CN-122020271-A - Rail wave mill identification method and device based on neural network fusion model

Abstract

The invention provides a rail wave mill identification method and device based on a neural network fusion model, which relate to the technical field of railway facility detection and comprise the steps of collecting vertical acceleration data of an axle box of a train; the method comprises the steps of performing discrete processing through Fourier transformation, filtering low-frequency signals, reducing time domain axle box acceleration data after trend filtering through inverse Fourier transformation, decomposing the time domain axle box acceleration data through a variation modal decomposition processing method to obtain multi-resolution inherent modal function components, selecting a plurality of components with highest signal energy proportion as input samples of a model, synchronously inputting the input samples into a convolutional neural network and a two-way long-short-term memory network, capturing local features through convolution operation, capturing two-way time sequence data dependency relationship of the axle box acceleration data by combining a two-way mechanism, and combining the convolution local features with the time sequence features through a feature fusion layer of a neural network fusion model to obtain a recognition result of a rail wave mill section.

Inventors

• HE WENXUAN
• LIU JINCHAO
• SHAO QI
• GUO JIANFENG
• XUE RUI
• ZHANG YU
• DAI CHUNPING
• LIU JINGYUAN
• SHI JING

Assignees

• 中国铁道科学研究院集团有限公司
• 中国铁道科学研究院集团有限公司基础设施检测研究所
• 北京铁科英迈技术有限公司

Dates

Publication Date

20260512

Application Date

20241112

Claims (13)

1. 1. A rail wave mill identification method based on a neural network fusion model is characterized by comprising the following steps: acquiring vertical acceleration data of an axle box of a train; performing discrete processing on the axle box vertical vibration acceleration data in a frequency domain through Fourier transformation, filtering low-frequency signals, and restoring time domain axle box acceleration data after trend filtering by utilizing inverse Fourier transformation; Decomposing the time domain axle box acceleration data through a variation modal decomposition processing method to obtain multi-resolution inherent modal function components, selecting a plurality of components with highest signal energy proportion in the inherent modal function components, and taking the selected components as input samples of a neural network fusion model, wherein the neural network fusion model is formed by fusion of a convolutional neural network and a bidirectional long-short-term memory network, and is obtained through training of axle box acceleration data corresponding to a steel rail wave grinding section and a steel rail non-wave grinding section; and synchronously inputting the input samples into a convolutional neural network and a two-way long-short-term memory network, capturing local features through convolution operation, capturing two-way time sequence data dependency relationship of the acceleration data of the axle box by combining a two-way mechanism, and combining the convolutional local features output by the convolutional neural network with the time sequence features output by the two-way long-short-term memory network through a feature fusion layer of a neural network fusion model to obtain the identification result of the rail wave mill section.
2. 2. The neural network fusion model-based rail grinds recognition method according to claim 1, characterized by collecting axle box vertical acceleration data of a train, comprising: the sliding window is adopted to slide and take the number of the axle box vertical vibration acceleration data of the train, and each preset number of the axle box vertical vibration acceleration data form one sample of data.
3. 3. The rail wave mill identification method based on the neural network fusion model according to claim 1, wherein the discrete processing is performed on the axle box vertical vibration acceleration data in a frequency domain through fourier transformation, the low-frequency signals are filtered, and the time domain axle box acceleration data after trend filtering is restored by inverse fourier transformation, and the method comprises the following steps: Converting the time domain signal into the frequency domain through Fourier transformation, wherein the transformation formula is as follows: Wherein X k is a discrete sample of the frequency domain signal, representing the complex amplitude of the kth frequency component, X N is a discrete sample of the original time domain signal, N is the number of sample points, representing the length of the signal, k is a discrete variable in the frequency domain, j is an imaginary unit; N is a discrete variable in the time domain; And restoring the time domain axle box acceleration data after trend filtering by utilizing inverse Fourier transform, wherein a transformation formula is as follows: wherein X [ N ] is a discrete sample of the reconstructed time domain signal, X [ k ] is a discrete sample of the frequency domain signal, k and N are discrete variables in the frequency domain and the time domain, j is an imaginary unit, and N is the number of sample points.
4. 4. The rail wave mill identification method based on the neural network fusion model according to claim 1, wherein the time domain axle box acceleration data is decomposed by a variation modal decomposition processing method to obtain multi-resolution intrinsic mode function components, a plurality of components with highest signal energy proportion in the intrinsic mode function components are selected, and the selected components are used as input samples of the neural network fusion model, and the method comprises the following steps: Constructing a variation model, enabling an original time sequence signal to be f (t), and solving K inherent mode function components by using Hilbert transformation, wherein the calculation formula is as follows: Wherein u k is the K-th intrinsic mode function component, t represents time, w k is the frequency center of the K-th mode, and K is the total number of modes; representing the derivative of the function with respect to time t, |·| 2 representing the L 2 norm, i being the imaginary unit, s.t. representing the constraint, f (t) representing the original time series signal; solving the variational model, and converting the variational model by using an augmented Bragg function, wherein the expression is as follows: wherein L represents an augmented Lagrangian function, alpha is a quadratic penalty factor, and lambda is a Lagrangian multiplier; and decomposing the time domain axle box acceleration data by a variation mode decomposition processing method according to the variation model to obtain a multi-order inherent mode function component.
5. 5. The rail wave mill identification method based on the neural network fusion model according to claim 4, wherein the time domain axle box acceleration data is decomposed by a variation modal decomposition processing method to obtain multi-resolution intrinsic mode function components, a plurality of components with highest signal energy ratio in the intrinsic mode function components are selected, and the selected components are used as input samples of the neural network fusion model, and the method comprises the following steps: and calculating the corresponding energy duty ratio for the multi-order inherent mode function component, wherein the energy of the inherent mode function component is calculated by adopting an energy formula of the signal, and the calculation formula is as follows: Wherein E i represents the energy of the ith order natural mode function component, S j is the jth sample value in the natural mode function component, N' is the number of samples of the natural mode function component, P i represents the proportion of the energy of the ith order natural mode function component to the total energy, E j represents the energy of the jth natural mode function component, and M is the total number of the natural mode function components; And carrying out variable-division modal decomposition on the axle box acceleration data of multiple groups of wave mills and non-wave mills, sorting according to the energy duty ratio, and selecting multiple components with the highest signal energy duty ratio as input samples of the neural network fusion model.
6. 6. The rail wave mill identification method based on the neural network fusion model according to claim 1, wherein the input samples are synchronously input into a convolutional neural network and a two-way long-short-term memory network, local features are captured through convolution operation, bidirectional time sequence data dependency relationship capturing is carried out on the acceleration data of the axle box by combining a two-way mechanism, the convolution local features output by the convolutional neural network are combined with the time sequence features output by the two-way long-short-term memory network through a feature fusion layer of the neural network fusion model, and an identification result of a rail wave mill section is obtained, and the method comprises the following steps: The selected components are synchronously input into a convolutional neural network and a two-way long-short-term memory network, wherein the convolutional neural network adopts 4 layers of convolution to capture the characteristic of the intrinsic mode function component crossing space-time points; and constructing a fusion processing layer of the convolutional neural network and the two-way long-short-term memory network, fusing output characteristics of the convolutional neural network and the two-way long-short-term memory network, superposing 4 layers of full-connection layers for characteristic fusion and dimension reduction calculation, extracting space-time characteristics under different characteristic frequency modes, returning data category probability by using a Sigmoid function by the output layer, and identifying a rail corrugation section to obtain a rail corrugation section and a rail non-corrugation section.
7. 7. The rail wave mill identification method based on the neural network fusion model according to claim 6, further comprising: And (3) tiling and expanding the high-dimensional features obtained by the convolutional neural network and the two-way long-short-term memory network into one-dimensional features, aligning the output feature dimensions of the convolutional neural network and the two-way long-short-term memory network, splicing the output features of the convolutional neural network and the two-way long-term memory network in the vertical dimension to form multi-dimensional features, performing fusion calculation and feature dimension reduction by adopting 4 full-connection layers, selecting hyperbolic tangent activation functions as activation functions for each full-connection layer, and selecting Sigmoid functions at the output layer to perform probability judgment on data types to obtain identification results.
8. 8. The rail wave mill identification method based on the neural network fusion model of claim 6, wherein the convolutional neural network consists of a plurality of superimposed convolutional layers, a pooling layer, a full-connection layer and an output layer, the convolutional layers extract multi-class characteristic information of original data through a convolutional kernel and perform dimension reduction processing on the original high-dimensional input data, and the calculation formula is as follows: RELU(x)=max(0,x); pooling(x)=max(x); In the formula, Is the calculated value of the ith neuron in the kth layer of the convolution layer; Convolving the neuron outputs of local j for the upper k-1 layer; is a k-th layer convolution kernel parameter; Is a bias term, M i is convolution output of a k-1 layer of the upper layer, f (x) is a nonlinear activation function, a RELU function is selected, x represents neuron output, pooling is pooling operation, and pooling (x) =max (x) is maximum pooling operation.
9. 9. The rail wave mill identification method based on the neural network fusion model according to claim 6, wherein the bidirectional long-short term memory network is formed by adding bidirectional on the basis of the long-short term memory network, the long-short term memory network is formed by adding a forgetting gate, an input gate and an output gate structure on the basis of the cyclic neural network, and a dependency relationship between time sequence data is established, wherein one unit state information is as follows: f t ＝σ(W f ·[h t-1 ,x t ]+b f ); i t ＝σ(W i ·[h t-1 ,x t ]+b i ); g t ＝Tanh(W g ·[h t-1 ,x t ]+b g ); o t ＝σ(W o ·[h t-1 ,x t ]+b o ); C t ＝f t *C t-1 +i t *g t ; h t ＝Tanh(C t )*o t ; Where f t is the forgetting gate output of time step t, σ is the Sigmoid activation function, W f is the state transition matrix of the forgetting gate, h t-1 is the hidden state of the previous time step, x t is the input of the current time step, b f is the bias term of the forgetting gate, i t is the input gate output of time step t, W i is the state transition matrix of the input gate, b i is the bias term of the input gate, g t is the candidate memory cell of time step t, tan h is the hyperbolic tangent activation function, W g is the weight matrix of the candidate memory cell, b g is the bias term of the candidate memory cell, o t is the output of the output gate of time step t, W o is the state transition matrix of the output gate, b o is the bias term of the output gate, C t is the cell state of time step t, C t-1 is the cell state of the previous time step t, and h t is the hidden state of time step t.
10. 10. Rail ripples grinds recognition device based on neural network fuses model, its characterized in that includes: the data acquisition module is used for acquiring vertical acceleration data of the axle box of the train; The data processing module is used for carrying out discrete processing on the vertical vibration acceleration data of the axle box in a frequency domain through Fourier transformation, filtering low-frequency signals, and restoring the time domain axle box acceleration data after trend filtering by utilizing inverse Fourier transformation; The system comprises a variation modal decomposition module, a multi-resolution intrinsic modal function component, a neural network fusion model and a data processing module, wherein the variation modal decomposition module is used for decomposing the time domain axle box acceleration data through a variation modal decomposition processing method to obtain multi-resolution intrinsic modal function components, selecting a plurality of components with the highest signal energy proportion in the intrinsic modal function components, and taking the selected components as input samples of the neural network fusion model, wherein the neural network fusion model is formed by fusion of a convolutional neural network and a two-way long-short-term memory network, and is obtained by training the axle box acceleration data corresponding to a steel rail wave grinding section and a steel rail non-wave grinding section; and the fusion identification module is used for synchronously inputting the input samples into the convolutional neural network and the bidirectional long-short-term memory network, capturing local features through convolution operation, capturing bidirectional time sequence data dependency relationship of the acceleration data of the axle box by combining a bidirectional mechanism, and combining the convolutional local features output by the convolutional neural network with the time sequence features output by the bidirectional long-short-term memory network through a feature fusion layer of a neural network fusion model to obtain the identification result of the rail wave grinding section.
11. 11. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the method of any of claims 1 to 9 when executing the computer program.
12. 12. A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program which, when executed by a processor, implements the method of any of claims 1 to 9.
13. 13. A computer program product, characterized in that the computer program product comprises a computer program which, when executed by a processor, implements the method of any of claims 1 to 9.

Description

Rail wave mill identification method and device based on neural network fusion model Technical Field The invention relates to the technical field of railway facility detection, in particular to a rail wave mill identification method and device based on a neural network fusion model. Background This section is intended to provide a background or context for embodiments of the invention. The description herein is not admitted to be prior art by inclusion in this section. At present, a backbone network of a high-speed railway is basically built, the operation mileage of the high-speed railway is increased, various rail diseases occur, and the wave-shaped abrasion of a steel rail (called as rail wave abrasion for short) is an important disease affecting the health state of the rail. The rail wave grinding is a phenomenon of uneven abrasion of the rail working surface caused by poor contact state of the wheel and the rail for a long time under the action of train load, wherein the rail surface is in a wave shape and similar in interval along the longitudinal direction. The rail wave mill not only causes abnormal vibration and noise in the running of the train, but also causes deterioration of the performance of the vehicle and the rail. The accurate identification of rail corrugation is one of the key problems of line maintenance and repair, and various methods for solving the problems are proposed in the prior art, for example, 1, by combining variation modal decomposition and correlation coefficient, decomposing and screening the axle box acceleration IMF component with the maximum correlation coefficient, and further calculating rail corrugation related parameters by using smooth pseudo-wiener-Wiley distribution to perform corrugation identification. 2. The train axle box vibration acceleration signal is decomposed into relatively stable IMFs through an adaptive noise empirical mode decomposition method, MPE and correlation coefficients are introduced to clean components with low signal to noise ratio, IMF entropy value characteristics are extracted, the energy value and total energy value of the remaining IMFs are combined, and rail wave grinding is judged through a threshold value. 3. Through the decomposition of the acceleration signal of the shaft box, the association relation between noise and the wave mill is researched, and the wave mill index for recognizing the high-speed rail wave mill is provided. 4. And (3) processing acceleration data of the train bogie by a WT method to realize wave grinding detection. 5. And the amplitude, the power spectral density and the frequency distribution range of the acceleration of the axle box are analyzed through Fourier transformation, so that the rail wave grinding detection is realized. 6. And simulating the coupling running state of the train and the steel rail, and acquiring the acceleration data of the simulation axle box to analyze the frequency spectrum of the wave grinding section. However, the detection result obtained by the method has larger error with the actual situation. With the development of artificial intelligence, machine learning and deep learning methods are also gradually applied to the field of wave mill identification, for example, 7, aiming at the problem of small acceleration data quantity of an inter-city railway steel rail wave mill axle box, a wave mill data set is expanded through an ID-GAN. 8. The subway rail wave grinding identification is carried out through 1D-CNN, however, the method only considers the local characteristics of the axle box acceleration, and does not consider the successive dependence characteristics of the time sequence data of the axle box acceleration. 9. And generating wheel track vibration acceleration simulation data through a coupling model, combining integrated empirical mode decomposition to extract bispectral features of signals, and combining a support vector machine to identify wave grinding wave depths on the basis of the bispectral features. 10. And classifying and identifying the axle box acceleration of the subway train by combining IDCNN, identifying whether rail wave grinding exists, simulating the axle box acceleration with fixed wave grinding depth and fixed length by simulation data, and identifying the wave grinding length and depth based on KSM and PSO algorithms. The research method is mainly used for obtaining higher identification accuracy based on simulation data, and effective research is not carried out based on actual measurement axle box acceleration data during operation of the high-speed railway vehicle. In addition, the wave mill identification method based on signal processing needs to manually extract the characteristics and manually set a threshold value, and the characteristic extraction process is complicated. In view of the foregoing, a technical solution is needed to overcome the above-mentioned drawbacks and improve the rail corrugation identification method and the identif