Research On Defect Recognition Algorithm Of Ground Penetrating Radar Image Of Railway Subgrade Detection And Risk Assessment Model

Railway subgrade defect is a major safety issue in the field of railway transportation.By the end of 2017,the railway operating mileage of our country exceeded 127,000 kilometers,ranking the second in the world.The railway construction has entered the stage of the high speed development.However,with the increase of operating time and operating mileage of the railway system,the railway subgrade defect is becoming more and more serious,which has brought a great threat to the safe and continuous operation of trains.As a rapid,non-destructive and efficient detection method,the vehicle-mounted geological radar method has become the main method for the railway subgrade defect detection.However,with the increase of the amount of subgrade detection data,the current way of artificial defect recognition has become a serious obstacle to the further promotion and application of geological radar technology in the field of railway subgrade detection.At present,there are mainly the following problems: 1)Artificial defect recognition method is inefficient and has a big error,which cannot timely provide scientific basis for the large-scale subgrade maintenance and repair;2)at present,the researches on recognition algorithms of railway subgrade defect mainly focus on the method of combining artificial design features with classification algorithms such as support vector machine and artificial neural network.However,the subgrade defects are various in shape and size,and the artificial design features are limited in feature representation ability.Therefore,the artificial design features are difficult to deal with the complicated and changeable subgrade defect environments.Besides,due to the lack of standard samples,the support vector machine and artificial neural network cannot optimize parameters sufficiently,so their recognition accuracy is difficult to meet the needs of practical application;3)because the influencing factors of railway subgrade defect are numerous,and the influence mechanism of various factors on the defect occurrence and development process has not been fully revealed.Therefore,the risk quantification of subgrade defect has characteristics of fuzziness,randomness and dynamics.However,the existing risk assessment model of railway subgrade defect mainly solves the fuzziness of influencing factors quantification,without considering the dynamics of defect development and the uncertainties of interactions between influencing factors.Therefore,the existing models are still not perfect in risk assessment systems and risk quantification.Therefore,the existing models are not perfect in risk assessment system and risk quantification.Based on the research and development project of China Railway General Corporation,“the key technology of railway roadbed geological radar rapid detection”（2017G003-H）,this paper studies the recognition algorithm of ground penetrating radar detection image of railway subgrade defects and the risk assessment model.The research content includes the following three aspects: 1)a fast recognition algorithm of railway subgrade defects,which is used for the preliminary rapid recognition of subgrade defects;2)a high-accuracy recognition algorithm for railway subgrade defects,which is used for secondary fine-grained recognition of subgrade defects;3)a risk assessment model of railway subgrade defects,which is used to assess the risk levels of identified defects and provide corresponding disposal measures.The research of this topic provides technical support for the rapid and accurate positioning of railway subgrade defect and the subgrade maintenance and repair.Therefore,it has important practical significance for the safe operation of railways.The main research work and innovations of this paper are as follows:（1）Aiming at the low efficiency of artificial recognition,a rapid recognition algorithm for railway subgrade defects based on SSD（Single Shot MultiBox Detector）is proposed,which can rapidly detect subgrade defects from detection images of ground penetrating radar.The SSD,a kind of target detection algorithm,has the detection speed to meet the real-time detection requirements,which can greatly improve the defect recognition efficiency.Besides,the way to use multiple convolution layer features to predict the classification and position correction of default box can well solve the problem of large changes in the size and shape of railway subgrade defects.Therefore,SSD is suitable for rapid recognition of subgrade defects.In view of the low recognition accuracy of small targets of basic version SSD,an improved method based on Atrous Spatial Pyramid Pooling（ASPP）and multi-scale feature fusion is proposed.Firstly,the ASPP feature extraction method is introduced into SSD.Using the characteristic of Atrous convolution to convolutional receptive field without reducing the resolution,the feature extraction ability of the shallow conv4₃ layer is fully exploited.By the fusion of high-resolution features with different receptive fields,the recognition accuracy of small objects is improved.Secondly,in order to make full use of the high-resolution information in conv3₃ layer and the rich semantic information in conv5₃ layer,the pooling layer and the deconvolution layer are respectively used to unify the conv3₃ and conv5₃ layer to the same scale and both of them are further merged with output feature of ASPP to form a multi-scale feature through the dot product of the elements,in order to further improve detection accuracy.In addition,by introducing the migration learning,the overfitting problem caused by insufficient training samples is solved.Through comparison experiments with the basic version SSD and other deep neural network models,it is proved that the improved SSD improves recognition accuracy while maintaining high recognition speed.（2）Aiming at the insufficient recognition accuracy of existing subgrade defect recognition algorithms,a high-precision recognition algorithm for railway subgrade defects based on Faster RCNN（Regions with Convolutional Neural Network）is proposed for the secondary detailed investigation of subgrade defects.Aiming at the problems in railway subgrade defect recognition tasks,the improvement strategies of feature cascade,online hard example mining,ASDN（Adversarial Spatial Dropout Network）,Soft-NMS（Non-maximum Suppression）and data augmentation are integrated to improve the basic version of Faster RCNN on algorithm recognition accuracy of railway subgrade defects.First of all,aiming at the size diversity of railway subgrade defect,an improved feature extraction method called feature cascade is proposed to improve the model detection accuracy,which connects features from multiple shallow layers and deep layers to form a new multi-scale feature for model detection.Then,according to the disparity in proportion between the target and background in model training process,the online hard example mining technique is applied to filter hard samples and retrain model,so as to balance the positive and negative sample proportions.Next,the Adversarial Spatial Dropout Network is used to generate hard positive samples by adversarial learning to enhance the model robustness.In addition,in order to solve the problem that the target defects closer to the one with highest confidence degree were suppressed,the Soft-NMS method was used to replace the traditional NMS to reduce the occurrence of false-positive results.Finally,aiming at the problem of small sample size of self-made training set,a variety of data enhancement techniques were integrated to prevent the model from over-fitting and to improve the model robustness.Additionally,the effectiveness of the improvement strategies adopted was verified by an ablation experiment.In order to comprehensively evaluate the performance of the improved model,a comparison experiment with traditional approaches was conducted.The experimental results indicated that the improved model achieved the greatest mAP of 83.6%,8.8% greater than Faster R-CNN and much greater than traditional SVM + HOG method with the mAP of 38.6%.Finally,the robustness of the improved model was verified by another comparison experiment of model detection performance for various defect types.（3）Aiming at the demand of defect risk grading of Beijing-Kowlow railway subgrade detection and the problems in the existing risk assessment models,a risk assessment model of railway subgrade defects is established based on cloud theory,set pair analysis theory and dynamic tracking thoughts.Taking into account the dynamic development characteristics of subgrade defects,a risk assessment method based on dynamic tracking thoughts is proposed,which takes the change rate of defect basic attributes into risk assessment system and making full use of multiple detection data to get a more realistic assessment result.Meanwhile,by introducing the cloud model which is one of the uncertainty artificial intelligence models to study the relation of the fuzziness and randomness,a risk assessment model of railway subgrade defects is established.Firstly,the asessment indicators are transformed into indicator clouds by inverse Gauss cloud algorithm based on the second-order and fourth-order center distance.Secondly,aiming at the subjectiveness of the existing weight allocation methods,a cloud-critic coupling weight optimized by least squares method is proposed to distribute weights for indicators scientifically,in which subjective and objective weights are taken into account.Finally,in view of the insufficiency of traditional similarity calculation methods of the cloud model,by introducing the theory of set pair analysis,a new calculation method of cloud model similarity based on “3En” rule and the set pair potential of set pair theory of Gauss cloud model is proposed to improve the discrimination accuracy of defect risk level.（4）A large number of experiments are designed based on the large-scale subgrade detection engineering on the Beijing-Kowloon Railway to verify the effectiveness of the proposed recognition algorithms and risk assessment model of railway subgrade defect.Firstly,the rapid recognition algorithm based on SSD and the high-precision recognition algorithm based on Faster RCNN are used to recognize subgrade defects by two rounds of screening.A probabilistic neural network risk assessment model based on fuzzy comprehensive evaluation and traditional machine learning algorithm is implemented respectively.Secondly,a risk assessment model based on fuzzy comprehensive evaluation in the traditional fuzzy mathematics theory and another one based on probabilistic neural network in traditional machine learning algorithms are realized respectively as the comparative experiment groups of defect assessment.Finally,the proposed risk assessment cloud model integrated dynamic tracking is compared with the other two models.After comparison with the excavation verification results,it is proved that the proposed model has higher accuracy and reliability.This paper explores and studies the recognition algorithm and risk assessment for railway subgrade defects.However,the following deficiencies still need to be further studied: 1)since the currently used geological radar detection equipment encrypts the original radar signal data,the research on recognition algorithm cannot be directly based on the original signal data.Therefore,the geological radar profile used in this paper loses part of the data accuracy during the radar signal generation process,which affects the defect recognition effect.2)The current recognition algorithm can identify four types of defects such as the mud pumping,subgrade settlement,ballast fouling and water abnormality,so the promotion of recognition algorithms to other types of subgrade defects needs to be further studied.3)The parameter and threshold setting of the risk assessment model need to be continuously optimized and improved in the future application practice.