provide a new way of thinking for metal corrosion fatigue prediction life.At the same time,the study has an important scientific value and theoretical significance for the establishment of life evaluation guidelines for slide plate in China. |
PDF Full Text Request