Method for indirectly identifying track irregularity based on acceleration response of commercial running train

Abstract

The invention relates to a method for indirectly identifying track irregularity based on acceleration response of a commercial running train, and belongs to the field of railway track state monitoring. The method comprises the following steps: constructing a VBI system model, and constructing a BiLSTM-AM model; inputting different types of track irregularities into the VBI system model as excitation to obtain train acceleration response; training the BiLSTM-AM model by taking the train acceleration response as the input of the model and the corresponding track irregularity as the output of the model; and inputting acceleration data acquired by an acceleration sensor of a commercial running train into the BiLSTM-AM model, and predicting the irregularity of a train running track. The method has excellent performance in both the spatial domain and the frequency domain, can predict different types of track irregularity, and does not need a professional test train and a large number of sensors.

Description

Technical Field

[0001] The invention belongs to the field of railway track status monitoring and relates to a method for indirectly identifying track irregularities based on the acceleration response of a commercial train. Background Art

[0002] Track irregularity, as the most critical excitation source of the vehicle-bridge interaction (VBI) system, has a significant impact on train safety and passenger comfort. The surface roughness of the rail deteriorates under the influence of wear of track components caused by long-term wheel-rail interaction, roadbed and subgrade deformation. Nowadays, due to the development trend of light-weight and high-speed trains, the requirements for railway track smoothness are more stringent. Therefore, it is very necessary to measure and maintain the railway regularly to ensure the safe operation of trains.

[0003] So far, some methods have been proposed to measure track irregularity in time. At present, the detection methods of track irregularity are divided into static detection technology and dynamic detection technology. Static detection technology obtains track irregularity by directly measuring the spatial geometric relationship of the track. Since static detection technology cannot consider the inertial effect of the real train, it will induce large errors in measurement. Dynamic detection technology obtains track irregularity indirectly from the dynamic response of the vehicle, including high-speed integrated track recording vehicles and in-service vehicles. High-speed integrated track recording vehicles widely use laser cameras and inertial reference devices to detect dynamic track irregularity, and provide real-time feedback of track geometry through image processing technology. Famous high-speed integrated track recording vehicles in the world include NMT in the UK, CRH380A-001 in China, Doctor Yellow in Japan, ARCHIMEDE in Italy, IRIS320 in France, etc. Although high-speed integrated track recording vehicles have some obvious advantages in identifying track irregularity, track inspection vehicles are expensive and will interfere with the operation of other vehicles during work, thus limiting their high-frequency application in track inspection.

[0004] Therefore, track anomaly detection methods based on on-track trains have attracted much attention. Vehicle acceleration data is collected by installing accelerometers on the axle or bogie, and then the track irregularities are retrieved through inverse analysis techniques. For example, the document "Estimation of long wavelength track irregularities from on board measurement. 2008 4th IET International Conference on Railway Condition Monitoring, 2008: 1-6" estimates the track long wavelength irregularities in the frequency domain under the acceleration of in-service trains through a model-based identification procedure. The document "Track geometry estimation from car-body vibration. Vehicle System Dynamics, 2014, 52 (sup1): 207-219" proposes a Kalman filter method that indirectly identifies track irregularities using vehicle body acceleration. The document "Urban rail track condition monitoring based on in-service vehicle acceleration measurements. Measurement, 2016, 80: 217-228" measures the acceleration of the bogie and car body floor in an in-service subway, and identifies track irregularities using the state equation of the vertical suspension system.

[0005] Although the method of combining in-service trains has been successfully applied to identify track irregularities, most theories ignore bridges in their track models due to the complex vehicle-bridge interaction effect. In fact, bridges account for a large proportion of railways, especially high-speed railways. Therefore, developing accurate and effective methods to identify track irregularities on bridges has great engineering value. The paper "Scanning of Bridge Surface Roughness from Two-Axle Vehicle Response by EKF-UI and Contact Residual: Theoretical Study. Sensors, 2022" uses the Kalman filter algorithm to identify track irregularities on bridges using two-axle vehicle responses. The paper "Simultaneous identification of the frequencies and track irregularities of high-speed railway bridges from vehicle vibration data. Mechanical Systems and Signal Processing, 2021, 152: 107412" combines vehicle dynamic response with the Kalman filter algorithm to achieve track irregularity identification on railway bridges. However, considering the uncertainty of structural parameters and the random speed of in-service trains, the identification accuracy of track irregularities needs to be further improved.

[0006] Deep learning has been shown to be an effective method for mapping nonlinear relationships of complete data sets and has been widely used for real-time prediction of bridge and vehicle responses. However, research on identifying bridge track irregularities based on deep learning models has only been reported in limited literature. The paper "Reconstruction of road defects and road roughness classification using artificial neural networks simulation and vehicle dynamic responses: Application to experimental data. Journal of Terramechanics, 2014, 53: 1-18" uses vehicle acceleration responses to identify different categories of road roughness through a NARX network. The paper "Black-box modelling of nonlinear railway vehicle dynamics for track geometry assessment using neural networks. Vehicle System Dynamics, 2019, 57 (9): 1241-1270" develops a black-box model for predicting the dynamic response of railway vehicles and track irregularities. The paper "Real-time prediction of dynamic irregularity and acceleration of HSR bridges using modified LSGAN and in-service train. Smart Structures and Systems, 2023, 31(5): 501-516" proposed an improved LSGAN method for predicting the dynamic response of high-speed railway bridges and identifying track irregularities. The paper "Trackvertical irregularity estimation for railway bridges using a novel and lightweight deep-learning architecture. Vehicle System Dynamics: 2024, 1-28" developed a lightweight deep learning architecture for reconstructing track irregularities in railway bridges using the dynamic response of the bogie. Although the model can accurately predict track irregularities on bridges, the structural parameters are treated as constants and the random characteristics are ignored. Summary of the invention

[0007] In view of this, the object of the present invention is to provide a method for indirectly identifying track irregularities based on the acceleration response of commercial trains, combining an attention mechanism (AM) and a BiLSTM model to indirectly identify track irregularities through the acceleration response of in-service trains.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] A method for indirectly identifying track irregularity based on the acceleration response of a commercial train, the method comprising:

[0010] S1. Build VBI system model and BiLSTM-AM model;

[0011] The BiLSTM-AM model includes an input layer, a BiLSTM layer, an AM layer, a fully connected layer and an output layer connected in sequence. In the model, different weights are dynamically assigned to the data output from the BiLSTM layer to the fully connected layer at each moment through the AM layer.

[0012] S2, input different types of track irregularities as excitations into the VBI system model to obtain the train acceleration response;

[0013] S3, using the train acceleration response as the input of the BiLSTM-AM model and the corresponding track irregularity as the output of the BiLSTM-AM model to train the BiLSTM-AM model;

[0014] S4. The acceleration data of the train during on-track operation is collected through the acceleration sensor of the commercial train, and the acceleration data is input into the trained BiLSTM-AM model to obtain the unevenness of the train track, thereby realizing indirect identification of track unevenness based on the acceleration response of the commercial train.

[0015] Furthermore, in step S3, the training process of the BiLSTM-AM model includes:

[0016] S31, preparing a data set, which is a train acceleration response generated by a VBI system model under different types of track irregularity excitations;

[0017] S32, dividing the data set into a training set and a test set, using the training set to train the BiLSTM-AM model, and the loss function in the iterative training is MAE;

[0018] S33. After training, the prediction accuracy of the BiLSTM-AM model was tested using the test set, using RMSE, MAE, and R 2 To evaluate the prediction accuracy of the BiLSTM-AM model.

[0019] Furthermore, the acceleration data collected by the acceleration sensor of the commercial train includes one or more of the body vertical acceleration, body rotational acceleration, front bogie vertical acceleration, front bogie rotational acceleration, rear bogie vertical acceleration and rear bogie rotational acceleration.

[0020] Furthermore, when predicting track irregularity through commercial train acceleration data, at least one type of acceleration data is input into the BiLSTM-AM model.

[0021] Furthermore, the constructed VBI system model is expressed as:

[0022]

[0023] Where M, C, and K represent the mass, damping, and stiffness matrices, respectively; Z(t) represents acceleration, velocity and displacement vector respectively; F(t) represents load vector.

[0024] The beneficial effects of the present invention are:

[0025] (1) The present invention combines BiLSTM and AM. AM emphasizes the importance of key information that has a significant impact on the recognition results by assigning different weights to the input vehicle acceleration, enabling the model to focus on information that is strongly related to track irregularities, thereby improving the accuracy and robustness of the model.

[0026] (1) The proposed BiLSTM-AM model has excellent performance in both spatial and frequency domains and can predict different types of track irregularities, including the German low-interference and high-interference spectra, and the Chinese high-speed rail spectrum.

[0027] (2) The BiLSTM-AM model proposed in the present invention can accurately identify track unevenness by only providing the vertical acceleration of the vehicle body, without the need for professional test trains and a large number of sensors.

[0028] (3) When the running train speed is 350 km / h and the bridge damping ratio is 0.05, the BiLSTM-AM model proposed in the present invention is also highly suitable for indirect identification of track irregularities.

[0029] (4) When the train acceleration is contaminated by noise levels as high as 20%, the BiLSTM-AM model proposed in the present invention can still accurately identify track irregularities and has good noise resistance.

[0030] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0032] Figure 1 It is a schematic diagram of the vehicle-bridge system model structure;

[0033] Figure 2 It is a schematic diagram of the LSTM storage unit structure;

[0034] Figure 3 Schematic diagram of BiLSTM network structure;

[0035] Figure 4 It is a schematic diagram of the AM structure;

[0036] Figure 5 Schematic diagram of the recognition process of the BiLSTM-AM model;

[0037] Figure 6 Three spatial domain samples for simulating track irregularities;

[0038] Figure 7 For target and simulation comparison of track irregularity, Figure 7 (a) is the low-interference spectrum in Germany, Figure 7 (b) is the high interference spectrum in Germany. Figure 7 (c) is the spectrum of China’s high-speed rail;

[0039] Figure 8 The training and test data loss curves of the BiLSTM-AM model;

[0040] Fig. 9 The comparison results of the simulated and predicted values ​​of three types of track irregularities are shown in Figure 2. Fig. 9 (a) is the low-interference spectrum in Germany, Fig. 9 (b) is the high interference spectrum in Germany. Fig. 9 (c) is the spectrum of China’s high-speed rail;

[0041] Fig.10 Comparison of acceleration response results of track irregularity under different input numbers. Fig.10 (a) is the amplitude, Fig.10 (b) is the residual;

[0042] Fig.11For the comparison of recognition results at different vehicle speeds, Fig.11 (a) is 207 km / h, Fig.11 (b) is 272 km / h, Fig.11 (c) 349 km / h;

[0043] Fig.12 The results of different track irregularity noise levels are compared. Fig.12 (a) is the amplitude, Fig.12 (b) is the residual;

[0044] Fig.13 The results of track irregularity under different bridge damping ratios are compared. Fig.13 (a) is the amplitude, Fig.13 (b) is the residual. DETAILED DESCRIPTION

[0045] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0046] An embodiment of the present invention proposes a BiLSTM model combined with an attention mechanism, which indirectly identifies track irregularities through the acceleration response of commercial trains. In this model, the BiLSTM model is used to capture the two-directional dependencies between input and output data, and the attention mechanism is used to improve the focusing ability of the BiLSTM model and reduce the interference of redundant information on the prediction results.

[0047] The method proposed in this embodiment is as follows:

[0048] 1. Establish a physical model of the vehicle-bridge system and use it to generate a dataset to train the proposed proxy model.

[0049] In this embodiment, it is assumed that the wheelset is in rigid contact with the rails and the bridge is simulated as an Euler-Bernoulli beam. The mechanical VBI model is as follows: Figure 1As shown. It can be seen that the train is modeled as a multi-rigid body system consisting of a car body (denoted as "c"), two bogies (denoted as "b") and four wheelsets (denoted as "w"), each rigid body having vertical and rotational degrees of freedom (denoted as "z, θ" respectively). The two suspension subsystems simulated by linear spring-damper elements connect the adjacent rigid bodies, where the spring stiffness and damping coefficient are denoted as "k" and "c" respectively. The bogie and wheelset are connected through the primary suspension (denoted by "p"), and the car body and bogie are connected through the secondary suspension (denoted by "s"). In addition, m and J represent the mass and mass moment of inertia, respectively; l c is the semi-longitudinal distance between the front bogie and the rear bogie; l r Indicates the half distance of the bogie; l g 、E g I g , are the length, bending stiffness and mass per unit length of the bridge respectively.

[0050] Figure 1 The equation for the VBI system shown can be expressed as:

[0051]

[0052]

[0053] In the formula, M, C, and K are mass, damping, and stiffness matrices, respectively; Z(t) represents the acceleration, velocity and displacement vector respectively; F(t) represents the load vector. In the formula, the parameter subscripts "v" and "b" represent the train and bridge respectively.

[0054] 2. Combine the attention mechanism and BiLSTM to build the BiLSTM-AM model.

[0055] (1) BiLSTM model

[0056] LSTM is a neural network that processes sequence data and is widely used in tasks such as time series prediction, speech recognition, and natural language processing. LSTM automatically controls the flow of information through a gating mechanism, solving the gradient vanishing or exploding problem of RNN. Compared with traditional RNN, LSTM has better tracking performance in processing long-term dependencies. The LSTM storage unit structure is as follows: Figure 2 As shown in the figure, it can be seen that the memory cell includes three gates: forget gate, input gate and output gate. These three gates can jointly control state update, information preservation, and the final transmission of memory cells.

[0057] The information transfer formula of LSTM storage unit is as follows:

[0058] Forget Gate:

[0059] r j =σ(w r [h j-1 ,X j ]+b r ) (4)

[0060] Among them, w r and b r Respectively represent the weight and bias of the forget gate; σ represents the sigmoid function, which combines the current input X j and the moment before j-1 The hidden layer state produces a value between 0 and 1 to determine which information should be retained.

[0061] Input Gate:

[0062] i j =σ(w i [h j-1 ,X j ]+b i ) (5)

[0063]

[0064] Among them, I j represents the output of the input gate; w i and b i Represent the weight and bias of the input gate respectively; represents the candidate state of the current step calculated by the tanh(·) function. Then, the forgotten part calculated by integration and new information To update the current memory unit.

[0065] Output Gate:

[0066] o j =σ(w o [h j-1 ,X j ]+b o ) (8)

[0067]

[0068] Among them, j is the output of the output gate; w o and b o Represent the weight and bias of the output gate respectively; the current hidden state h j By tanh(c j ) and element o j Multiply updates.

[0069] BiLSTM is an improved LSTM model that connects the forward LSTM and backward LSTM networks to capture data dependencies in two directions. The structure of the BiLSTM network is as follows: Figure 3 shown.

[0070] (2) Attention Mechanism

[0071] In order to further improve the ability of LSTM to capture data features, the attention mechanism is used to dynamically assign different weights to each moment of LSTM input, where the AM structure is as follows: Figure 4 The specific calculation process is summarized as follows: First, define the trainable weights as Where h represents the size of the BiLSTM hidden layer. Secondly, the weight vector w and the hidden layer output vector (denoted as [y 1 ,y 2 ,…,y v ]) to calculate the attention score α, which is shown as:

[0072]

[0073] Use the softmax function to calculate the attention score α for each time step i Perform normalization operation, expressed as:

[0074]

[0075] Next, the BiLSTM outputs of each time step of the hidden layer are weighted summed to get the final output, as shown below:

[0076]

[0077] Then, a fully connected layer is added to map the final output, where W fc and b represent the weight and bias of the fully connected layer, respectively, expressed as:

[0078] y finally =W fc c+b(13)

[0079] (3) BiLSTM-AM model and its recognition process

[0080] The recognition process of the BiLSTM-AM model proposed in this embodiment is as follows: Figure 5As shown. It can be seen that the recognition process mainly includes two main steps: generating a data set, training and evaluating the model. Among them, in order to obtain the data set, a physical VBI model considering the random damping parameters of the bridge and the random train speed is established, and the train acceleration is used as the input, and the track irregularity (track irregularity) is obtained as the corresponding output. The data set is divided into a training set and a test set. Using the obtained samples, the BiLSTM-AM model is established and trained.

[0081] The BiLSTM-AM proposed in this embodiment includes five layers, namely, input layer, BiLSTM layer, attention mechanism (AM) layer, fully connected (FC) layer and output layer. The unknown train acceleration from the test set is input into the proposed model to evaluate the prediction performance of the BiLSTM-AM model. During the test, the typical indicators such as root mean square error (RMSE), mean absolute error (MAE) and R square (R 2 ) is used to evaluate the prediction accuracy of the BiLSTM-AM model. The smaller the RMSE and MAE values ​​are, the smaller the error of the model in identifying track irregularities is. 2 The closer the value is to 1, the higher the correlation between the model identification and the true value. The expression of the above indicator is as follows:

[0082]

[0083] Among them, y i is the basic truth value; represents the predicted value; represents the average value of the ground truth; m represents the length of track irregularity.

[0084] 3. Simulation verification based on the constructed VBI system and BiLSTM-AM model

[0085] (1) Characteristics of vehicles and bridges

[0086] Considering the layout of bridges in actual railway projects, this embodiment uses 11 simply supported beam bridges with a length of 32m and 8 vehicles to estimate the BiLSTM-AM model. Among them, the main parameters of the VBI system are shown in Table 1, and Rayleigh damping is used to consider the influence of bridge damping. In order to consider the influence of random structural parameters and vehicle speed, the damping ratio and vehicle speed are regarded as random parameters that obey the uniform distribution of 0.01~0.05 and 200~350km / h respectively. In addition, the train runs 40m on the roadbed in advance to ensure that the train is stable before going on the bridge. The Newmark method is used to solve the motion equation step by step, and the time step is set to 0.01s.

[0087] Table 1

[0088]

[0089] (2) Track irregularities

[0090] The generation of track irregularity is the first step to obtain the train response dataset. Due to the randomness of track irregularity, three different types of track irregularity spectra include the German low-interference spectrum and high-interference spectrum, as well as the Chinese high-speed rail spectrum. To this end, the trigonometric series method is used to simulate the track irregularity samples, which is defined as follows:

[0091]

[0092] ΔΩ=(Ω u -Ω l ) / N (19)

[0093] Where r(x) is the vertical track irregularity at position x; Ω i is the spatial frequency; ΔΩ represents the frequency increment; θ i Represents a random phase angle between 0 and 2π; Ω u and Ω l are the upper and lower limits of the frequency; N represents the number of discrete frequencies, S V (Ω i ) is the power spectral density (PSD) function. The function S of the German vertical interference orbit spectrum GV (Ω) is expressed as:

[0094]

[0095] Among them, A v is the surface roughness constant; Ω r ,Ω c and Ω s is the cut-off frequency. The specific spectrum parameters of the German interference orbit are shown in Table 2:

[0096] Table 2

[0097]

[0098] Function S of China's vertical high-speed rail spectrum CV (Ω) is expressed as follows:

[0099]

[0100] In the formula, A, B, C, D, E, and F are 0.1270, -2.1531, 1.5503, 4.9835, 1.3891, -0.0327, and 0.0018 respectively.

[0101] Figure 6 and Figure 7The simulated spatial domain samples and their PSD spectra of three different track irregularities are shown respectively. It can be seen that the frequency domain results of simulated track irregularities are in good agreement with the target power spectrum, indicating that the spatial domain results of simulated track irregularities are reliable.

[0102] (3) Parameter setting of BiLSTM-AM model

[0103] In the BiLSTM-AM model, the train acceleration dataset is divided into a training set and a test set, with 80% allocated to the training set and 20% allocated to the test set. Before training the network, all datasets are normalized to [0,1] to ensure that each feature has the same scale. The main hyperparameters of the proposed model are shown in Table 3:

[0104] Table 3

[0105]

[0106] Figure 8 The loss curves of the training and test data of the BiLSTM-AM model are depicted. It can be seen that the loss curves of both drop rapidly at the beginning of training and then gradually tend to zero. This means that the proposed BiLSTM-AM model is close to the true value during training and also performs well on unseen test data.

[0107] (4) Comparison of prediction results

[0108] The BiLSTM-AM model is used to predict various types of track irregularities. The input of the model is the train acceleration response under several different types of track irregularities (German low-disturbance spectrum and high-disturbance spectrum, and Chinese high-speed rail spectrum). In addition, the number of samples has a significant impact on the prediction effectiveness of the deep learning network.

[0109] The prediction performance of the BiLSTM-AM model using different sample numbers is shown in Table 4:

[0110] Table 4

[0111]

[0112] It should be noted that the sample size in Table 4 is the train acceleration sample size for a certain type of track irregularity spectrum, which means that the total number of samples is three times the value listed in Table 4. In addition, the evaluation indicator is the average result of 100 validation sets. It can be seen from Table 4 that the prediction accuracy of the BiLSTM-AM model increases with the increase of sample size. For the case where the sample size is greater than 1500, the model achieves satisfactory accuracy in predicting all three types of track irregularities. Therefore, in order to balance accuracy and efficiency, a total of 6000 samples were used when training and testing the BiLSTM-AM model.

[0113] Three different types of track irregularity predictions are randomly selected from the validation set to show the accuracy of the BiLSTM-AM model under different train speeds and bridge damping ratios. Fig. 9 The figure shows the comparison between the simulated and predicted values ​​of the three different types of track irregularities. It can be seen that except for some local extreme points with drastic changes, the simulated and predicted values ​​of all three types of track irregularities on the bridge are well matched in the spatial domain. The RMSE values ​​of the German low and high interference spectrum and the Chinese high-speed rail spectrum are 4.08E-04, 5.28E-04 and 1.73E-04, respectively, and the corresponding R 2 The values ​​are 0.9798, 0.9843, and 0.9855, respectively. In addition, in the short wavelength range, the PSD of the predicted track irregularity is slightly lower than the simulated value, which shows that although the proposed BiLSTM-AM model performs well in identifying track irregularity using train acceleration, there are still some challenges in predicting the high-frequency components of track irregularity.

[0114] 4. Impact of different parameters on the BiLSTM-AM model

[0115] (1) Influence of acceleration response input number

[0116] The number of train acceleration responses as the input of the BiLSTM-AM model is crucial for identifying track irregularities. In actual engineering, the train acceleration response is generally obtained by accelerometers. In theory, the acceleration response of the train should be large enough to contain more comprehensive track irregularity information, but a large number of accelerometers face many challenges in field measurements. To this end, this embodiment studies the impact of the number of acceleration inputs on the accuracy of track irregularity identification, and defines four distribution positions of different accelerometers, as shown in Table 5.

[0117] Table 5

[0118]

[0119] Based on the BiLSTM-AM model, the recognition results and their residuals are shown in Figure 2 under different acceleration response input numbers. Fig.10As shown in , where the residual is defined as the difference between the predicted value and the true value. It can be seen that the track irregularity identified under different numbers of accelerations is in good agreement with the true value, and the identification residual decreases slightly with the increase in the number of acceleration inputs. This shows that the proposed BiLSTM-AM model has a good ability to identify track irregularity, even when only one vertical acceleration is input. In addition, Table 6 lists the evaluation indicators of the recognition performance of the proposed BiLSTM-AM model under different numbers of sensors. A similar conclusion can be drawn that the accuracy of identifying track irregularity is slightly improved with the increase in the number of input accelerations. Overall, when using one acceleration input, the accuracy of the BiLSTM-AM model in identifying track irregularity is satisfactory as expected. The results of one sensor listed in Table 6 show that the RMSE and MAE values ​​are 3.9571E-04 and 3.1270E-04, respectively, and the R value reaches 20.9782.

[0120] Table 6

[0121]

[0122]

[0123] (2) Impact of running speed

[0124] The running speed of the test train is one of the most important factors in evaluating the efficiency of the BiLSTM-AM model in identifying track irregularities. Therefore, this example selects five random speeds, namely 201, 235, 272, 315 and 349 km / h, to demonstrate the impact of speed on the BiLSTM-AM model. Fig.11 The figure shows the spatial and frequency domain prediction results of the BiLSTM-AM model at different train speeds. Fig.11 It can be seen that, except for the small differences at those sharp change points, the identification of track irregularity is basically consistent with the real value (simulation). For the spectrum subgraph, similarly, the identification values ​​at different train speeds are in good agreement with the simulation values, even at speeds as high as 349 km / h.

[0125] Table 7 lists the evaluation indicators of the BiLSTM-AM model at different train speeds. Obviously, the prediction error increases slightly with the increase of speed. Among them, when the speed increases from 201km / h to 349km / h, the RMSE and MAE only increase by 16.61% and 14.77% respectively, and R2 only decreases by 1.99%. The above results prove that the proposed BiLSTM-AM model has excellent applicability in identifying track irregularities at different operating speed levels.

[0126] Table 7

[0127]

[0128] (3) Impact of noise level

[0129] In the field test, the acceleration response recorded by the test train is inevitably affected. Therefore, the proposed BiLSTM-AM model should be verified to be noise-resistant to ensure sufficient prediction accuracy. In order to analyze the impact of noise, four levels of noise are artificially added to the simulated train acceleration, such as 2%, 5%, 10%, and 20%. The recognition results and their residuals under different noise levels are shown in Figure 2. Fig.12 As shown. It can be seen that the track irregularities identified at different noise levels are basically the same, and the residual of the identified track irregularities has almost no change as the noise level increases. Table 8 lists the evaluation indicators at different noise levels. It can be seen that the error index increases slightly with the increase of the noise level, but at a noise level of 20%, the accuracy of the identified track irregularities is still acceptable, such as RMSE ~ 4.9170E-04 and MSE ~ 3.8612E-04. The comparison results mentioned above show that the proposed BiLSTM-AM model has excellent noise resistance, even at a 20% noise level of the train acceleration response.

[0130] Table 8

[0131]

[0132] (4) Influence of bridge damping ratio

[0133] Five unknown train acceleration samples were selected considering different bridge damping ratios (1%, 2%, 3%, 4%, and 5%) to demonstrate the influence of the bridge damping ratio. The track irregularity prediction results and residual performance under different bridge damping ratios are shown in Figure 2. Fig.13 As shown. Obviously, the track irregularity identification results under different bridge damping ratios are basically consistent with the true values, and the identified track irregularity residuals are difficult to distinguish from each other. Table 9 gives the evaluation indicators of the BiLSTM-AM model under different bridge damping ratios. A similar phenomenon can be observed, that is, the bridge damping ratio has a limited effect on the identification accuracy of track irregularities. This is mainly attributed to the fact that the change in bridge damping has the least effect on the acceleration response of in-service trains running at higher speeds, which in turn shows that the influence of the bridge damping ratio on the indirect identification of track irregularities using train acceleration is negligible.

[0134] Table 9

[0135]

[0136] In summary, the BiLSTM-AM model proposed in the present invention uses the acceleration response of commercial trains to indirectly identify track irregularities, which has the following advantages:

[0137] (1) The BiLSTM-AM model has excellent performance in both spatial and frequency domains and can predict different types of track irregularities, including the German low-interference and high-interference spectra, and the Chinese high-speed rail spectrum.

[0138] (2) The BiLSTM-AM model can accurately identify track unevenness by providing only the vertical acceleration of the vehicle body, without the need for professional test trains and a large number of sensors.

[0139] (3) The BiLSTM-AM model is highly suitable for indirect identification of track irregularities even when the running train speed is 350 km / h and the bridge damping ratio is 0.05.

[0140] (4) Even if the train acceleration is contaminated by noise levels up to 20%, the BiLSTM-AM model can still accurately identify track irregularities, indicating that the proposed model has good noise resistance.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for indirectly identifying track irregularities based on the acceleration response of commercial trains, characterized in that: The method comprises: constructing a VBI system model and a BiLSTM-AM model; the BiLSTM-AM model comprises an input layer, a BiLSTM layer, an AM layer, a fully connected layer and an output layer connected in sequence, and dynamically allocating different weights to the data outputted from the BiLSTM layer to the fully connected layer at each moment through the AM layer; Different types of track irregularities are input as excitations into the VBI system model to obtain train acceleration responses; The BiLSTM-AM model is trained by taking the train acceleration response as the input of the BiLSTM-AM model and taking the corresponding track irregularity as the output of the BiLSTM-AM model; The acceleration data of the train while it is running on track is collected by the acceleration sensor of the commercial train, and the acceleration data is input into the trained BiLSTM-AM model to obtain the roughness of the train track, thereby realizing the indirect identification of track roughness based on the acceleration response of the commercial train.

2. The method according to claim 1, characterized in that The training process of the BiLSTM-AM model includes: preparing a data set, which is a train acceleration response generated by the VBI system model under different types of track irregularity excitation; The data set is divided into a training set and a test set, and the BiLSTM-AM model is trained using the training set, and the loss function in the iterative training is MAE; After the training is completed, the prediction accuracy of the BiLSTM-AM model is tested using the test set, and RMSE, MAE, and R 2 To evaluate the prediction accuracy of the BiLSTM-AM model.

3. The method according to claim 1, characterized in that The acceleration data collected by the acceleration sensor of the commercial train includes one or more of the body vertical acceleration, body rotational acceleration, front bogie vertical acceleration, front bogie rotational acceleration, rear bogie vertical acceleration and rear bogie rotational acceleration.

4. The method according to claim 1 or 3, characterized in that: When predicting track irregularity using commercial train acceleration data, at least one acceleration data is input into the BiLSTM-AM model.

5. The method according to claim 1, characterized in that The VBI system model is expressed as: Where M, C, and K represent the mass, damping, and stiffness matrices, respectively; Z(t) represents acceleration, velocity and displacement vector respectively; F(t) represents load vector.

Images