CN114386707A - A method and device for predicting track irregularity - Google Patents

Abstract

The present disclosure relates to the field of transportation, and in particular, to a method and an apparatus for predicting track unevenness. The method comprises the steps of obtaining vehicle operation data, wherein the vehicle operation data comprise vehicle acceleration and vehicle speed; inputting the vehicle operation data into a pre-established irregularity prediction model to obtain irregularity data, wherein the irregularity prediction model comprises an attention network, a convolution neural network and a circulation neural network, the attention network is used for determining attention weight of the vehicle operation data, and the convolution neural network is used for determining waveform characteristics of the vehicle operation data with the attention weight; and the cyclic neural network is used for determining the rugged data according to the waveform characteristics. According to the scheme, the track long wave height and the medium wave height of different railway lines and different under-rail foundations can be estimated in real time, the change of the track state is sensed in time, powerful support is provided for adjusting track equipment and ensuring the safe operation of a train, and the safety of the train operation is improved.

Description

A method and device for predicting track irregularity

technical field

This paper relates to the field of transportation and can be used in the field of high-speed railways, especially a method, device, computer equipment and storage medium for predicting track irregularity.

Background technique

At present, high-speed trains have become people's daily travel mode. The safe operation of trains is the basic requirement and necessary guarantee for railway transportation. Regular inspection of trains and tracks is particularly important in transportation. When the train passes through the relatively uneven track section, various parts of the vehicle will generate corresponding vibration, which will affect the driving safety and passenger experience. Therefore, it is increasingly important to stably monitor the real-time geometry of railway track lines and to rectify the geometric smoothness of the lines.

In the prior art, the measurement of the geometric irregularity of the track mainly relies on the comprehensive detection train or the track inspection vehicle for regular detection. However, the detection method of rail inspection vehicles can only be carried out in a fixed time outside the vehicle operation time, and the detection frequency of rail inspection vehicles is insufficient, which makes it difficult to effectively obtain the deterioration characteristics and development laws of track geometry; Mutational diseases can only be detected at the time of testing. The Kalman filter method is difficult to establish and invert complex nonlinear dynamic models, and it lacks practicability.

Aiming at the problems of low efficiency of track irregularity test and complicated inversion at present, a method and device for predicting track irregularity are required.

SUMMARY OF THE INVENTION

In order to solve the above problems in the prior art, the embodiments herein provide a method, device, computer equipment and storage medium for predicting track irregularity, which solve the problems in the prior art.

The embodiments herein provide a method for predicting track irregularity, including: acquiring vehicle operation data, the vehicle operation data including vehicle acceleration and vehicle speed; inputting the vehicle operation data into a pre-established roughness prediction model, Acquiring roughness data, the roughness prediction model includes an attention network, a convolutional neural network and a recurrent neural network, the attentional network is used to determine the attention weight of the vehicle operation data, and the convolutional neural network used to determine the waveform characteristics of vehicle operation data with attention weights; the cyclic neural network is used to determine the roughness data according to the waveform characteristics.

According to an aspect of the embodiments herein, the training process of the roughness prediction model includes: determining a sample data set of the vehicle operation data, the sample data set including the historical acceleration of the vehicle, the historical speed of the vehicle and the corresponding historical roughness data; according to the sample data set, train the parameters in the unevenness prediction model.

According to an aspect of the embodiments herein, the training of parameters in the roughness prediction model according to the sample data set includes: determining the initial value of an attention network, a convolutional neural network, and a recurrent neural network in the roughness prediction model. parameters; input the sample data set into the roughness prediction model to obtain the roughness forecast data; construct a loss function according to the roughness forecast data and the historical roughness data; use the loss function to train The parameters of each network in the pre-established unevenness prediction model.

According to an aspect of the embodiments herein, after acquiring the vehicle operation data, the method further includes: performing data preprocessing on the vehicle operation data, wherein the data preprocessing at least includes: trend item filtering, data standardization, and reverse order processing; The inputting of the vehicle operation data into the pre-established roughness prediction model further includes inputting the vehicle operation data obtained after preprocessing into the pre-established roughness prediction model.

According to an aspect of the embodiments herein, performing reverse order processing on the vehicle operation data includes: setting vehicle operation data of every adjacent L+1 time interval as a data group; and performing reverse order processing on each data group.

According to one aspect of the embodiments herein, performing trend term filtering on the vehicle operation data includes utilizing least squares fitting, wavelet decomposition, convex optimization, smooth prior, variational modal decomposition, Fourier transform At least one method determines trend items of the vehicle operating data and filters out the trend items.

According to an aspect of the embodiments herein, after training the parameters in the roughness prediction model according to the sample data set, the method further includes: identifying the vehicle running data in the test sample by using the roughness prediction model for which the parameters have been determined; Calculate the evaluation index according to the prediction result of the vehicle operation data in the test sample, wherein the evaluation index is at least one of mean absolute error, root mean square error, Hill inequality coefficient, and correlation coefficient; determine whether the evaluation index is The preset conditions are met, and if not, the structure of the unevenness prediction model is adjusted.

The embodiments herein also provide an apparatus for predicting track irregularity, including: a vehicle operation data acquisition module, configured to acquire vehicle operation data, where the vehicle operation data includes vehicle acceleration and vehicle speed;

The determination module is used to input the vehicle operation data into a pre-established roughness prediction model to obtain the roughness data, and the roughness prediction model includes an attention network, a convolutional neural network and a recurrent neural network, so The attention network is used to determine the attention weight of the vehicle operation data, and the convolutional neural network is used to determine the waveform characteristics of the vehicle operation data of the attention weight; the recurrent neural network is used to, according to the waveform characteristics, Determine the jagged data.

The embodiments herein also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the above method when executing the computer program.

The embodiments herein also provide a computer-readable storage medium on which computer instructions are stored, and when the computer instructions are executed by a processor, implement the above-mentioned method.

The method and device for predicting track irregularity in this scheme can real-time predict the long-wave and medium-wave heights of different railway lines and foundations under different tracks, and have good prediction effects on bridge sections and foundation sections. Effectively predict the periodic irregularity caused by simply supported beam span, track slab length, etc., so as to quickly know the track irregularity state in real time, provide safety support for adjusting track equipment and train operation, save labor and material costs, and improve operation efficiency. Improve the safety of train operation.

Description of drawings

In order to more clearly illustrate the embodiments or the technical solutions in the prior art, the following briefly introduces the accompanying drawings that are used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only For some embodiments herein, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative effort.

FIG. 1 is a flowchart of a method for predicting track irregularity according to an embodiment of this paper;

FIG. 2 is a flowchart of a training method for a roughness prediction model according to an embodiment of this paper;

3 is a flowchart of a method for training a roughness prediction model according to an embodiment of this paper;

FIG. 4 is a flow chart of a method for adjusting the structure of an unevenness prediction model according to an embodiment of this paper;

FIG. 5 is a schematic structural diagram of an attention network according to an embodiment of this paper;

FIG. 6 is a schematic structural diagram of a convolutional neural network according to an embodiment of this paper;

FIG. 7 is a schematic structural diagram of a threshold cycle unit according to an embodiment of this document;

FIG. 8 is a schematic structural diagram of a device for predicting unevenness according to an embodiment of the present disclosure;

FIG. 9 is a schematic diagram showing the specific structure of the device for predicting the unevenness according to the embodiment of this paper;

FIG. 10 is a schematic structural diagram of a roughness prediction model according to an embodiment of this paper;

FIG. 11 is a schematic structural diagram of a computer device according to an embodiment of this document.

Description of the symbols in the drawings:

801. A vehicle operation data acquisition unit;

8011. A vehicle acceleration acquisition module;

8012. The vehicle speed acquisition module;

802. A unit for obtaining data for prediction of unevenness;

8021. The training module of the unevenness prediction model;

8022. Data preprocessing module;

8023. Model adjustment module;

1102. Computer equipment;

1104. processor;

1106. memory;

1108. Drive mechanism;

1110. Input/output module;

1112. Input device;

1114. Output device;

1116. Presentation equipment;

1118. Graphical user interface;

1120. network interface;

1122. Communication link;

1124. A communication bus.

Detailed ways

In order to make those skilled in the art better understand the technical solutions in this specification, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings in the embodiments. Obviously, the described implementation Examples are only some of the embodiments herein, but not all of the embodiments. Based on the embodiments herein, all other embodiments obtained by persons of ordinary skill in the art without creative efforts shall fall within the scope of protection herein.

It should be noted that the terms "first", "second" and the like in the description and claims herein and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific sequence or sequence. It is to be understood that data so used may be interchanged under appropriate circumstances such that the embodiments herein described can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having", and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, apparatus, product or device comprising a series of steps or units is not necessarily limited to those expressly listed Rather, those steps or units may include other steps or units not expressly listed or inherent to these processes, methods, products or devices.

This specification provides method operation steps as described in the embodiments or flow charts, but more or less operation steps may be included based on routine or non-creative work. The sequence of steps enumerated in the embodiments is only one of the execution sequences of many steps, and does not represent the only execution sequence. When an actual system or device product is executed, the methods shown in the embodiments or the accompanying drawings may be executed sequentially or in parallel.

It should be noted that the track irregularity prediction method in this paper can be used in the field of transportation, and the application field of the track irregularity prediction method and device is not limited in this paper.

FIG. 1 is a flowchart of a method for predicting track irregularity according to an embodiment of this paper, which specifically includes the following steps:

Step 101: Obtain vehicle operation data, where the vehicle operation data includes vehicle acceleration and vehicle speed. In this step, the vehicle operation data includes vehicle acceleration and vehicle operation speed. The vehicle acceleration includes vehicle lateral acceleration (LVBA, Lateral Vehicle Body Acceleration) and vehicle vertical acceleration (VVBA, Vertical Vehicle Body Acceleration). Among them, the vehicle acceleration and vehicle speed can be regarded as the spatial sequence data corresponding to the mileage sequence data, and are related to the track geometric irregularity. In this step, vehicle operation data may be acquired through signal devices or sensors installed on the vehicle or trackside (eg, gyroscopes or accelerometers installed on the vehicle bogie). In addition, vehicle operation data can be collected at certain mileage intervals or sampling intervals. For example, vehicle acceleration and vehicle speed are obtained at a sampling interval of 0.25 m/time. The vehicle running data obtained in this step corresponds to the mileage data one-to-one. For example, if a line is 10 kilometers long, 40,000 pieces of vehicle operation data (including 40,000 pieces of vehicle acceleration and 40,000 pieces of vehicle speed) can be obtained according to the sampling interval of 0.25m/time. The 40,000 pieces of vehicle operation data and mileage data One-to-one correspondence. That is, each mileage data corresponds to one vehicle acceleration and vehicle running speed. Thus, each mileage data and each corresponding vehicle running data (including vehicle speed and vehicle acceleration) can form an array. The mileage data and vehicle operation data on the 10-kilometer line can form 40,000 array sequences.

Step 102: Input the vehicle operation data into a pre-established roughness prediction model to obtain the roughness data, wherein the roughness prediction model includes an attention network, a convolutional neural network and a recurrent neural network, so The attention network is used to determine the attention weight of the vehicle operation data, and the convolutional neural network is used to determine the waveform characteristics of the vehicle operation data with the attention weight; the recurrent neural network is used to determine the waveform characteristics according to the waveform characteristics Uneven data.

In this step, the roughness prediction model is used to predict the roughness data of the input vehicle operation data. The input vehicle operation data may include vehicle operation data collected in real time on the spot when the rail inspection vehicle is running. In some embodiments of this specification, the unevenness refers to the vertical unevenness of the top surface of the rail along the extension direction. Track irregularity is a random process that changes with mileage, and the amplitude and wavelength of track irregularity at different locations may not be the same. In general, the higher the driving speed, the greater the wavelength range of track irregularities. The roughness can be divided into short wave roughness, medium wave roughness and long wave roughness according to the wavelength. Since the excitation frequency caused by short-wave irregularities is high and has little effect on vehicle acceleration, this application ignores the impact of track irregularities with smaller wavelengths on vehicle acceleration, and this application mainly focuses on long-wave irregularities and medium-wave irregularities , so the long-wave unevenness and the medium-wave unevenness are used as examples to discuss. In some embodiments of the present specification, the long-wave roughness is set as the roughness with a wavelength between 70 meters and 120 meters, and the medium-wave roughness can be the unevenness with a wavelength between 42 meters and 70 meters. smooth. Specifically, the wavelength ranges of the long-wave roughness and the medium-wave roughness can be adjusted according to the change of the train running speed in the specific application scenario, which is not limited in this application.

In some embodiments of this specification, the roughness prediction model includes an attention network, a convolutional neural network, and a recurrent neural network. Among them, the attention network is used to input vehicle operation data, so that the model can automatically learn the input data that contributes more to the output unevenness. The back end of the attention network is a combination of convolutional neural network and recurrent neural network, which are used for Learn the waveform characteristics and hidden sequence information and trend information of vehicle operation data. Wherein, the recurrent neural network includes but does not participate in one of Gated Recurrent Units (GRU, Gated Recurrent Units), Long Short-Term Memory (LSTM, LongShort-Term Memory), etc. or any combination thereof. The structure of the unevenness prediction model can be seen in Figure 10.

FIG. 2 is a flowchart of a training method for a roughness prediction model according to an embodiment of this document. Specifically include:

Step 201: Determine a sample data set of vehicle operation data, where the sample data set includes historical vehicle acceleration, historical vehicle speed and corresponding historical high and low roughness data.

In this step, the historical acceleration of the vehicle in the sample data set is the vehicle acceleration obtained in the historical track detection test, the historical vehicle speed is the vehicle speed obtained in the historical track detection test, and the historical high and low roughness data is the historical vehicle speed in the historical test, The historical high and low roughness data corresponding to the historical acceleration of the vehicle. For example, the sample data set is the vehicle speed, vehicle acceleration and corresponding high and low unevenness data obtained in a 10-kilometer detection section of the Beijing-Shanghai line in December 2021. Among them, the sample data in the sample data set can come from one or more times of training and testing based on different routes and different tracks.

Step 202, according to the sample data set, train parameters in the unevenness prediction model. In this step, the parameters of the attention network, convolutional neural network and recurrent neural network in the roughness prediction model are trained according to the historical acceleration of the vehicle, the historical speed of the vehicle and the corresponding historical roughness data to complete the roughness prediction Model training. In some embodiments of this specification, attention networks include but are not limited to attention mechanisms and self-attention mechanisms, convolutional neural networks include but are not limited to CNN neural networks, and recurrent neural networks include but are not limited to gated recurrent units (GRU, Gated Recurrent Unit) and LSTM Long Short-Term Memory (LSTM, Long Short-Term Memory). Detailed descriptions of the attention network, convolutional neural network, and recurrent neural network in this step are shown in Figures 5 to 7.

FIG. 3 shows a flowchart of a method for training an unevenness prediction model according to an embodiment of this paper, which specifically includes:

Step 301: Determine the initial parameters of the attention network, the convolutional neural network, and the recurrent neural network in the unevenness prediction model. In this step, conventional attention network, convolutional neural network and recurrent neural network can be used to form an initial roughness prediction model. The initial parameters of the attention network, the convolutional neural network and the cyclic neural network in the unevenness prediction model are the parameters initially specified for the attention network, the convolutional neural network and the cyclic neural network respectively. Specific parameters include but are not limited to: network structure, number of network layers, weights and offsets of each layer of network, etc. For example, the parameters of the attention network in the unevenness prediction model are: the number of layers is 2, the weight of the first layer is W1, and the weight of the second layer is W2; for another example, the convolution in the unevenness prediction model The parameters of the neural network are: used to be 3 layers, the offset of the first layer is b1, the offset of the second layer is b2, and the offset of the third layer is b3. The initial parameters of each model in the roughness prediction model in this application may have other arbitrary deformations, which are not limited in this application.

Step 302: Input the sample data set into the roughness prediction model to obtain the roughness prediction data. Input the historical vehicle acceleration, vehicle speed and corresponding historical roughness data in the sample data set into the roughness prediction model, and the predicted roughness data can be obtained one-to-one with each input sample data. The roughness data obtained by the prediction may have a certain gap with the actual roughness data. Therefore, it is necessary to further train the roughness prediction model.

Step 303 , constructing a loss function according to the prediction data of unevenness and the historical data of unevenness. In this step, the loss function represents the difference between the predicted unevenness data and the actual unevenness data obtained in the history. A loss function is constructed according to the difference between the predicted unevenness data and the actual unevenness data obtained in history. In some embodiments of this specification, the loss function includes, but is not limited to, a mean square error loss function, a cross entropy loss function, an exponential loss function, and the like.

Step 304, using the loss function to train the parameters of each network in the pre-established roughness prediction model. The loss function is used to calculate the difference between the predicted roughness data obtained in step 302 and the predicted data of historical irregularities. When the difference does not meet the preset difference range, continuously adjust the parameters of each network in the unevenness prediction model, and further continue to train the unevenness prediction model until the value of the loss function is less than a certain range, and determine the trained level. Roughness prediction model. Specifically, in the model training process, when the difference of the loss function does not meet the predetermined difference range, the loss function is minimized by continuously optimizing the weights and offsets between each layer of each network in the model. In some embodiments, hyperparameter optimization can also be performed on the trained model. Specifically, the hyperparameters may include parameters such as the learning rate, the number of iterations, and the batch size. Among them, the learning rate refers to the magnitude of updating the network weights in the optimization algorithm, the number of iterations refers to the number of times the entire training set samples are input into the neural network for training, and the batch size refers to the number of samples that are fed into the model each time the neural network is trained.

In some embodiments of this specification, after acquiring the vehicle operation data, the method further includes: performing data preprocessing on the vehicle operation data, wherein the data preprocessing at least includes: trend item filtering, data normalization, and reverse order processing ; inputting the vehicle operation data into the pre-established roughness prediction model and further inputting the vehicle operation data obtained after preprocessing into the pre-established roughness prediction model.

In some embodiments of this specification, in order to improve the convergence speed of the roughness prediction model, in order to improve the convergence speed of the roughness prediction model, the maximum-minimum standardization can be used for the phenomenon that different vehicle operation data are quite different, and the vehicle operation data in the sample data set can be adjusted according to A certain relationship is constrained to be in the range 0 to 1. Specifically, it can be processed by the following formula:

where x is the vehicle running data in the sample data set, x _norm is the standardized data, and x _max and x _min are the maximum and minimum values of the vehicle running data in the sample data set, respectively.

In some embodiments of this specification, performing reverse order processing on the vehicle operation data includes: setting the vehicle operation data of every adjacent L+1 time interval as a data group; and performing reverse order processing on each data group. In some embodiments of the present specification, track roughness and vehicle acceleration are not synchronized in the time domain. The acceleration of the vehicle depends on the track roughness at a certain moment and the track roughness at a past moment. Therefore, reversing the order of vehicle operation data (eg, vehicle acceleration) in the time domain can capture the causal relationship between input and output.

其中，可以设定当i为1时，X¹可以表示车辆垂向加速度，当i为2时，X²可以表示车辆横向加速，当i为3时，X³可以表示车辆速度，t表示当前时刻，L表示预先设定的时间间隔。其中，每一个输入向量中包括多个数组，每一个数组中包含L+1个时刻对应的数据。对应的，输出数据由Y表示，

j表示向量Y的2个维度，分别为中波高低不平顺和长波高低不平顺。将输入的车辆横向加速度、车辆垂向加速度和车辆速度数据与输出的高低不平顺数据按照一定步长进行滑动，可以预测出整条线路的高低不平顺。例如，设置L为5。则输入向量为车辆垂向加速度：

表示车辆垂向加速度中包括多个数组。其中第一个数组为某一t时刻的车辆垂向加速度、该t时刻后的第1个时刻的垂向加速度……、该t时刻后的第5个车辆垂向加速度。其中第二个数组为某一t+1时刻的车辆垂向加速度、该t+1时刻后的第1个时刻的车辆垂向加速度……该t+1时刻后的第5个车辆垂向加速度。输入向量为车辆横向加速度：

表示车辆横向加速度中包括某一时刻t的车辆垂向加速度、该t时刻后的第1个时刻的垂向加速度……、该t时刻后的第5个车辆垂向加速度。该时刻在本申请中，L的具体数值可以预先由系统设定，也可以根据线路情况、预测情况实时调整。Specifically, the input data is represented by a vector X. The input data are vehicle vertical acceleration, vehicle lateral acceleration and vehicle speed, respectively

Among them, it can be set that when i is 1, X ¹ can represent the vertical acceleration of the vehicle, when i is 2, X ² can represent the lateral acceleration of the vehicle, when i is 3, X ³ can represent the vehicle speed, and t represents the current Time, L represents a preset time interval. Among them, each input vector includes multiple arrays, and each array includes data corresponding to L+1 moments. Correspondingly, the output data is represented by Y,

j represents the two dimensions of the vector Y, which are medium wave roughness and long wave roughness. By sliding the input vehicle lateral acceleration, vehicle vertical acceleration and vehicle speed data with the output unevenness data according to a certain step size, the unevenness of the entire line can be predicted. For example, set L to 5. Then the input vector is the vertical acceleration of the vehicle:

Indicates that multiple arrays are included in the vertical acceleration of the vehicle. The first array is the vertical acceleration of the vehicle at a certain time t, the vertical acceleration of the first time after the time t..., the vertical acceleration of the fifth vehicle after the time t. The second array is the vertical acceleration of the vehicle at a certain time t+1, the vertical acceleration of the vehicle at the first time after the time t+1...the fifth vertical acceleration of the vehicle after the time t+1 . The input vector is the lateral acceleration of the vehicle:

Indicates that the lateral acceleration of the vehicle includes the vertical acceleration of the vehicle at a certain time t, the vertical acceleration of the first time after the time t, and the fifth vertical acceleration of the vehicle after the time t. In this application, the specific value of L can be set in advance by the system, and can also be adjusted in real time according to the line conditions and forecast conditions.

As described above, the mileage data is processed into a mileage sequence, and each vehicle vertical acceleration, vehicle lateral acceleration and vehicle speed on the route can be in a one-to-one correspondence with the mileage sequence. The vehicle operation data within a certain length range and the roughness prediction data predicted by the roughness prediction model are composed of multiple sequence data, which can be used as the roughness of the entire line. According to the mileage sequence, the amplitude of the high and low irregularity output corresponding to the vehicle running data collected every 0.25 meters along the line is connected into a waveform curve, and the waveform of the track irregularity curve can be obtained.

In some embodiments of the present specification, performing trend term filtering on the vehicle operating data includes utilizing at least one of least squares fitting, wavelet decomposition, convex optimization, smooth prior, variational modal decomposition, Fourier transform A method determines a trending item of the vehicle operating data; filtering out the trending item. In some embodiments of this specification, the vehicle operation data collected in step 101 may deviate from the baseline due to unstable low frequency performance outside the vehicle sensor frequency range and interference from the surrounding environment of the sensor. Further, the degree to which the collected vehicle operation data deviates from the baseline may vary over time. In some embodiments of the present specification, it is necessary to filter out the part that affects the correctness of the collected data in the process that the vehicle operation data deviates from the baseline and changes over time, that is, trend item filtering. The signals collected in the time domain can be further purified to obtain more accurate vehicle operation data. In some embodiments of the present specification, at least one of least squares fitting, wavelet decomposition, convex optimization, smooth prior, variational modal decomposition, and Fourier transform may be used for the trend term of the vehicle operating data Perform filtering to obtain the purified vehicle speed, vehicle lateral acceleration and vehicle vertical acceleration.

FIG. 4 is a flow chart of a method for adjusting the structure of a prediction model for unevenness according to an embodiment of this document. Specifically include the following steps:

Step 401 , identifying the vehicle running data in the test sample by using the roughness prediction model whose parameters have been determined. In this step, the test sample is the data sample selected when the model is actually applied, and the test sample includes the vehicle test speed, the vehicle test lateral acceleration and the vehicle test vertical acceleration. Using the trained roughness prediction model to identify the vehicle running data in the test sample, the roughness prediction result corresponding to the test sample can be obtained.

Step 402: Calculate the evaluation index according to the test result of the vehicle operation data in the test sample. Wherein, the evaluation index is at least one of mean absolute error, root mean square error, Hill inequality coefficient, and correlation coefficient. In this step, the test result obtained after inputting the test sample into the vehicle operation data is the roughness data corresponding to the test sample. In this step, one or more evaluation indicators are used to evaluate the accuracy of the uneven data output by the model. For example, if the roughness prediction model outputs roughness data, the mean absolute error and the root mean square error can be used to judge the correct rate of the output roughness data. That is, the ratio of the number of correctly classified data to the total number of output data.

Step 403, judging whether the evaluation index satisfies the preset condition, if not, adjusting the structure in the unevenness prediction model. In this step, corresponding to one or more evaluation indicators, there are one or more preset conditions in this step. For example, the preset condition is that the correct rate of the classification result is greater than a certain threshold or the error rate of the classification result is less than or equal to a certain threshold, or the like. When the evaluation index meets the threshold condition of the preset accuracy rate, the unevenness prediction model is not adjusted; when the evaluation index does not meet the threshold condition of the preset accuracy rate, the attention network and convolutional neural network in the unevenness prediction model are adjusted. The network structure or parameters of a network or recurrent neural network.

FIG. 5 is a schematic structural diagram of an attention network according to an embodiment of this paper. The attention network learns and assigns different weights to the input data according to the importance or contribution of the input vehicle operation data to the output data, pays attention to the feature dimension of the input data, and acts on the feature dimension of the input. In some embodiments of this specification, it can be known from historical experiments that the correlation between the vertical acceleration of the vehicle and the roughness is the greatest, or the vertical acceleration of the vehicle has the greatest importance on the roughness. Therefore, the attention network is learned and trained. , we will focus on the parameter of lateral acceleration of the vehicle that contributes the most to the roughness.

It is assumed that vectors X ¹ , X ² , and X ³ are used to represent the vehicle vertical acceleration, vehicle lateral acceleration and vehicle speed in the vehicle operation data, respectively. Then layer 1 attention:

α ₁ =[α ₁₁ ,α ₁₂ ,α ₁₃ ]=Softmax(XW ₁ ^T )

s ₁ =[s ₁₁ ,s ₁₂ ,s ₁₃ ]=Multiply(Xα ₁ ^T )=[X ¹ α ₁₁ ,X ² α ₁₂ ,X ³ α ₁₃ ]

The second layer of attention is:

α ₂ =[α ₂₁ ,α ₂₂ ,α ₂₃ ]=Softmax(s ₁ W ₂ ^T )

s ₂ =[s ₂₁ ,s ₂₂ ,s ₂₃ ]=Multiply(s ₁ α ₂ ^T )=[X ¹ α ₁₁ α ₂₁ ,X ² α ₁₂ α ₂₂ ,X ³ α ₁₃ α ₂₃ ]

In the above formulas: W ₁ , W ₂ represent the weight matrix to be learned by the attention network; α ₁ , α ₂ represent the attention weight vector automatically learned by the attention network; Multiply represents the multiplication of the corresponding elements of the matrix or vector; s ₁ , s ₂ represents the output of the Multiply operation. In this application, the number of layers and the specific structure of the attention machine network can be any other arbitrary number, which is not limited in this application.

FIG. 6 is a schematic structural diagram of a convolutional neural network according to an embodiment of this document. As shown in the figure, the convolutional neural network mainly includes a one-dimensional convolution layer and a one-dimensional pooling layer. The convolutional neural network takes the output of the attention network in Figure 5 [X ¹ α ₁₁ X ₂₁ , X ² α ₁₂ α ₂₂ , X ³ α ₁₃ α ₂₃ ] as input, so that the convolutional neural network can effectively extract the vehicle acceleration and Waveform characteristics of vehicle speed. Among them, the waveform features are contained in the local vehicle acceleration and vehicle speed waveforms, and are related to the track irregularity. Since the vibration of high-speed vehicles is usually sensitive to the multi-wavelength components of track irregularities, stacked convolutional layers are used to balance and extract the waveform features of vehicle acceleration. In some embodiments of this specification, the convolutional neural network CNN is set as a 2-layer convolutional layer (Conv1D), the number of convolution kernels is 16 and 32, the size of the convolution kernel is 1×5, the stride is 1, and the number of convolution kernels is 0. filling. Among them, the output of the convolutional layer adopts the Tanh activation function to increase the nonlinear expression ability of the network. In the pooling layer (MaxPoolin1D), a 2-layer max-pooling alternating with the convolutional layers is adopted, the pool size is 1×2, and the stride is 2. feature and reduce the output dimension size. After two consecutive alternating convolution and maximum pooling operations, the correlation between the data is mined and the noise and unstable components are removed from it, and the relatively stable information after processing is passed into the recurrent neural network as a whole to further learn sequence features and predict. Out of the uneven.

FIG. 7 is a schematic structural diagram of a threshold cycle unit according to an embodiment of the present disclosure. Among them, the information propagation of the GRU network is represented by the following formula:

r _t =σ(W _rx x _t +W _rh h _t-1 )

z _t =σ(W _zx x _t +W _zh h _t-1 )

为点乘，σ为sigmoid函数，tanh为双曲正切函数，r_t为控制重置的门控，即为重置门；z_t为控制更新的门控，即为更新门；h_t为当前时刻的输出值，包含上一个时刻的输出值h_t-1和记忆了当前时刻的状态值

其中，σ函数可以将数据变换为0-1之间的数值，从而充当门控信号；tanh激活函数将数据变换为-1至1之间的数值，W_rx表示输入值x_t与重置门之间的权重；W_rh表示重置门与上一个时刻的输出值h_t-1之间的权重；W_zx表示输入值x_t与更新门之间的权重；W_zh表示更新门与上一个时刻的输出值h_t-1之间的权重；

表示输入值x_t与记忆了当前时刻的状态值之间的权重；

表示记忆了当前时刻的状态值与上一个时刻的输出值h_t-1之间的权重。Among them, x _t is the input value at the current moment, W is the weight matrix,

is the point product, σ is the sigmoid function, tanh is the hyperbolic tangent function, r _t is the gate that controls the reset, which is the reset gate; z _t is the gate that controls the update, which is the update gate; h _t is the current gate The output value of the moment, including the output value h _t-1 of the previous moment and the state value of the current moment memorized

Among them, the σ function can transform the data into a value between 0-1, thus acting as a gate signal; the tanh activation function transforms the data into a value between -1 and 1, and W _rx represents the input value x _t and the reset gate W _rh represents the weight between the reset gate and the output value h _t-1 at the previous moment; W _zx represents the weight between the input value x _t and the update gate; W _zh represents the update gate and the previous The weight between the output value h _t-1 at the moment;

Represents the weight between the input value x _t and the state value that has memorized the current moment;

Indicates that the weight between the state value at the current moment and the output value h _t-1 at the previous moment is memorized.

The input of the threshold loop unit is the input x _t at the current moment and the output value h _t at the previous moment. The threshold loop control unit mainly includes update gate and reset gate. The update gate z _t is used to control the degree to which the state information of the previous moment is brought into the current state. The larger the value of the update gate is, the more the state information of the previous moment is brought in. The reset gate _rt is used to control the extent to which state information from the previous moment is ignored. Since the stacked gated recurrent control units can learn higher-level sequence features, single-layer or multi-layered gated recurrent control networks can be built. Among them, each layer of threshold loop control unit has a certain number of neurons. In some embodiments of this specification, a layer 2 GRU (as shown in Figure 7) is established. The number of GRU neurons in the second layer is 64 and 128 respectively. Activation functions in GRU include but are not limited to: Tanh, ReLu, Leaky ReLu, etc. In addition, in order to prevent the GRU model from overfitting, a random dropout (Dropout) method can be used for each layer of GRU, and an appropriate ratio of Dropout can be set, for example, the ratio of Dropout is set to 0.2. In some embodiments of the present specification, a fully connected layer is set after the threshold loop control unit, and finally the track irregularity data is output, that is, the track irregularity sequence corresponding to the mileage data one-to-one.

FIG. 8 is a schematic structural diagram of an apparatus for predicting roughness according to the embodiment of this paper. The basic structure of the apparatus for predicting roughness is described in this figure, and the functional units and modules therein can be implemented by software, or by A general-purpose chip or a specific chip is implemented, and some or all of the functional units and modules can be implemented on static detection and dynamic detection hardware, or a part of them can also be implemented on static detection and dynamic detection hardware. Specifically include:

a vehicle operation data acquisition unit 801, configured to acquire vehicle operation data, the vehicle operation data including vehicle acceleration and vehicle speed;

The roughness prediction data acquisition unit 802 is used for inputting the vehicle operation data into a pre-established roughness prediction model to obtain the roughness data, and the roughness prediction model includes an attention network, a convolutional neural network and a recurrent neural network, the attention network is used to determine the attention weight of the vehicle operation data, and the convolutional neural network is used to determine the waveform characteristics of the vehicle operation data with the attention weight; The waveform features determine the roughness data.

This scheme can predict in real time the heights of long-wave and medium-wave of different railway lines and foundations under different tracks, and has a good prediction effect on bridge sections and foundation sections. Effectively predict the periodic irregularity caused by simply supported beam span, track slab length, etc., so as to quickly know the track irregularity state in real time, provide safety support for adjusting track equipment and train operation, save labor and material costs, and improve operation efficiency. Improve the safety of train operation.

As an embodiment of this document, reference may also be made to FIG. 9 , which is a schematic diagram of a specific structure of the apparatus for predicting unevenness according to the embodiment of this document.

As an embodiment of this document, the vehicle operation data acquisition unit 801 further includes: acquiring vehicle acceleration and vehicle speed. Therefore, the vehicle operation data acquisition unit 801 further includes:

The vehicle acceleration acquisition module 8011 is used to acquire the vehicle lateral acceleration and the vehicle vertical acceleration;

The vehicle speed obtaining module 8012 is used to obtain the running speed of the vehicle.

As an embodiment of this document, the roughness prediction data acquisition unit 802 further includes: training a roughness prediction model, performing data preprocessing on vehicle operation data, and adjusting the model according to the prediction results and evaluation indicators. Therefore, the roughness prediction data acquisition unit 802 further includes:

The roughness prediction model training module 8021 is used to train the roughness prediction model according to the sample data set;

a data preprocessing module 8022, configured to perform data preprocessing on the vehicle operation data;

The model adjustment module 8023 is configured to adjust the structure of the unevenness prediction model according to the prediction result and the evaluation index output by the model.

As shown in FIG. 11 , for a computer device provided by the embodiments herein, the computer device 1102 may include one or more processors 1104 , such as one or more central processing units (CPUs), each processing unit may implement One or more hardware threads. The computer device 1102 may also include any memory 1106 for storing any kind of information such as code, settings, data, and the like. Without limitation, for example, memory 1106 may include any one or a combination of the following: any type of RAM, any type of ROM, flash memory devices, hard disks, optical disks, and the like. More generally, any memory can use any technology to store information. Further, any memory can provide volatile or non-volatile retention of information. Further, any memory may represent a fixed or removable component of computer device 1102. In one instance, when the processor 1104 executes the associated instructions stored in any memory or combination of memories, the computer device 1102 may perform any operation of the associated instructions. The computer device 1102 also includes one or more drive mechanisms 1108 for interacting with any memory, such as a hard disk drive mechanism, an optical disk drive mechanism, and the like.

Computer device 1102 may also include an input/output module 1110 (I/O) for receiving various inputs (via input device 1112 ) and for providing various outputs (via output device 1114 ). A specific output mechanism may include presentation device 1116 and associated graphical user interface (GUI) 1118 . In other embodiments, the input/output module 1110 (I/O), the input device 1112 and the output device 1114 may not be included, and only serve as a computer device in the network. Computer device 1102 may also include one or more network interfaces 1120 for exchanging data with other devices via one or more communication links 1122 . One or more communication buses 1124 couple together the components described above.

Communication link 1122 may be implemented in any manner, eg, through a local area network, a wide area network (eg, the Internet), a point-to-point connection, etc., or any combination thereof. Communication links 1122 may include any combination of hardwired links, wireless links, routers, gateway functions, name servers, etc. governed by any protocol or combination of protocols.

Corresponding to the methods in FIG. 1 to FIG. 4 , the embodiments herein also provide a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and the computer program is executed by a processor to execute the steps of the above method. .

Embodiments herein also provide computer-readable instructions, wherein when a processor executes the instructions, the program therein causes the processor to perform the methods shown in FIGS. 1 to 4 .

It should be understood that, in the various embodiments herein, the size of the sequence numbers of the above-mentioned processes does not imply the sequence of execution, and the execution sequence of each process should be determined by its functions and internal logic, rather than the implementation of the embodiments herein The process constitutes any qualification.

It should also be understood that, in the embodiments herein, the term "and/or" is only an association relationship for describing associated objects, indicating that there may be three kinds of relationships. For example, A and/or B can mean that A exists alone, A and B exist at the same time, and B exists alone. In addition, the character "/" in this document generally indicates that the related objects are an "or" relationship.

Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two, in order to clearly illustrate the differences between hardware and software Interchangeability, the above description has generally described the components and steps of each example in terms of function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of this document.

Those skilled in the art can clearly understand that, for the convenience and brevity of description, for the specific working process of the above-described systems, devices and units, reference may be made to the corresponding processes in the foregoing method embodiments, and details are not repeated here.

In the several embodiments provided herein, it should be understood that the disclosed system, apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or Can be integrated into another system, or some features can be ignored, or not implemented. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may also be electrical, mechanical or other forms of connection.

The units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solutions in the embodiments herein.

In addition, each functional unit in each of the embodiments herein may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

The integrated unit, if implemented in the form of a software functional unit and sold or used as an independent product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solutions in this article are essentially or part of contributions to the prior art, or all or part of the technical solutions can be embodied in the form of software products, and the computer software products are stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments herein. The aforementioned storage medium includes: U disk, removable hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

The principles and implementations of this paper are described by using specific examples in this paper, and the descriptions of the above examples are only used to help understand the methods and core ideas of this paper; , there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to this article.

Claims (10)

1. A method for predicting track irregularities, wherein the method comprises: acquiring vehicle operation data, where the vehicle operation data includes vehicle acceleration and vehicle speed; inputting the vehicle operation data into a pre-established roughness prediction model to obtain the roughness data; Wherein, the unevenness prediction model includes an attention network, a convolutional neural network and a recurrent neural network; the attention network is used to determine the attention weight of the vehicle operation data; The convolutional neural network is used to determine the waveform characteristics of the vehicle operation data with the attention weight; The cyclic neural network is used for determining high and low unevenness data according to the waveform characteristics. 2. The track irregularity prediction method according to claim 1, wherein the training process of the irregularity prediction model comprises: determining a sample data set of the vehicle operation data, the sample data set including the historical acceleration of the vehicle, the historical speed of the vehicle and the corresponding historical high and low roughness data; According to the sample data set, the parameters in the roughness prediction model are trained. 3. The track irregularity prediction method according to claim 2, wherein the parameters in the training irregularity prediction model according to the sample data set include: Determine the initial parameters of the attention network, the convolutional neural network and the recurrent neural network in the unevenness prediction model; inputting the sample data set into the roughness prediction model to obtain the roughness prediction data; constructing a loss function according to the roughness prediction data and the historical roughness data; The parameters of each network in the pre-established roughness prediction model are trained by using the loss function. 4. The method for predicting track irregularity according to claim 1, characterized in that, after acquiring the vehicle operation data, further comprising: Data preprocessing is performed on the vehicle operation data, wherein the data preprocessing at least includes: trend item filtering, data standardization, and reverse order processing. 5. The method for predicting track irregularity according to claim 4, wherein performing reverse order processing on the vehicle operation data comprises: Set the vehicle operation data of every adjacent L+1 time interval as a data group; Process each data set in reverse order. 6. The method for predicting track irregularity according to claim 4, wherein performing trend item filtering on the vehicle operation data comprises: Determine the trend term of the vehicle operation data by using at least one method of least square fitting, wavelet decomposition, convex optimization, smooth prior, variational modal decomposition, and Fourier transform; The trending item is filtered out. 7. The track irregularity prediction method according to claim 2, wherein, according to the sample data set, after training the parameters in the irregularity prediction model, the method further comprises: Identify the vehicle running data in the test sample by using the parameterized roughness prediction model; Calculate the evaluation index according to the prediction result of the vehicle operation data in the test sample, wherein the evaluation index is at least one of mean absolute error, root mean square error, Hill inequality coefficient, and correlation coefficient; It is judged whether the evaluation index satisfies the preset condition, and if not, the unevenness prediction model is adjusted. 8. A device for predicting track irregularities, wherein the device comprises: a vehicle operation data acquisition unit, configured to acquire vehicle operation data, the vehicle operation data including vehicle acceleration and vehicle speed; a roughness prediction data acquisition unit, configured to input the vehicle operation data into a pre-established roughness prediction model to obtain the roughness data; Wherein, the unevenness prediction model includes an attention network, a convolutional neural network and a recurrent neural network; the attention network is used to determine the attention weight of the vehicle operation data; The convolutional neural network is used to determine the waveform characteristics of the vehicle operation data with the attention weight; The cyclic neural network is used for determining high and low unevenness data according to the waveform characteristics. 9. A computer device comprising a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor implements any of claims 1-7 when the processor executes the computer program one of the methods described. 10. A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, implements the method of any one of claims 1-7.

Images