Track static irregularity trend prediction method and system, electronic device and storage medium

Abstract

The application provides a track static irregularity trend prediction method, system, electronic equipment and storage medium. The method comprises the following steps: performing similarity test on a plurality of acquired historical data by using a dynamic time warping algorithm, and representing the similarity test result by using a path matrix; fitting and correcting the path matrix by using a robust regression method to obtain fitting data, and repeatedly performing similarity test on the fitting data until the fitting data meets a preset condition, and then ending the fitting and correction to obtain a dynamic warping path matrix; rechecking necessary mileages in the path matrix and the dynamic warping path matrix by using a correlation coefficient method; aligning the plurality of historical data when the rechecking requirement is met; and training the aligned plurality of historical data by using a long short-term memory network model with preset hyperparameters. The application can make accurate prediction on small-scale static single data in the track irregularity degradation state within a range of several meters.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of railway track irregularity prediction, and particularly relates to a track static irregularity trend prediction method, system, electronic device and storage medium. BACKGROUND

[0002] The railway is developing rapidly, but the trend of "high speed" and "heavy load" brought by its rapid development also puts higher requirements on the safety maintenance of the track. With the continuous development of high-speed and heavy-load of railway transportation, the track repeatedly bears the load of locomotives and vehicles, gradually changing the track gauge, level and height of the track geometry and spatial position, causing track irregularity; the track irregularity directly threatens the safety of train operation. However, the track maintenance work not only faces the pressure of huge task quantity, but also has to deal with the limitation of compressed window time due to high operation density. In order to solve this contradiction, industry experts are working to change the "periodic repair" of the track according to the periodic maintenance to the "state repair" of the track according to the track irregularity state. In order to implement and popularize the concept of "state repair", how to accurately predict the future irregularity state of the track is one of the key problems.

[0003] At present, one type of track irregularity state prediction work is to explain the change of the overall quality of the track through the prediction of the comprehensive index, and the other type is to explain the deterioration degree of a certain type of irregularity state through the prediction of a single index. The former emphasizes the integrity, the latter emphasizes the pertinence, and in theory, the relevant position maintenance work should be guided according to the prediction results. But most of the existing researches on the prediction of a single index to explain the deterioration degree of a certain type of irregularity state are based on the research object of hundreds of meters or even larger mileage, which is relatively limited in helping the maintenance personnel to accurately locate the problem. SUMMARY

[0004] In order to solve the above technical problems, the present application provides a track static irregularity trend prediction method, system, electronic device and storage medium, which adjusts and optimizes the mileage deviation of track historical data, and uses a long short-term memory network model to train the aligned historical data, so as to make a small-range accurate prediction of the static single data in the track irregularity deterioration state within a number of meters, and effectively help the maintenance personnel to quickly and accurately locate the track irregularity position.

[0005] In a first aspect, the present application provides a track static irregularity trend prediction method, comprising:

[0006] The obtained historical data is subjected to similarity test by using a dynamic time warping algorithm, and the result of the similarity test is represented by a path matrix;

[0007] The path matrix is fitted and corrected by using a robust regression method to obtain fitting data, and the fitting data is repeatedly subjected to the similarity test until the fitting data reaches a preset condition, and then the fitting correction is ended to obtain a dynamic regular path matrix;

[0008] The necessary mileage in the path matrix and the dynamic regular path matrix is rechecked by a correlation coefficient method;

[0009] When the data of the necessary mileage does not meet the rechecking requirements, the necessary mileage is re-fitted and corrected by using a starting point resetting method or a scaling method to realize alignment of the several pieces of historical data;

[0010] The aligned several pieces of historical data are trained by a long short-term memory network model with preset hyperparameters, so as to predict future data of track static irregularities, and the prediction effect is characterized by a root mean square error and an average deviation.

[0011] Preferably, the step of performing similarity test on the obtained several pieces of historical data by using a dynamic time warping algorithm, and characterizing the result of the similarity test by a path matrix, specifically comprises:

[0012] Obtaining several pieces of historical data, taking any one of the several pieces of historical data as reference data, and taking other historical data in the several pieces of historical data as to-be-tested data;

[0013] Combining and pairing each to-be-tested data with the reference data;

[0014] Performing similarity test on the combined and paired to-be-tested data and reference data by using a dynamic time warping algorithm, and characterizing the test result of the similarity test by a path matrix.

[0015] Preferably, the preset condition is any one of recurrence, adjacent recurrence, and alternate recurrence.

[0016] Preferably, in the fitting process of the robust regression method, a double-threshold method is used to make symmetric optimization of the dynamic regular path matrix, form a hexagonal region, and save the calculation amount in the generation process of the dynamic regular path matrix.

[0017] Preferably, the preset hyperparameters include at least one or a combination of several of a training ratio, a batch quantity, a loss function, an optimization function, a learning rate, a precision requirement, a number of hidden layers, and a number of hidden layer neurons.

[0018] Preferably, the root mean square error represents the fitting degree of the measured value and the true value curve, which is used to measure the accuracy of the measurement; the average deviation refers to the arithmetic mean of the random error of all measured values in the equal accuracy measurement, which is used to reflect the deviation of the data in the interval.

[0019] In a second aspect, the application provides a track static irregularity trend prediction system, comprising:

[0020] The test module is configured to perform similarity test on the obtained historical data by using a dynamic time warping algorithm, and to represent the result of the similarity test by a path matrix.

[0021] The fitting module is configured to perform fitting correction on the path matrix by using a robust regression method to obtain fitting data, and to repeatedly perform the similarity test on the fitting data until the fitting data meets a preset condition, and then to end the fitting correction to obtain a dynamic warping path matrix.

[0022] The retest module is configured to retest necessary mileages in the path matrix and the dynamic warping path matrix by using a correlation coefficient method.

[0023] The alignment module is configured to perform re-fitting correction on the necessary mileages by using a starting point resetting method or a scaling method when the data of the necessary mileages does not meet the retest requirement, so as to align the historical data.

[0024] The prediction module is configured to train the aligned historical data by using a long short-term memory network model with preset hyperparameters, so as to predict future data of the track static irregularity, and the prediction effect is represented by a root mean square error and an average deviation.

[0025] Preferably, the test module further comprises:

[0026] The definition unit is configured to obtain a plurality of historical data, and to take any one of the historical data as reference data and the other historical data as test data.

[0027] The pairing unit is configured to pair each of the test data with the reference data.

[0028] The test unit is configured to perform similarity test on the test data and the reference data by using a dynamic time warping algorithm, and the test result of the similarity test is represented by a path matrix.

[0029] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the track static irregularity trend prediction method according to the first aspect.

[0030] In a fourth aspect, the present application provides a storage medium having a computer program stored thereon, and the program is executable on a processor to implement the track static irregularity trend prediction method according to the first aspect.

[0031] Compared with the prior art, the present application provides a track static irregularity trend prediction method, system, electronic device and storage medium; based on a plurality of collected track measurement geometry data, first, taking any one of the historical data as the baseline data for deviation correction, and taking the other baseline data as the to-be-corrected data, pairing each piece of data with the baseline data, and calculating the similarity of the two paired historical data by using the dynamic time warping algorithm; secondly, combining dynamic time warping with robust regression method, calculating the mileage deviation of the two paired historical data and correcting to obtain fitting data, and again judging the similarity of the fitting data until the preset condition is met to end the fitting and correction to obtain the dynamic warping path matrix; performing a rechecking mechanism on the dynamic warping path matrix, and regarding the alignment of the plurality of historical data as meeting the rechecking requirement; otherwise, the start point resetting method or the scaling method needs to be used to re-correct the necessary mileage; training the aligned plurality of historical data through a long short-term memory network model with a preset hyperparameter, so as to predict the future data of the track static irregularity; the present application adjusts and optimizes the mileage deviation of the track historical data, and trains the aligned historical data by using the long short-term memory network model, so as to make a small-range accurate prediction on the static single-item data in the track irregularity deterioration state within a number of meters, and effectively helps the track maintenance personnel to quickly and accurately locate the track irregularity position. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0033] Figure 1 A perspective view of the track static irregularity trend prediction method provided for the first embodiment of the present application;

[0034] Figure 2 A phase sequence of two time series;

[0035] Figure 3A specific flow chart of step S101 provided for the embodiment one of the present application is shown in the following table:

[0036] Figure 4 A left high-low initial data comparison chart of January and February provided for the embodiment one of the present application is shown in the following table:

[0037] Figure 5 A left high-low initial data comparison chart of January and March provided for the embodiment one of the present application is shown in the following table:

[0038] Figure 6 A left high-low first matching path chart of January and February provided for the embodiment one of the present application is shown in the following table:

[0039] Figure 7 A partial area enlarged view of Figure 6

[0040] Figure 8 A left high-low first matching path chart of January and March provided for the embodiment one of the present application is shown in the following table:

[0041] Figure 9 A partial area enlarged view of Figure 8

[0042] Figure 10 A left high-low first robust regression fitting chart of January and February provided for the embodiment one of the present application is shown in the following table:

[0043] Figure 11 A left high-low first robust regression fitting chart of January and March provided for the embodiment one of the present application is shown in the following table.

[0044] Figure 12 A first correction result chart of January and February provided for the embodiment one of the present application is shown in the following table:

[0045] Figure 13 A first correction result chart of January and March provided for the embodiment one of the present application is shown in the following table:

[0046] Figure 14 A left high-low second robust regression fitting chart of January and February provided for the embodiment one of the present application is shown in the following table:

[0047] Figure 15 A left high-low second robust regression fitting chart of January and March provided for the embodiment one of the present application is shown in the following table:

[0048] Figure 16 A left high-low prediction effect chart provided for the embodiment one of the present application is shown in the following table:

[0049] Figure 17 A left track prediction effect chart provided for the embodiment one of the present application is shown in the following table:

[0050] Figure 18 ​​A track static irregularity trend prediction system structure block diagram corresponding to the method of example one is provided for example two of the present application.

[0051] Figure 19 A hardware structure schematic diagram of an electronic device is provided for example three of the present application.

[0052] Label explanation:

[0053] 101 - inspection module, 1011 - definition unit, 1012 - pairing unit, 1013 - inspection unit, 102 - fitting module, 103 - re-inspection module, 104 - alignment module, 105 - prediction module;

[0054] 20 - bus, 21 - processor, 22 - memory, 23 - communication interface. DETAILED DESCRIPTION

[0055] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is described and explained below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. Based on the examples provided in the present application, all other examples obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0056] In the present application, the phrase "example" means that the specific features, structures or characteristics described in combination with the example can be included in at least one example of the present application. The appearance of this phrase at various places in the specification does not necessarily mean the same example, nor is it an independent or alternative example that is not mutually exclusive with other examples. It is explicitly and implicitly understood by those of ordinary skill in the art that the examples described in the present application can be combined with other examples without conflict.

[0057] The various techniques described in this application can be used in various wireless communication systems such as 2G, 3G, 4G, 5G communication systems and next generation communication systems, for example, a Global System for Mobile communications (GSM), a Code Division Multiple Access (CDMA) system, a Time Division Multiple Access (TDMA) system, a Wideband Code Division Multiple Access (WCDMA) system, a Frequency Division Multiple Access (FDMA) system, an Orthogonal Frequency-Division Multiple Access (OFDMA) system, a Single Carrier-FDMA (SC-FDMA) system, a General Packet Radio Service (GPRS) system, a Long Term Evolution (LTE) system, a 5G New Radio (NR) system, and other such communication systems.

[0058] Under the conditions of "heavy load" and "high speed", once the track is put into use, the deformation and wear of the track will continue to accumulate, and such deformation and wear of the track are inevitable, which will cause vehicle vibration during train operation and affect the safety of people's life and property. In order to ensure the safety of rail transit and improve the quality of train operation, it is necessary to deeply study the track irregularity problem.

[0059] Generally, due to the wheel-rail force between the wheel and the rail, the track irregularity state will deteriorate, and the wheel-rail force is affected by the track load, train speed, track size, etc. In addition to the effect of wheel-rail force on the deterioration of track irregularity state, artificial factors include the number of train operations per unit time (train operation density), track structure design, etc., and natural factors include rainfall, geological structure, air temperature, etc. In order to deal with the safety hazards that may be caused by the deterioration of the track irregularity state, according to the different degrees of deterioration, corresponding lines will be subjected to maintenance operations such as tamping, track shifting, track changing, and screen cleaning. Among the operations for geometric size, from the waveform generated by historical data, it can be clearly found that the track irregularity state has a sudden change.

[0060] Embodiment One

[0061] This embodiment provides a method for predicting the static irregularity trend of a track. Figure 1 This is a flowchart of a method for predicting the static irregularity trend of a track according to an embodiment of this application, as follows: Figure 1 As shown, the process includes the following steps:

[0062] S101, The acquired historical data are subjected to a similarity test using a dynamic time warping algorithm, and the result of the similarity test is represented by a path matrix;

[0063] Firstly, historical data comes from a track inspection instrument, which consists of hardware and software. The hardware is responsible for data acquisition, while the software handles data processing. It can simultaneously measure the track while in motion, measuring track geometry data including horizontal and vertical alignment, gauge, lateral alignment, and triangular irregularities (torsion). Based on these data, irregularities within a 0.125m range can be detected. The track inspection instrument, through its track data processing software, can detect not only the track's internal geometric parameters but also its external geometric parameters, achieving the dual function of controlling line shape and track smoothness. It possesses rich charting and data management functions and can interface with other database software to export measurement data.

[0064] Secondly, the dynamic time warping algorithm can "bend" time series, that is, compress or extend the time domain in an "n-to-1" manner, so that two time series can achieve similarity matching. It can find the point-to-point correspondence between series well, such as for... Figure 2 The phase transition time series shown in the figure has a phase change between the poles A of sequence a and B of sequence b. This phase change can be achieved by finding the offset positions in the two sequences using an "n-to-1" approach. For two given time series P = {p1, p2, ..., pn} and Q = {q1, q2, ..., qm}, the distance matrix D between them is established as shown in the following formula.

[0065]

[0066] The distance between any two points is shown below, where D(i,j) represents the distance between pi and qj, i = 1, 2, ..., n, j = 1, 2, ..., m.

[0067] D(i,j)=||s i -q j ||W

[0068] Suppose there is a path matrix L = {l1, l2, …, lk}, where max(n, m)≤k≤n+m-1, which can represent that the cumulative distance from P to Q reaches the minimum value, and lk represents the position of the element in the distance matrix composed of the corresponding time sequence, which can be expressed as lk = (i, j)k.

[0069] Further, as Figure 3 shown, the specific steps of step S101 include:

[0070] S1011, obtaining a plurality of historical data, taking any one of the plurality of historical data as reference data, and taking other historical data in the plurality of historical data as to-be-verified data.

[0071] S1012, combining and pairing each of the to-be-verified data with the reference data.

[0072] S1013, performing similarity verification on the combined and paired to-be-verified data and reference data by using a dynamic time warping algorithm, and the verification result of the similarity verification is represented by a path matrix.

[0073] In this embodiment, seven static inspection data from January to July of a certain year on the Shanghai-Kunming line are collected as historical data. The relative measurement method is used to set the historical data in January as the reference data, and the six static inspection data from February to July as the to-be-verified data. The left high-low at K442+100-K442+600 is taken as an example. In order to make the results more clearly, the comparison chart of January and February and the comparison chart of January and March are displayed respectively.

[0074] As Figure 4 shown is a comparison chart of the initial data of the left high-low of January and February, in which the x-axis represents 4001 sequence points (the distance between each point is 0.125 meters), and the total distance is 500 meters; the y-axis represents the measured left high-low value, with mm as the unit, the dotted line represents the left high-low value of January, and the solid line represents the left high-low value of February. As Figure 5 shown is a comparison chart of the initial data of January and March, the interval, the meaning of the x-axis, and the meaning of the y-axis are the same as Figure 4 . Table 1 lists the first matching result of the left high-low data of January and February (the first and last 50 of the best distance matrix are selected as the display), and the numbers in the table represent the matching result of the sequence numbers of the two months, for example, the combination of (1, 4) in the first 50 positions represents that the sequence position of January 1 corresponds to the sequence position of February 4. Table 2 is the matching result of the left high-low data of January and March. Therefore, it can be found that a plurality of “one-to-many” phenomena appear, which also embodies the function of the dynamic time warping algorithm in processing variable sequences.

[0075] Table 1: the first matching result of left high-low in January and February

[0076]

[0077] Table 2: the first matching result of left high-low in January and March

[0078]

[0079] As shown in Figure 1 is the first matching path diagram of left high-low in January and February of the complete 4001 best path matrix points, the horizontal coordinate represents the sequence number in January, and the vertical coordinate represents the sequence number in February, the full length of the matching path is 5795, in the matching matrix of mxn dimension, the full length of the matching path is closer to max{m,n}, the better the matching effect. Figure 6 is a partial enlarged view of Figure 7 . Figure 6

[0080] As shown in Figure 2 is the first matching path diagram of left high-low in January and March of the complete 4001 best path matrix points, the horizontal coordinate represents the sequence number in January, and the vertical coordinate represents the sequence number in March, the full length of the matching path is 6251. Figure 8 is a partial enlarged view of Figure 9 . Figure 8

[0081] In summary, in order to promote the subsequent prediction, the dynamic time warping algorithm is used to preprocess the static inspection single data, and how to realize the similarity judgment of two historical data is described in detail, and based on the fact that there are a large number of variable sequences in the single data of the track, the reason for using the dynamic time warping algorithm as the preprocessing of the mileage deviation of the application is explained. In order to better explain the similarity judgment result of the dynamic time warping algorithm, taking the left high-low static inspection data of K442+100~K442+600 of Shanghai-Kunming line as an example, the similarity test of multiple historical data is carried out, and the matching effect diagram of January and February and January and March is selected to show, the full length of the two groups of matching paths is 5795 and 6251 respectively, from the length of the matching path, it can be roughly estimated that the sequence similarity of January and February is better than that of January and March.

[0082] S102, using robust regression method to fit and correct the path matrix to obtain fitting data, and repeating the similarity test on the fitting data until the fitting data reaches the preset condition, and ending the fitting correction to obtain the dynamic warping path matrix.

[0083] ​​The robust regression method is a method of rough error reduction and adjustment. The rough error is observed and eliminated by iteration. In each iteration, one rough error is eliminated, and then the adjustment is recalculated. Then, the rough error is observed and eliminated again. After multiple iterations, all possible rough errors are eliminated. The preset condition is any one of the following: repetition, adjacent repetition, and interval repetition. Repetition means that the actual mileage after the current correction offset (assuming the nth correction) is the same as the actual mileage after the last correction (the (n-1)th correction). Adjacent repetition means that the actual mileage after the (n-1)th correction is the same as the actual mileage after the (n+1)th correction, and the actual mileage after the nth correction is less than the minimum interval unit. Interval repetition means that the actual mileage after the (n-1)th correction is the same as the actual mileage after the (n+1)th correction, and the actual mileage after the nth correction is greater than the minimum interval unit (0.125 m) and less than twice the minimum interval unit (0.25 m). In this case, the actual mileage is selected as the middle position of the two.

[0084] Further, in the fitting process of the robust regression method, a double threshold method is used to optimize the dynamic regular path matrix to form a hexagonal region, thereby saving the calculation amount in the generation process of the dynamic regular path matrix. The range of the path matrix is limited by using the double threshold method, and 83% of the calculation amount in the generation process of the path matrix is saved. For the matching demand of "long period" and "many times", the resource consumption is greatly saved and the work efficiency is improved.

[0085] Specifically, based on the best path calculated by the dynamic time warping algorithm from the 4001 data of the left high low in January and February at K442+100-K42+600 of the Shanghai-Kunming line, a first robust regression fitting is performed as shown in Figure 10 , wherein the bold line represents the best path calculated by the dynamic time warping algorithm, and the thin line represents the fitting curve after the robust regression. Similarly, the fitting curve of the 4001 data of the left high low in January and March at the same railway mileage section is as shown in Figure 11 . Thus, the scatter point model in the best matching path matrix of January and February and the scatter point model in the best matching path matrix of January and March have been fitted. According to the fitting model and the static test data, the mileage offset is calculated by the formula: delta = p / t; in the formula, delta represents the size of the mileage deviation, unit: km; p represents the intercept in the robust regression model, unit: sequence points (or no unit); and t represents how many intervals are divided in unit length, unit: sequence points (or no unit) / km.

[0086] Since the static inspection data is every 0.125 meters, t is 8000, the intercept of the fitting curve of January and February is p2=-1.4543, which corresponds to delta2=-0.0002km, the intercept of the fitting curve of January and March is p3=50.0867, which corresponds to delta3=0.0063km, according to the delta data, the mileage is corrected, the first correction result of January and February is shown in Figure 12 , and the first correction result of January and March is shown in Figure 13 .

[0087] The similarity test of dynamic time warping algorithm and robust regression method is performed again on the January-February data and the January-March data after the first correction, and the results are shown in Figure 14 , Figure 15 . By calculating the mileage deviation delta, it is still necessary to perform the second alignment. The alignment results of the complete Shanghai-Kunming line K442+100-K442+600 from February to July are shown in Table 3. At this point, the seven pieces of left high-low data obtained by static detection from January to July on the Shanghai-Kunming line K442+100-K42+600 are all aligned, and the reference data for alignment is the January data, and the remaining six data from February to July are used as the data to be aligned.

[0088] Table 3: February-July left high-low alignment process

[0089]

[0090] S103, by correlation coefficient method, the necessary mileage in the path matrix and the dynamic warping path matrix is rechecked.

[0091] Specifically, the Pearson correlation coefficient (PCCs) in the correlation coefficient method is used to analyze the alignment effect of six times of data with January as the reference data, and the results are shown in Table 4. According to the results, for most cases, the correlation coefficient of the curve mileage deviation after the dynamic time warping algorithm correction will be closer to 1, that is, the corrected data can achieve better similarity.

[0092] Table 4: February-July alignment effect

[0093] Month Correlation coefficient before correction Correlation coefficient after correction February 0.933808332 0.940479117 March 0.662976581 -0.146107693 April 0.862253587 0.923175299 May 0.878055623 0.899950871 June 0.886405726 0.893340008 July 0.796473471 0.79648917

[0094] S104, when the data of the necessary mileage does not meet the rechecking requirements, the start point resetting method or the scaling method is used to re-fit and correct the necessary mileage, so as to realize the alignment of the several pieces of historical data.

[0095] In the experiment, there are some cases that the Pearson correlation coefficient decreases significantly after matching. In order to ensure the rigor of railway operation, the Pearson correlation coefficient test should be performed on the matched results. If there is an anomaly (such as a significant decrease in the Pearson correlation coefficient), the line should be re-matched.

[0096] Specifically, the abnormality of the Pearson correlation coefficient before and after matching can be re-matched by two ideas: 1. Re-matching by simultaneously translating the line data of the matching group, which requires the inclusion of the key line data segment to be used subsequently. This method is called start point resetting method. 2. Re-matching by simultaneously scaling the length of the data segment, which also requires the preservation of the key data segment. This method is called scaling method. For example: data A and data B are matched from 442+000 to 442+500, and the result shows that the Pearson correlation coefficient is abnormal. It is determined that the data segment to be used subsequently is 442+200 to 442+300. The start point resetting method can try to adjust the starting position of the two data from 442+100 to 442+600 for re-matching. The scaling method can try to match from 442+100 to 442+400 or 441+900 to 442+600.

[0097] S105, training the aligned historical data through a long short-term memory network model with preset hyperparameters to predict future data of the track static irregularity, and the prediction effect is characterized by root mean square error and average deviation.

[0098] The preset hyperparameters refer to the parameters artificially set before training the machine in machine learning. The preset hyperparameters of the embodiment include training ratio, batch size, loss function, optimization function, learning rate, precision requirement, number of hidden layers, and number of hidden layer neurons. The root mean square error represents the fitting degree of the measured value and the true value curve, which is used to measure the accuracy of measurement. The average deviation refers to the arithmetic mean of random errors of all measured values in equal accuracy measurement, which is used to reflect the deviation size of the data in the interval.

[0099] This embodiment uses the left high-low prediction case to predict and analyze future data. Specifically, because the object is determined to be within a few meters, the range of final deterioration trend prediction is limited to 1 meter. The selection of each hyperparameter of this embodiment needs to be analyzed according to the specific situation as follows:

[0100] 1. Determination of training ratio

[0101] Since the interval between every two adjacent data of left height is 0.125 meters, and the minimum unit of the experiment is limited to meters, the proportion of the selected data should be selected as a multiple of 8 to determine the boundary point. A total of three groups of proportions were selected, which were 50%, 75% and 90%. Under the condition that other variables were unchanged, the root mean square error was analyzed, and 0.687411, 0.332209 and 0.769631 were obtained respectively. Therefore, the training proportion of 75% was selected.

[0102] 2. Batch size, loss function, optimization function and learning rate

[0103] The batch data is the length of the sequence input each time, and theoretically the longer the batch data, the more sufficient the application of historical data each time. Combined with the left height data, the first six data are used as historical data input, and the seventh data is used as the prediction value output, so the batch length is set to 6; the commonly used mean square error loss function MSELoss() is selected as the loss function, the Adam() with smaller memory requirement and adaptive learning rate according to different parameters is selected as the optimization function, and the learning rate is selected as 1×10 -3 .

[0104] 3. Precision requirement, number of hidden layers and number of hidden layer neurons

[0105] According to prior knowledge or feasibility in practice, these parameters are arranged, the precision requirement is selected from 1×10 -3 , 1×10 -4 and 1×10 -5 , the number of hidden layers is selected from 64, 128 and 256, and the number of hidden layer neurons is selected from 1 or 2. According to the advantages and disadvantages of the experimental results, the specific parameters are selected, and the experimental results are shown in Table 5. It needs to be explained that due to the small batch data, the results of training under the same conditions also have differences, and the best data in multiple experiments is selected as the comparison in the table. Finally, it is determined that the number of hidden layers is set to 2, the number of hidden layer neurons is set to 128, and the precision requirement is set to 1×10 -3 , the best prediction result is obtained, the corresponding root mean square error value is 0.03948, and the average deviation value is 7.03301. The intuitive effect is shown in Figure 16 .

[0106] Table 5: Comparison of left height experimental results

[0107]

[0108] Of course, other embodiments can also use the left track prediction case to make prediction analysis of future data. Specifically, the left track mileage deviation correction idea, the prediction model construction steps, the hyperparameter selection scheme, etc. are consistent with the left height prediction case, and are not repeated here. The correction process is shown in Table 6.

[0109] Table 6: Left track mileage deviation correction process

[0110]

[0111] The selection criteria of the number of hidden layers, the number of hidden layer neurons, and the accuracy requirement of the parameter selection table are shown in Table 7 according to the experimental results. It should be noted that during the experiment, it was found that the limit of the training accuracy for the left track was all in the order of 1x10 -3 , and this level was still maintained after fine-tuning the learning rate. Therefore, for the left height prediction, the accuracy requirement was uniformly set to 1x10 -3 . Finally, it was determined that the number of hidden layers was 1, the number of hidden layer neurons was 64, and the accuracy requirement was set to 1x10 -3 , which achieved the best prediction results. At this time, the corresponding root mean square error value was 0.00659, and the average deviation value was 0.325645. The intuitive effect is shown in Figure 17 .

[0112] Table 7: Comparison of left track experimental results

[0113]

[0114] In summary, according to the learning of six historical data, the prediction effect of the left height and left track 1 meter data of the static inspection data is shown in Table 8. From Table 8, it can be found that: by selecting the left height and left track two static inspection single indicators as the prediction experiment object, setting the loss function, optimization function, learning rate, etc. through prior experience, optimizing the training proportion, the number of hidden layers, the number of hidden layer neurons, etc. through pre-experiment, the left height data of the final prediction experiment section determines the training proportion 75%, the number of hidden layers 2, the number of hidden layer neurons 128, and the accuracy requirement 1x10 -3 , which achieves the best prediction results of the root mean square error 0.03948 and the average error 7.03301; the left track data determines the training proportion 75%, the number of hidden layers 1, the number of hidden layer neurons 64, and the accuracy requirement 1x10 -3 , which achieves the best prediction results of the root mean square error 0.00659 and the average error 0.325645. Overall, the long short-term memory network model has good prediction effect on the left height and left track in the single indicator of the track meter-level static data.

[0115] Table 8: Comparison of case prediction effects

[0116]

[0117] Embodiment two

[0118] The embodiment provides a structural block diagram of a system corresponding to the method in embodiment one. Figure 18 is a structural block diagram of a track static irregularity trend prediction system according to the embodiment of the application, as shown in the figure, the system comprises: Figure 18 a verification module 101, configured to perform similarity verification on a plurality of pieces of historical data by using a dynamic time warping algorithm, and to represent a result of the similarity verification by a path matrix;

[0119] a fitting module 102, configured to perform fitting correction on the path matrix by using a robust regression method to obtain fitting data, and to repeatedly perform the similarity verification on the fitting data until the fitting data reaches a preset condition, and then to end the fitting correction to obtain a dynamic warping path matrix;

[0120] a re-inspection module 103, configured to re-inspect necessary mileages in the path matrix and the dynamic warping path matrix by using a correlation coefficient method;

[0121] an alignment module 104, configured to perform re-fitting correction on the necessary mileages by using a starting point resetting method or a scaling method when data of the necessary mileages does not meet re-inspection requirements, so as to realize alignment of the plurality of pieces of historical data;

[0122] a prediction module 105, configured to train the plurality of pieces of aligned historical data by using a long short-term memory network model with preset hyperparameters, so as to predict future data of track static irregularities, and a prediction effect is represented by a root mean square error and an average deviation.

[0123] Further, the verification module 101 further comprises:

[0124] a definition unit 1011, configured to obtain a plurality of pieces of historical data, take any one of the plurality of pieces of historical data as reference data, and take other historical data in the plurality of pieces of historical data as to-be-verified data;

[0125] a pairing unit 1012, configured to pair each piece of the to-be-verified data with the reference data one by one;

[0126] a verification unit 1013, configured to perform similarity verification on the to-be-verified data and the reference data which are combined and paired by using a dynamic time warping algorithm, and a verification result of the similarity verification is represented by a path matrix.

[0127]

[0128] ​It should be noted that the above various modules can be functional modules or program modules, which can be implemented by software or hardware. For the modules implemented by hardware, the above various modules can be located in the same processor; or the above various modules can also be located in different processors in any combination.

[0129] Embodiment three

[0130] In combination Figure 1 The track static irregularity trend prediction method of the application described can be implemented by an electronic device. Figure 19 The hardware structure schematic diagram of the electronic device according to the embodiment of the application.

[0131] The electronic device can include a processor 21 and a memory 22 storing computer program instructions.

[0132] Specifically, the processor 21 can include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the application.

[0133] The memory 22 can include mass storage for data or instructions. By way of example, and not limitation, the memory 22 can include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, a USB drive, a magneto-optical disk, optical disks, such as CDs and DVDs, magnetic tapes, or other types of storage media, which can be removable or non-removable (or fixed). In some embodiments, the memory 22 is non-volatile memory. In some embodiments, the memory 22 includes read-only memory (ROM) and random-access memory (RAM). The ROM can be mask- programmed ROM, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), electrically alterable ROM (EAROM), or flash memory, or a combination of two or more of these. The RAM can be static RAM (SRAM) or dynamic RAM (DRAM), which can be Fast Page Mode DRAM (FPM DRAM), Extended Data Output DRAM (EDO DRAM), synchronous DRAM (SDRAM), etc.

[0134] The memory 22 can be used to store or buffer various data files required for processing and / or communication, and possible computer program instructions executed by the processor 22.

[0135] The processor 21 realizes the track static irregularity trend prediction method of the above-mentioned embodiment one by reading and executing the computer program instructions stored in the memory 22.

[0136] In some embodiments, the electronic device can further include a communication interface 23 and a bus 20. In which, as shown in the figure, the processor 21, the memory 22, the communication interface 23 are connected through the bus 20 and complete the communication between each other. Figure 19

[0137] The communication interface 23 is used to realize the communication between each module, device, unit and / or equipment in the embodiments of the present application. The communication interface 23 can also realize the data communication between other components, such as: external equipment, image / data acquisition equipment, database, external storage and image / data processing workstation, etc.

[0138] ​Bus 20 includes hardware, software, or both, to couple components of the electronic device to each other and to couple components of the electronic device to other components, including those within the electronic device and outside the electronic device. While bus 20 is shown for the sake of clarity as a single bus, bus 20 can include one or more buses operating together. Bus 20 can include, without limitation, a data bus, an address bus, a control bus, an expansion bus, a local bus, etc. In some embodiments, bus 20 can include a Graphics Accelerated Peripherals (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or another suitable bus or combination of two or more of these. Where appropriate, bus 20 can include one or more buses. Although this application describes and shows a particular bus, this application contemplates any suitable bus or interconnect.

[0139] The electronic device can acquire the track static irregularity trend prediction system, and perform the track static irregularity trend prediction method of the embodiment one of the application.

[0140] In addition, in combination with the track static irregularity trend prediction method of the above-mentioned embodiment one, the embodiment of the application can provide a storage medium for implementation. The storage medium has computer program instructions stored thereon; the computer program instructions are executed by a processor to implement the track static irregularity trend prediction method of the above-mentioned embodiment one.

[0141] The above merely describes preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement and improvement within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method of predicting the trend of track static irregularities, characterized in that, The method comprises the following steps: The similarity of the obtained historical data is tested by using a dynamic time warping algorithm, and the results of the similarity test are represented by a path matrix; The path matrix is fitted and corrected by using a robust regression method to obtain fitting data, and the fitting data is repeatedly subjected to the similarity test until the fitting data meets a preset condition, and then the fitting correction is ended to obtain a dynamic warping path matrix; The necessary mileage in the path matrix and the dynamic warping path matrix is rechecked by using a correlation coefficient method; When the data of the necessary mileage does not meet the rechecking requirement, the necessary mileage is re-fitted and corrected by using a starting point resetting method or a scaling method, so as to align the historical data; The aligned historical data is trained by using a long short-term memory network model with preset hyperparameters, so as to predict future data of track static irregularities, and the prediction effect is represented by a root mean square error and an average deviation; In the fitting process of the robust regression method, a double-threshold method is used to make symmetric optimization of the dynamic warping path matrix, form a hexagonal region, and save the calculation amount in the generation process of the dynamic warping path matrix; The preset condition is any one of recurrence, adjacent recurrence and interval recurrence.

2. The rail static irregularity trend prediction method of claim 1, wherein, The step of testing the similarity of the obtained historical data by using a dynamic time warping algorithm, and representing the results of the similarity test by a path matrix comprises the following steps: Obtain historical data, and take any one of the historical data as reference data and the other historical data as test data; Combine and pair each test data with the reference data; Test the similarity of the combined and paired test data and reference data by using a dynamic time warping algorithm, and represent the test results by a path matrix.

3. The rail static irregularity trend prediction method of claim 1, wherein, The preset hyperparameters include at least one of a training ratio, a batch size, a loss function, an optimization function, a learning rate, a precision requirement, a number of hidden layers and a number of hidden layer neurons, or a combination of several thereof.

4. The rail static irregularity trend prediction method of claim 1, wherein, The root mean square error represents the fitting degree of the measured value and the true value curve, and is used to measure the accuracy of the measurement; the average deviation is the arithmetic mean of the random errors of all measured values in the equal-precision measurement, and is used to reflect the deviation of the data in the interval.

5. A track static irregularity trend prediction system employing the track static irregularity trend prediction method as claimed in claim 1, characterized by, The method comprises the following steps: The test module is configured to test the similarity of the obtained historical data by using a dynamic time warping algorithm, and represent the results of the similarity test by a path matrix; The fitting module is configured to fit and correct the path matrix by using a robust regression method to obtain fitting data, repeatedly test the fitting data, and end the fitting correction when the fitting data meets a preset condition to obtain a dynamic warping path matrix; The rechecking module is configured to recheck the necessary mileage in the path matrix and the dynamic warping path matrix by using a correlation coefficient method; The re-fitting module is configured to re-fit and correct the necessary mileage by using a starting point resetting method or a scaling method when the data of the necessary mileage does not meet the rechecking requirement. An alignment module is configured to perform re-fitting correction on the necessary mileage by using a starting point resetting method or a scaling method when data of the necessary mileage does not meet the re-checking requirement, so as to realize alignment of the historical data. A prediction module is configured to train the aligned historical data by using a long short-term memory network model with preset hyperparameters, so as to predict future data of the track static irregularity, and a prediction effect is characterized by a root mean square error and an average deviation.

6. The rail static irregularity trend prediction system of claim 5, wherein, The inspection module further includes: A definition unit is configured to obtain historical data, take any one of the historical data as reference data, and take other historical data as to-be-inspected data; A pairing unit is configured to pair each to-be-inspected data with the reference data; An inspection unit is configured to perform similarity inspection on the to-be-inspected data and the reference data by using a dynamic time warping algorithm, and an inspection result of the similarity inspection is characterized by a path matrix.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the track static irregularity trend prediction method in any one of claims 1 to 4.

8. A storage medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the track static irregularity trend prediction method in any one of claims 1 to 4.

Images