A method for extracting geometric irregularities of rail welds

Abstract

The present invention provides a method for extracting geometric irregularities of a rail weld, comprising the following steps: S1: preparing a rail corrugation instrument for obtaining rail short-wave data, using the rail corrugation instrument of a railway maintenance section to collect the short-wave data of the rail; and using the rail short-wave irregularity data to manually mark the center position of the rail weld; S2: using a convolutional neural network and a fully connected network to extract the main features of the rail short-wave geometric data; S3: encoding the position of the rail weld, calculating a loss function with the encoding features and the features extracted by the neural network, and optimizing the parameters of the neural network through continuous training; S4: using a tile-coding decoding method to obtain the accurate position of the rail weld, and extracting the specific geometric waveform of the rail weld in the rail short-wave geometric detection data. The present application can use the existing detection data to mine the geometric information of the rail weld, improve the detection efficiency of the rail weld, save the detection cost, and deeply mine the hidden information of the existing detection data.

Description

Technical Field

[0001] The present invention relates to a method for detecting geometric waveforms of rail welds, and in particular to a method for extracting geometric irregularities of rail welds, which can apply the method to detection data of a corrugator, extract the geometric waveforms of rail welds in track irregularity data, deeply mine hidden information of existing detection data, and improve the efficiency of using detection data. Background Art

[0002] With the development of seamless railway tracks, compared with ordinary lines, the welded joints of seamless lines replace the bolted joints, providing continuous, smooth and stable support and guidance for vehicles, significantly reducing wheel-rail impact force and noise.

[0003] However, due to the continuous influence of train loads and various environmental factors, the welded joints gradually suffer fatigue damage and cause their structural degradation, eventually leading to catastrophic failure. The metallurgical and geometric discontinuities of the rail welded joints destroy the smoothness of the rails, increase the running resistance of the train, and are one of the sources of vibration of locomotives and vehicles, which is directly related to the smoothness, safety and comfort of train operation.

[0004] Due to the complex wheel-rail interaction between trains and rails, it is difficult to control the feature extraction, cause analysis and evolution law of the weld geometry under the repeated action of wheel-rail force. In addition, the irregular grinding of the rail surface can easily cause the contour of the weld joint to deteriorate, resulting in morphological wear, plastic deformation, fatigue cracks and even failure and broken rails. If there is a lack of corresponding effective weld geometry detection measures, the damage to the weld joint will be further aggravated and the maintenance cost will rise sharply.

[0005] At present, the detection of geometric irregularities in rail welds mainly relies on manual operation. The main equipment for collecting the geometric amplitude of the weld is a feeler gauge. When using the feeler gauge for collection, it is necessary to judge whether the geometric amplitude of the weld exceeds the limit based on the number of steel sheets that can be inserted. However, this method has low working efficiency, and the detection accuracy is not easy to guarantee. In addition, it can only judge the maximum amplitude of the weld, and it is difficult to extract the true waveform of the weld.

[0006] In addition, some existing technical patent documents use the principle of laser ranging. Although it can improve the detection accuracy, it is expensive and difficult to achieve large-scale application. Summary of the invention

[0007] The problems existing in the prior art solved by this application are:

[0008] The existing technology for detecting geometric unevenness of rail welds has problems such as low work efficiency, unstable detection accuracy, and high cost.

[0009] The solution to the technical problem of the present invention is:

[0010] A method for extracting geometric irregularities of a rail weld is provided, comprising the following steps:

[0011] S1: Prepare a rail corrugation instrument for obtaining rail short-wave data, use the rail corrugation instrument of the railway maintenance section to collect rail short-wave data; and use the rail short-wave irregularity data to manually mark the center position of the rail weld;

[0012] S2: Convolutional neural network and fully connected network are used to extract the main features of rail shortwave geometry data;

[0013] S3: Encode the rail weld position, calculate the loss function by combining the encoded features with the features extracted by the neural network, and optimize the neural network parameters through continuous training;

[0014] S4: Use the Tile-coding decoding method to obtain the exact position of the rail weld and extract the specific geometric waveform of the rail weld in the rail short-wave geometric detection data.

[0015] Preferably, in step S1, the weld positions in the collected rail shortwave data need to be manually marked in the early stage, and each manually marked weld position is taken as the center, and the length Lm is taken as the window length, and the left and right random translation is performed. Generate different data sets; the short-wave waveform of the rail with a length of Lm is the input of the model, and the label is the distance A of the random translation to the left and right.

[0016] Preferably, in step S4, for the judgment of the rail weld position, the continuously corresponding weld center position offsets are divided into multiple Tiles; the Tiles are regarded as multiple categories, each category is associated with a unique range; a finite number of Tiles constitute a Tiling, representing a layer of the rail weld center position; multiple Tilings with different offsets are used to describe a specific range of rail wave weld center position offsets; in each Tiling, except for the last Tile, the wave depth mean range length of all tiles is the same; wherein, the weld center position offset starts from zero and tends to infinity, and the last Tile in each Tiling is regarded as a Tile whose amplitude range extends infinitely into the future.

[0017] Preferably, there are m Tilings, each Tiling has n Tiles, and the binary vector of the offset A of the rail weld after encoding is recorded as , where each element It can be expressed as:

[0018] ,

[0019] In the formula, The amplitude variation range for each Tile; The offset of Tiling, the value range is ; Represents the index of n tiles, The index of m Tilings; the offset A of the weld center is converted into a one-dimensional vector of length MxN.

[0020] Preferably, in step S3, a deep learning training model is built, the input of the model is a short-wave waveform of a rail with a length of L, and the relevant features of the waveform are extracted by a one-dimensional convolution kernel. Finally, a one-dimensional vector B with a length of MxN (consistent with the dimension of T(A)) is output by a fully connected layer; MSE (mean root mean square error) is used as the loss function to calculate the loss function of the output vector B and T(A), and the parameters of the deep learning model are updated by the stochastic gradient descent method until the model converges.

[0021] Preferably, after the model is completed, the corresponding rail shortwave waveform is input to obtain the output vector B, and the output vector B needs to be further decoded to obtain the rail weld position predicted by the model; the decoding steps are step 1: divide L into K segments, that is, , and perform Tile-coding on Lˊ to obtain a matrix of dimension Kx(MxN), denoted as P, with K rows and MxN columns; step 2: Calculate the probability of each weld offset after decoding: S=PxB; S is a Kx1 vector, corresponding to the probability of rail welds appearing in Lˊ at K positions; select the maximum index value of the weld offset probability: ; At this time, the first element in Lˊ is recorded as the position of the rail weld, that is, .

[0022] The technical effects of solving the technical problem in this application are as follows:

[0023] Compared with the prior art, the present invention provides a method for extracting geometric irregularities of rail welds by adopting the following steps: S1: preparing a rail corrugation instrument for obtaining rail short-wave data, using the rail corrugation instrument of the railway maintenance section to collect the short-wave data of the rail; and using the rail short-wave irregularity data to manually mark the center position of the rail weld; S2: using a convolutional neural network and a fully connected network to extract the main features of the rail short-wave geometric data; S3: encoding the rail weld position, calculating the loss function by combining the encoded features with the features extracted by the neural network, and optimizing the neural network parameters through continuous training. S4: Use the Tile-coding decoding method to obtain the exact position of the rail weld, extract the specific geometric waveform of the rail weld in the rail short-wave geometric detection data, and use the existing rail wave grinding instrument (basically all railway engineering sections are equipped with this equipment) to collect enough rail short-wave geometric irregularities. On the basis of this, a deep learning algorithm is used to automatically identify and extract the rail weld geometric waveform in the geometric data, which can greatly improve the rail weld detection efficiency and save detection costs. At the same time, it can deeply mine the hidden information of the existing detection data and improve the utilization rate of the existing detection data. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the process of a method for extracting geometric irregularities of a rail weld according to the present invention;

[0025] Figure 2 It is a schematic diagram of extracting a rail weld waveform from a rail short wave waveform in a method for extracting geometric irregularity of a rail weld of the present invention;

[0026] Figure 3 It is a Tile-coding coding example diagram in a method for extracting geometric irregularities of a rail weld according to the present invention;

[0027] Figure 4 and Figure 5 It is a rail weld waveform extracted from a rail short wave waveform in a method for extracting geometric irregularity of a rail weld of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0029] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0030] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.

[0031] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there can be a central component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there can be a central component at the same time.

[0032] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0033] The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0034] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0035] See also Figures 1 to 5 The present invention provides a method 1 for extracting geometric irregularities of a rail weld, comprising the following steps:

[0036] S1: Prepare a rail corrugation instrument for obtaining rail short-wave data, use the rail corrugation instrument of the railway maintenance section to collect rail short-wave data; and use the rail short-wave irregularity data to manually mark the center position of the rail weld;

[0037] S2: Convolutional neural network and fully connected network are used to extract the main features of rail shortwave geometry data;

[0038] S3: Encode the rail weld position, calculate the loss function by combining the encoded features with the features extracted by the neural network, and optimize the neural network parameters through continuous training;

[0039] S4: Use the Tile-coding decoding method to obtain the exact position of the rail weld and extract the specific geometric waveform of the rail weld in the rail short-wave geometric detection data.

[0040] This patent adopts the following steps S1: prepare a rail corrugation instrument for obtaining rail short-wave data, and use the rail corrugation instrument of the railway maintenance section to collect the short-wave data of the rail; and use the short-wave unevenness data of the rail to manually mark the center position of the rail weld; S2: use a convolutional neural network and a fully connected network to extract the main features of the rail short-wave geometric data; S3: encode the position of the rail weld, calculate the loss function with the encoding features and the features extracted by the neural network, and optimize the neural network parameters through continuous training; S4: use the Tile-coding decoding method to obtain the accurate position of the rail weld, extract the specific geometric waveform of the rail weld in the rail short-wave geometric detection data, use the existing rail corrugation instrument (basically all railway maintenance sections are equipped with this equipment), on the basis of collecting enough short-wave geometric unevenness of the rail, use the deep learning algorithm to automatically identify and extract the geometric waveform of the rail weld in the geometric data, which can greatly improve the efficiency of rail weld detection and save detection costs. At the same time, it can deeply mine the hidden information of the existing detection data and improve the utilization rate of the existing detection data.

[0041] In some other embodiments, in step S1, the weld positions in the collected rail shortwave data need to be manually marked in the early stage, and each manually marked weld position is taken as the center, and the length is Lm as the window length, and the left and right random translation is performed. Generate different data sets; the short-wave waveform of the rail with a length of Lm is the input of the model, and the label is the distance A of the random translation to the left and right.

[0042] Tile-coding is a practical and computationally efficient tool. For the determination of rail weld locations, the continuously corresponding weld center position offsets are divided into multiple tiles, which are considered as multiple categories, each of which is associated with a unique range.

[0043] In step S4, for the determination of the rail weld position, the continuously corresponding weld center position offsets are divided into multiple Tiles; the Tiles are regarded as multiple categories, each category is associated with a unique range; a finite number of Tiles constitute a Tiling, representing a layer of the rail weld center position; multiple Tilings with different offsets are used to describe a specific range of rail wave weld center position offsets; in each Tiling, except for the last Tile, the wave depth mean range length of all tiles is the same; wherein, the weld center position offset starts from zero and tends to infinity, and the last Tile in each Tiling is regarded as a Tile whose amplitude range extends infinitely into the future.

[0044] There are m Tilings, each Tiling has n Tiles, and the binary vector of the offset A of the rail weld after encoding is recorded as , where each element It can be expressed as:

[0045] ,

[0046] In the formula, The amplitude variation range for each Tile; The offset of Tiling, the value range is ; Represents the index of n tiles, The index of m Tilings; the offset A of the weld center is converted into a one-dimensional vector of length MxN.

[0047] Typically, the binary vector of tile-coding is defined as follows: if two conditions are met, , is called the tile encoding vector with parameters m and n, the two conditions are as well as .

[0048] The attached drawings of the specification of this application show the offset of two tile-coded rail welds in three tiles (m=3), each containing four tiles (n=4). In the example of space (a) and (b), the rail weld offset (a) is located in Tile-1 of Tiling-1, and in Tile-2 of Tiling-2 and Tiling-3. Therefore, the encoding vector of the rail weld offset (a) is expressed as Similarly, the encoding vector of the rail weld offset (b) is In summary, this step forms a data set for rail weld identification, where the input is a short-wave rail waveform with a length of Lm, the label is the offset A of the rail weld position in Lm relative to the center position of the waveform, and the offset A is tile-coded for the next step of training. The offset A after tile-coding is recorded as T(A).

[0049] In the step S3, a deep learning training model is built, the input of the model is a short-wave waveform of a rail with a length of L, and the relevant features of the waveform are extracted through a one-dimensional convolution kernel. Finally, a one-dimensional vector B with a length of MxN (consistent with the dimension of T(A)) is output by a fully connected layer; MSE (mean root mean square error) is used as the loss function to calculate the loss function of the output vector B and T(A), and the parameters of the deep learning model are updated by the stochastic gradient descent method until the model converges.

[0050] After the model is completed, the corresponding rail shortwave waveform is input to obtain the output vector B, which needs to be further decoded to obtain the rail weld position predicted by the model; the decoding steps are step 1: divide L into K segments, that is, , and perform Tile-coding on Lˊ to obtain a matrix of dimension Kx(MxN), denoted as P, with K rows and MxN columns; step 2: Calculate the probability of each weld offset after decoding: S=PxB; S is a Kx1 vector, corresponding to the probability of rail welds appearing in Lˊ at K positions; select the maximum index value of the weld offset probability: ; At this time, the first element in Lˊ is recorded as the position of the rail weld, that is, .

[0051] This patent can use existing inspection data to mine the geometric information of rail welds, greatly improving the efficiency of rail weld inspection and saving inspection costs. At the same time, it can deeply mine the hidden information of existing inspection data and improve the utilization rate of existing inspection data.

[0052] The accuracy of weld center recognition can reach over 90%.

[0053] The technical effects of solving the technical problem in this application are as follows:

[0054] Compared with the prior art, the present invention provides a rail weld geometric irregularity extraction method 1 by adopting the following steps S1: preparing a rail corrugation instrument for obtaining rail short-wave data, using the rail corrugation instrument of the railway maintenance section to collect the rail short-wave data; and using the rail short-wave irregularity data to manually mark the center position of the rail weld; S2: using a convolutional neural network and a fully connected network to extract the main features of the rail short-wave geometric data; S3: encoding the rail weld position, calculating the loss function by combining the encoded features with the features extracted by the neural network, and adjusting the neural network parameters through continuous training. Optimization; S4: Use the Tile-coding decoding method to obtain the exact position of the rail weld, extract the specific geometric waveform of the rail weld in the rail short-wave geometric detection data, and use the existing rail corrugation instrument (basically all railway engineering sections are equipped with this equipment) to collect enough rail short-wave geometric irregularities. On the basis of this, a deep learning algorithm is used to automatically identify and extract the rail weld geometric waveform in the geometric data, which can greatly improve the rail weld detection efficiency and save detection costs. At the same time, it can deeply mine the hidden information of the existing detection data and improve the utilization rate of the existing detection data.

[0055] The above-described embodiments of the present invention do not constitute a limitation on the protection scope of the present invention. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for extracting geometric irregularities of rail welds, characterized in that: The following steps are included: S1: Prepare a rail corrugation instrument for obtaining rail short-wave data, use the rail corrugation instrument of the railway maintenance section to collect rail short-wave data; and use the rail short-wave irregularity data to manually mark the center position of the rail weld; S2: Convolutional neural network and fully connected network are used to extract the main features of rail shortwave geometry data; S3: Encode the rail weld position, calculate the loss function by combining the encoded features with the features extracted by the neural network, and optimize the neural network parameters through continuous training; S4: Using Tile-coding decoding method, the accurate position of rail weld is obtained, and the specific geometric waveform of rail weld in rail short-wave geometric detection data is extracted; In the step S3, a deep learning training model is built, the input of the model is a short-wave waveform of a rail with a length of L, and the relevant features of the waveform are extracted through a one-dimensional convolution kernel, and finally a one-dimensional vector B with a length of MxN is output by a fully connected layer; MSE (mean root mean square error) is used as a loss function to calculate the loss function of the output vector B and T (A), and the parameters of the deep learning model are updated by the stochastic gradient descent method until the model converges; After the model is completed, the corresponding rail shortwave waveform is input to obtain the output vector B, which needs to be further decoded to obtain the rail weld position predicted by the model; the decoding steps are step 1: divide L into K segments, that is, , and perform Tile-coding on Lˊ to obtain a matrix of dimension Kx(MxN), denoted as P, where the matrix P has K rows and MxN columns; step 2: Calculate the probability of each weld offset after decoding: S=PxB; S is a Kx1 vector, corresponding to the probability of rail welds appearing in Lˊ at K positions; Select the maximum index value of weld offset probability: ; At this time, the first element in Lˊ is recorded as the position of the rail weld, that is, ; In the step S4, for the determination of the rail weld position, the continuously corresponding weld center position offsets are divided into a plurality of tiles; the tiles are regarded as a plurality of categories, each category being associated with a unique range; a limited number of the tiles constitute a tile, representing a layer of the rail weld center position; Multiple Tilings with different offsets are used to describe a specific range of rail wave weld center position offsets; in each Tiling, except for the last Tile, the wave depth average range lengths of all Tiles are the same; wherein the weld center position offset starts from zero and tends to infinity, and the last Tile in each Tiling is regarded as a Tile whose amplitude range extends infinitely into the future; There are m Tilings, each Tiling has n Tiles, and the binary vector of the offset A of the rail weld after encoding is recorded as , where each element It can be expressed as: In the formula, The amplitude variation range for each Tile; The offset of Tiling, the value range is ; Represents the index of n tiles, The index of m Tilings; the offset A of the weld center is converted into a one-dimensional vector of length MxN.

Images