Research On Intelligent Diagnosis And Health Assessment Of Turnout Based On Adaptive Curve Segmentation

As a key component in the railway system,turnouts play a vital role in the stable and safe operation of trains.Ensuring the safe operation of turnouts and adopting intelligent means to detect and evaluate the status of turnouts is the intelligent development trend of railway,and it is also an important means to strengthen railway safety management and achieve controllable operational safety risks.Considering the problems of different multi-stage operation mechanisms,missing fault samples,unbalanced samples,multi fault concurrency and similar fault characteristics of turnout in reality,based on the in-depth analysis of turnout action principle and operation curve,this paper breaks through the problems of turnout novelty detection,condition assessment and fault identification,and forms an intelligent diagnosis and health assessment system of turnout based on adaptive curve segmentation,which realizes efficient and accurate intelligent fault detection and health condition assessment of turnouts.The specific research contents are as follows:(1)In response to the difficulty of adaptive segmentation feature extraction of electric power curve under the multi-stage operating characteristics of the turnout,this paper proposes a feature vector construction method of the turnout based on adaptive curve segmentation and m RMR algorithm.Firstly,the operation mechanism of typical electric turnout is deeply analyzed.Based on the characteristics of turnout power curves,curve segmentation indicators such as slope and coefficient of variation in different sliding windows are proposed.On this basis,the electric power curve is adaptively segmented and extracted,and the curve is adaptively divided into four stages: start,unlocking,conversion and locking according to the similarities and differences of the operation mechanism.Then,the time domain analysis is combined with the Maximum Relevance and Minimum Redundancy(m RMR)algorithm to extract the effective health features of each stage,and the feature selection optimization based on m RMR under the adaptive curve segmentation is realized.Finally,the effectiveness of the proposed adaptive segmentation feature extraction method is verified by comparative experiments.(2)To solve the problem of fast detection of turnout fault in the absence of fault samples,on the basis of the construction of segmented m RMR features,this paper proposes a turnout novelty detection model based on kernel-target alignment hybrid kernel function SVDD(KASVDD).This method draws on the framework of novelty detection concept,and only uses normal samples for model generation,which can realize fast and efficient detection of unknown faults of turnout.In addition,this paper also optimizes the SVDD model from the perspective of kernel function optimization.In this paper,a new hybrid kernel function is proposed to replace the single kernel function in the traditional SVDD,so that the high-dimensional map restoration of the data distribution structure can be more comprehensively realized,and the model boundary can more effectively describe the intrinsic characteristics of the data.At the same time,the hybrid kernel function also avoids the difficulty of kernel parameter selection and the dependence on expert experience of traditional algorithms.In the construction of the hybrid kernel function,this paper also introduces the kernel-target alignment method to calculate the weight of each basic kernel function,and obtain a new hybrid kernel function through linear combination,so as to construct a self-consistent hybrid kernel function SVDD model for the training data.Finally,taking the harmonic F1-score of Recall and Specificity as the model evaluation indicators,the proposed turnout novelty detection model is verified.Compared with the optimal result of single kernel function,this method can achieve or even exceed the model performance of the optimal kernel parameters while avoiding the difficulty of kernel function selection,which verifies the effectiveness of this method in the application of turnout novelty detection.(3)In response to the problem of turnout fault diagnosis in the case of unbalanced samples and concurrent faults,this paper proposes a concurrent fault diagnosis model for turnouts based on cost-sensitive and fixed-input Extreme Learning Machine(cf-ELM).Firstly,in order to effectively overcome the performance instability of the ELM model,this paper proposes a constraint equation of input weight and bias.Through the definition of the constraint equation,the input weight and bias matrix with boundary are integrated,and a more stable and efficient input mapping is obtained,which fully optimizes the ELM network structure.Secondly,this paper also embeds the fault severity and sample type ratio into the cost calculation rules of the model,and proposes a cost-sensitive optimization method to solve the problem of sample imbalance.On this basis,aiming at the common concurrent multi-fault diagnosis problem of turnout,a phased hierarchical fault diagnosis model is established to realize the concurrent fault diagnosis of turnout.Finally,through comparative experiments,it is verified that the algorithm shows good stability and superior diagnostic accuracy in both single fault and concurrent fault diagnosis of turnouts.(4)In response to the problem of fast assessment of nonlinear degradation of turnout health state,this paper proposes a turnout health assessment model based on nonlinear Dynamic Time Warping(NLDTW).Firstly,in view of the concept of manifold learning,this paper introduces nonlinear distance measurement,which breaks through the limitation of traditional DTW that can only measure data points linearly,comprehensively improves and optimizes the ability of DTW to describe nonlinear trends,and better describes the nonlinear differences in different health states of turnouts.Secondly,considering the gradual evolution law of the health state of turnout,this paper proposes a new health assessment function based on improved tanh funtion,which makes it more sensitive to the early degradation of turnout,thus realizing the mapping characterization of nonlinear distance to health degree.Finally,through the comparison experiments with non-segmented and traditional DTW methods,the superiority of this method in turnout health assessment and early degradation identification in each stage is verified.