Method for improving track correction processes
Patent Information
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- HP3 REAL GMBH
- Filing Date
- 2024-08-27
- Publication Date
- 2026-07-08
Smart Images
Figure AT2024060326_06032025_PF_FP_ABST
Abstract
Description
[0001] Procedures for improving track alignment procedures
[0002] Technical area
[0003] The invention relates to a method for improving the track correction methods used to improve the track position by track tamping machines.
[0004] State of the art
[0005] Most railway tracks are constructed with a ballast superstructure. The sleepers lie in the ballast. The wheel forces exerted by trains passing over them cause irregular settlements in the ballast and shifts in the lateral geometry of the track. Settlement of the ballast bed results in errors in the longitudinal height, cant (in curves), torsion, and alignment. If certain comfort or safety limits for these geometric parameters are exceeded, maintenance work is required. A track tamping machine improves the track geometry, which has deteriorated due to train loads. For this purpose, the track is lifted and aligned to the desired position using electro-hydraulically controlled lifting and straightening devices. The track is then secured in this position by compacting (tamping) the ballast beneath the sleepers.
[0006] To guide the correction tools of the track machine, measurement and control systems based on the three-point method are predominantly used. In addition to the three-point method, other methods are also known, such as the secant method, the four-point method, or the two-chord method. Practice shows that while track geometry is improved, the theoretically possible improvements are far from being achieved. Track geometry errors are typically only reduced by between 30 and 50%. The shape and position of the track geometry errors usually remain unchanged, only the amplitudes of the errors decrease. Phase shifts (spatial displacement of the corrected track errors relative to their original position) also occur. The correction properties of the track correction methods depend on the wavelength and are described by a transfer function. This shows how the amplitude and phase position of the errors at specific wavelengths are changed.
[0007] To ensure that the track can be reopened for train service after such track geometry improvement work, the railway superstructure machines are equipped with so-called acceptance measuring systems and an acceptance recorder. This recorder records the remaining defects. For release, the track geometry defects must be below specified tolerances.
[0008] The smaller the residual defects after maintenance work, the lower the interactive forces between wheel and rail caused by trains, the slower the deterioration of the track geometry, and the longer the durability of the track geometry. It is therefore desirable to bring the track geometry as close to the target position as possible, as this can subsequently save considerable costs and effort.
[0009] The target geometries of the railway tracks are available as track layout plans and are used by the tamping machine's track geometry control computer to guide the machine. Before work begins, the current track layout is measured using various methods (mechanically or with hand-pushed devices). These are compared with the target layout, and the track geometry errors are determined from this. To correct these, lifting values and reference values are determined and transferred to the tamping machine's track geometry control computer. The mechanical track measuring system, for example, based on the three-point method, consists of three measuring carriages. The front and rear carriages span a chord (steel chord or optical chord). The middle measuring carriage, located in the area of the lifting and straightening unit, measures the current vertex height, longitudinal height, and superelevation. The track geometry computer specifies the target reference value (vertex height), target lift (longitudinal height), and target superelevation at this point.The machine control system now regulates the actuators of the lifting and straightening unit so that the difference between the target and actual values equals zero. Using the correction values for lift and direction, the front end of the machine's measuring device is guided along the target track geometry, while the rear end is guided along the already corrected track. The position of the tamping machine along the track's longitudinal axis is determined using an odometer or satellite measurement data. This method is called the three-point method.
[0010] A disadvantage of the current application of the three-point and other known methods is the unsatisfactory reduction of track geometry errors by only approximately 30-50%, contrary to the theoretically expected improvement in track geometry. This inadequate function of the known methods means that the potential for effort and cost savings that would result from better application is not fully exploited. One of the reasons for this inadequate function (as is the case with the three-point system) is that the rear chord end is not guided exactly along the desired track geometry, but has residual errors that are fed back into the system. These errors arise from irregular settlement of the track after lifting and from springback of the track after straightening, as well as from the feedback of these errors to the control loop.The resulting settlements depend on the ballast depth and condition, while the springback of the track depends on the straightening forces, the properties of the rail fastening, and the behavior of the track itself. The endlessly welded track exhibits compressive stress at high rail temperatures (approximately from T>20°C) and tensile stress at low temperatures (approximately from T<20°C). After straightening, these internal stresses can cause the track to spring back or rebound. Another reason is the transfer function of the three-point system, which leads to systematic errors that cannot be compensated for with current methods. Known solutions do not reflect the overall behavior of the machine, but only the indirect influence of individual track parameters (see EP3743561 A1).Other factors influencing the tamping machine's performance include the control behavior of the lifting and straightening unit—the lifting and straightening must be completed before tamping is completed; changes in the lifting and straightening corrections during the work process; the wavelengths of the track defects to be corrected; the magnitude of the lifting and straightening values (measured by the actuators of the lifting and straightening device); inaccuracies in the measuring system (mechanical, electronic, etc.); and the tamping mode. Tamping with eccentric shaft tamping units and fully hydraulic tamping drives (EP2770108A1) are well known. Fully hydraulic.
[0011] Tamping units measure the ballast bed properties and adjust the tamping parameters accordingly to achieve an optimal tamping result (AT520117B1).
[0012] The remaining errors are recorded on the acceptance measurement chart. If this is done with an inertial-based navigation measurement system (e.g., EP3358079A1), the errors are known up to wavelengths of more than 100 m with accuracies of less than 1 mm.
[0013] The application of artificial intelligence methods is state-of-the-art. The applied AI models can be divided into various categories. A distinction is made between artificial neural networks (ANNs), adaptive neural fuzzy interference systems (AN FIS), decision support systems (DSSs), and artificial learning models. Artificial intelligence models have the ability to model complex track geometry deterioration behavior, i.e., the type, location, extent, and wavelength of track geometry defects, with high accuracy. AI models must be trained using training data sets. They are tested using test data sets. Depending on the objective of the AI application and the nature of the available data, supervised, unsupervised, and reinforcement learning methods are used.
[0014] (reinforcement learning), or a combination of these methods is used.
[0015] Today's tamping machines are characterized by extensive data acquisition. The collected data is transferred to a database via mobile network and stored on a server. Database servers manage the data from numerous machines. In this way, an ever-growing database is constantly being created, mapping the input data from a multitude of tamping machines in conjunction with the track processing results of these machines.
[0016] The amount and nature of this data accruing over time allows the use of machine learning techniques to generate statistical models of tamping machine operation, which are subsequently used in real-time machine operation. The following strategies are used:
[0017] Unsupervised learning methods are characterized by the fact that data previously interpreted by human experts is not available as training data sets. Algorithms in this category can autonomously, without human intervention, reveal hidden patterns and associations in large volumes of data. Input data about the tamping machine, the tamping process and the track infrastructure are grouped and divided into clusters. The resulting clusters reveal relationships in the data that are hardly recognizable a priori for a human expert due to the large volume of data. A correlation analysis thus determines the influence of specific input parameters, such as the compaction force and adjustment path of the tamping units, on the track geometry quality. Generative AI systems, such as Generative Adversarial Networks (GAN), transform this input data iteratively, resulting in new, corrected images of the expected track geometry orof the respective target parameters after processing. The iteration is carried out until predefined criteria are reached, e.g. compliance with residual error curve upper limits. Supervised learning starts with a database that has already been interpreted by human experts. This data is used to train a model, which is confronted with unknown data in a second phase. The system uses this to calculate a prediction for a target variable, which is either discrete (classification) or numeric (regression). The differentiation of the ballast properties (hardness category, ballast height, material) at a given track position is solved by classification. The regression forecast provides the expected remaining track defects after processing with a tamping machine.
[0018] Reinforcement learning is based on an evaluation system that rewards favorable behavior and penalizes unfavorable behavior with regard to a specific objective (e.g., compliance with an upper limit for track geometry errors). The system operates autonomously and iteratively attempts to develop new solution strategies by varying its behavior in order to maximize the evaluation result. The system's behavior corresponds to state transitions at defined points in time, from which the individual evaluations are derived. The system interacts with an environment that may or may not be known in advance. The advantage of this method is that it works without sample data and can be simulated independently of the tamping machine, e.g., in the laboratory. This technique makes it possible to optimize a machine's tamping cycle, to support it in dealing with obstacles in the track, and to carry out optimized approach and de-ramping for given track geometry in the field.
[0019] Description of the invention
[0020] The invention is based on the object of providing a method that maps the overall functioning of the machine so that, given known correction values and machine parameters, the result of the tamping work can be predicted. The invention is intended to be able to detect any deviations from the norm in the overall machine, thus enabling timely maintenance or repair measures. If this transfer function of the overall machine is known, the quality of the work can be assessed.
[0021] The invention solves the problem with the features of claims 1, 3 and 5. Advantageous developments of the invention can be found in the subclaims.
[0022] Using artificial intelligence methods, a system is trained using known measurement data from a machine's past work. The goal of the training is to predict the outcome of the track alignment adjustment based on given input data.
[0023] The prediction of the correction result by the tamping machine with known machine and track parameters indicates the accuracy of the method. According to the invention, after a tamping job, the result is assessed to determine whether it lies within the predicted result with inaccuracies of the trained AI program. This tracks the machine's operation and provides indications if the machine's performance deteriorates over time. Reasons include, for example, malfunctions or faults in hydraulic valves, faults in the machine's measuring system, changing control behavior of the lifting and straightening units, etc. This enables condition monitoring of the entire machine with regard to its performance. Even a test tamping job on a new machine during the commissioning phase allows the trained model to be used to check whether the machine's operation and thus its control and regulation mechanisms are of the required quality.This makes it possible to prepare the tamping work to be carried out in order to achieve optimal results.
[0024] Brief description of the invention
[0025] The drawing shows the subject matter of the invention schematically by way of example.
[0026] Fig. 1 Schematic explanation of the operation of the three-point method, Fig. 2 Transfer function of the three-point method,
[0027] Fig. 3 Input and output variables during the training phase of an unsupervised Kl program,
[0028] Fig. 4 Input variables and predicted output variables of a trained unsupervised Kl program,
[0029] Fig. 5 Diagrams of track defects before work and predicted remaining track defects after work with error range of the Kl program
[0030] Fig. 6 Input and output variables during the training phase of an unsupervised AI program for the design lifting
[0031] Fig. 7 Input variables and predicted output variables of the trained unsupervised Kl program for the design lifting after load
[0032] Fig. 8 Diagram of track height errors, lifting values, settlement under load and overlift values
[0033] Fig. 9 Schematic histogram of the residual error distribution of a measured variable
[0034] Fig. 10 Input variables and assessment variables during the training phase of a supervised AI program
[0035] Fig. 11 Input variables and predicted assessment variables after the training phase of a supervised AI program
[0036] Fig. 12 a concrete implementation example
[0037] Ways to implement the invention
[0038] Fig. 1 shows a schematic representation of the three-point method for track correction. The actual geometry IG with the track defects is shown as a solid black line. The virtual ideal chord IS is shown as a dotted line. This is spanned between a measuring carriage at the front and a measuring carriage at the rear, for example via a steel chord. At the front and rear, it is guided along the target geometry SG for theoretically exact track defect correction. The target value center SM is specified via a measuring carriage in the middle. The electronics controls the lifting and straightening unit to this target value and thus corrects the track defect. The real chord, however, lies at the front on the faulty actual track. To avoid this error, the deviation between the actual position and the target position must be determined before tamping work. The real chord position RS is compensated for using the correction values K determined in this way to form the ideal chord position IS.A and b denote the chord pitch, L the chord length. Track kilometer s represents the arc length in the longitudinal direction of the track, and Y (m) represents the (height) deviation of the track from Y=0.
[0039] Fig. 2 shows an example of the transfer function of the three-point method. The track faults are amplified with a gain v depending on the wavelength (or reduced if v<1). The x-axis shows the wavelength A of the track faults Fi and the vertical axis the gain. The shape of the transfer function depends on the chord pitch a, b and the chord length L. In addition to the gain response, there is also a phase response. The transfer function shown shows that, for example, a fault with a wavelength of 7m is practically not eliminated (v=1). A fault with a wavelength of 30m is reduced by 70%. This means that with real faults that have a mixture of track faults with wavelengths from 3 to 150m, residual faults remain in the track. These track faults occur at the position of the rear end of the chord and are fed back to the measuring point in the middle via the resulting chord position (see Figure 1).This means that the correction process cannot theoretically correct track errors exactly; residual errors always occur depending on the wavelength of the original errors.
[0040] The following applies to the transfer function H(A) of the three-point method: x ... track distance (m)
[0041] A ... wavelength (m) The gain v is given by v(A) = a£>s(H(A)). Abs represents the absolute value of a complex function.
[0042] The phase shift <p ergibt sich zu Angle means the calculation of the phase angle of the complex function H.
[0043] Fig. 3 schematically shows how an unsupervised AI program US-KI is trained. Characteristic input variables Ei that influence or could influence the tamping machine's performance are fed into the AI program US-KI. The errors prior to work (such as the elevation error curve 101, the directional error curve 102, and the superelevation error curve 103) are usually measured independently. If the corrections to be made are excessive, they are processed by the responsible person. Further input variables Ei are planned lifting correction values 104, planned direction correction values 105, tamping mode 106, tamping time 107, control parameters for lifting / straightening 108, type of measuring system (parameters) 109, transfer function of track position correction method 110, actual lifting during work 111, actual direction during work 112 and further characteristic input parameters 113. This also applies to the start and end ramp - because the track must not be corrected abruptly.The lifting values and reference values are slowly built up via a ramp. The planned lifting and direction correction values follow from this processing. The tamping result is also influenced by the tamping mode, whether it is a new track being laid or a maintenance tamping. The tamping time can also have an impact. The parameters of the lifting and straightening controllers used are also important. The correction can only be fixed by tamping if the lifting and straightening processes are completed before the end of the tamping. The type of correction method used also has an influence. As already shown above, the transfer function is also an important connection. The actual lifting and reference values that the actuators on the lifting and straightening unit perform deviate from the specified correction values. During the lifting and straightening process, the track geometry changes and with it the correction values for the next work step.In addition, there are other characteristic input parameters 113 that could influence the tamping result (e.g., ballast bed hardness, compaction force, etc.). The unsupervised Kl also receives the results from the individual construction sites as a progression of the elevation, direction, and camber errors. The Kl is trained using the work results from numerous tamping operations. It learns which output variables Fi arise depending on the input variables. The output variables Fi include elevation errors after work 114, direction errors after work 115, and camber errors after work 116.
[0044] Fig. 4 shows a schematic of the function of an unsupervised Kl after training US-KI-T. If the trained Kl model US-KI-T is now applied to known characteristic input variables Ei, it provides a prediction of the result FKi, predicted elevation error after the work 114V, predicted direction error after the work 115V and predicted superelevation error after the work 116V. This has the following advantages and offers the following possibilities. After the work, the result can be compared with the prediction. For example, the differences can be calculated as the standard deviation. If the machine is functioning well, the differences will always lie within a certain confidence interval. If this confidence interval is exceeded, it can be deduced that the functionality of the machine is deteriorating. This can be due, for example, to a change in the control parameters or a deterioration of the measuring system.This makes it necessary to perform a machine service or a general check of the machine. When a machine is put into operation for the first time, the quality of the tamping machine can be assessed after just a few test runs. The class can be trained on a specific machine type.
[0045] Fig. 5 shows, according to the invention, the track error F of a track section to be worked on before work begins in the diagram above. The diagram below shows the prediction of the residual error FKi of the Kl for this track section (and the planned input variables). The Kl also provides the confidence interval of the prediction as a tolerance band, the FTOLKi. Every change to an input parameter leads to a modification of the prediction. Fig. 6 shows one possibility for training an unsupervised Kl program US-KI-DH when so-called design lifting is used. It is known from practice and studies that lifts of 15-20 mm cause the track to return to its original fault position under traffic load. The reason for this is that the ballast structure under the sleepers is not significantly changed. Only with larger lifts are new stones transported under the sleepers, which leads to the permanent elimination of the track defects.During design lifting, the track is overlifted in a targeted manner depending on the fault pattern. Under subsequent tensile loads, the track settles accordingly and strives for the desired, low-fault elevation. The resulting residual errors after loading, typically after 2 million tons of traffic load, serve as the output variables FBi. After this loading, the majority of settlements have developed: elevation errors after loading 114B, directional errors after loading 115B, and superelevation errors after loading 116B. The input variable, superelevation values 117, is characteristic of the heave allowances used. These can be calculated, for example, as a percentage addition to the planned heave values.
[0046] Fig. 7 shows the application of the trained Kl program US-KI-DH-T from Fig. 6 according to the invention. The prediction now provides information about the residual errors after loading FBKi. Predicted elevation error after loading 114VB, predicted directional error after loading 115VB, and predicted superelevation error after loading 116VB. Here, too, the lowest possible residual error after track loading can be achieved by optimizing the superelevation values.
[0047] Figure 8 shows a schematic of the design lifting method. Normally, the lifting values H are calculated from the track defects Fi by introducing a lifting reference - the track zero position. After tamping, the ideal track is one that shows no deviations from a reference line. Tamping compacts the ballast beneath the sleepers after lifting to fix this position. Since the ballast stones are not in their densest position after tamping, but are forced into place by the compaction forces, changes in the positioning occur during train travel. The denser positioning results in settlements S that depend on the lifting values H. This results in irregular settlements S - track defects reappear. The design lifting theoretically takes this expected settlement S into account. Ideally, the track will theoretically assume the ideal track zero position after sufficient loading.
[0048] Fig. 9 shows a schematic histogram (frequency FR) of the residual errors Fi (height, direction, cant). Namely, the standard deviation of the residual errors Fi of the measured variable [mm], classified according to equal cumulative frequency. Typically, the standard deviation of the residual errors is calculated from 200m track sections. The higher the quality of the track geometry, the lower the standard deviations. The measurement results from work carried out by identical machines are stored in databases. Experience has shown that these distributions can be represented by log-normal distributions. This residual error distribution can then be divided into classes (K1 - K10) of equal cumulative frequency. The lowest class K1 corresponds to the best track geometries, and the highest class K10 to the worst. Using this classification, a specific result after tamping work can be objectively assessed by classifying it as belonging to a specific class.
[0049] Fig. 10 shows a supervised Kl program S-Kl. Unlike unsupervised Kl programs, this program must be provided with a result assessment FKLi during learning. Each data set learned during training must therefore be assigned a prior assessment. This can be done by specifying the quality class (K1 - K10) of the residual errors Fi (assessment variables) of the section under consideration: Elevation error quality class 114Q, direction error quality class 115Q, and superelevation error quality class 116Q.
[0050] Fig. 11 schematically shows the application of the trained, supervised Kl program S-Kl-T. If the input parameters Ei of a track section to be assessed are fed to it, it provides a prediction of the assignment of the residual errors to a quality class FKLKi. Prediction of elevation errors quality class 114VQ, prediction of directional errors quality class 115VQ, and prediction of superelevation errors quality class 116VQ. The supervised Kl program S-Kl-T also allows the application of optimization methods by adjusting the input variables Ei before tamping begins in order to achieve the best possible result. If the tamping machine results deviate from the predictions, malfunctions can be assumed, and the machine can be serviced or checked.
[0051] Fig. 12 shows a possible implementation of an unsupervised AI program. The example is based on the normalized elevation error of a known track section as the input signal, designated "before work" 1. The input signal is a superposition of an unknown number of periodic track error wave signals, each representing a track elevation error with a characteristic wavelength (e.g., X = 20 m).
[0052] The encoder network 2 analyzes the input signal and divides it into signal components, weighted according to their relevance to the overall signal 3. Dominant wavelengths in the input signal section (e.g., X = 15m) receive a higher weighting than wavelengths with a lower contribution (e.g., X = 120m). The signal, encoded through several layers of the network, is forwarded directly to a decoder network 4 and, in parallel, to another, deep neural network (DNN) 6.
[0053] Decoder network 4 reassembles the individual signal components after weighting them and calculates an output signal. In step 5, a "before-work" output is generated that should correspond as closely as possible to the "before-work" input signal. The greater this output deviates from the original input, the worse the performance of encoder-decoder pipeline 3. This comparison is represented by a cost function or evaluation function, which calculates the differences between the "before-work" input signal and the "before-work" output signal. Using this cost function and the previously unevaluated training data, the encoder and decoder networks are trained (unsupervised learning). The quality of the overall system depends heavily on the size of the available training data set.
[0054] To obtain a forecast of the expected elevation error for the selected track section "after work," the signal is transformed by DNN 6 before being passed to the decoder network. In step 7, a modified representation of the data is created, which is then passed to decoder network 4. The decoder network calculates the forecast of the elevation error "after work" as output signal 8.
[0055] The DNN 6 also requires explicit training. The training dataset consists not only of the height error "before the work," but also of the recorded results, i.e., the residual height errors "after the work." These residual height errors represent the "expert knowledge" on which the DNN can be trained (supervised learning). A portion (5%-10%) of the total available training dataset is omitted. This portion is used to validate the DNN and thus reflects the quality of the prediction.
[0056] If the system is applied in practice, the expected result of tamping by a tamping machine can be calculated in advance for a given height error on a track section. The discrepancy between the tamping result and the predicted result provides insight into emerging problems with the tamping machine.
Claims
Patent claims 1 . Method for determining probable residual errors of a track position to be corrected with at least one track-movable tamping machine after the correction has been carried out, wherein the tamping machine having lifting, straightening and tamping units operates according to a track position correction method and stores input variables (Ei) and output variables (Fi) assigned to tamping operations in a database, wherein firstly an unsupervised artificial intelligence program (US-KI) of a track tamping machine is trained with stored input variables (Ei) and stored output variables (Fi) of previous tamping operations, after which the trained unsupervised artificial intelligence program (US-KI-T) before carrying out a tamping operation,Track position parameters of a track to be corrected are read in and the unsupervised artificial intelligence program (US-KI-T) calculates the residual errors (FKi) present after tamping work is carried out.
2. Method according to claim 1, characterized in that the unsupervised artificial intelligence program (US-KI-T) determines a confidence interval of the prediction of the residual errors (FKi) as a tolerance band (FTOLKi), wherein the width of the tolerance band (FTOLKi) serves as a quality measure for the working accuracy and the wear of the tamping machine.
3. Method for determining expected residual errors of a track position to be corrected with at least one track-movable track tamping machine after the correction has been carried out, wherein the track tamping machine having lifting, straightening and tamping units operates according to a track position correction process and stores input variables (Ei) and output variables (Fi) assigned to tamping operations in a database, wherein first a supervised artificial intelligence program (S-Kl) of a track tamping machine with stored input variables (Ei) and stored classified assessment variables (FKLi) of previous tamping operations, after which track position parameters of a track to be corrected are read into the trained supervised artificial intelligence program (S-Kl-T) before tamping work is carried out, and the supervised artificial intelligence program (S-Kl-T) pre-calculates the assessment variables (FKLi) available after tamping work has been carried out before tamping work is carried out.
4. Method according to claim 3, characterized in that the supervised artificial intelligence program (S-Kl-T) determines a confidence interval of the prediction of the assessment variables (FKLi), wherein the width of the confidence interval serves as a quality measure for the working accuracy and the wear of the tamping machine.
5. Method for determining expected residual errors of a track position to be corrected with at least one mobile track tamping machine after the correction has been carried out, wherein the track tamping machine having lifting, straightening and tamping units operates according to a track position correction process and stores input variables (Ei) assigned to tamping operations as well as output variables (FBi) loaded by train traffic after tamping operations in a database, wherein firstly an unsupervised artificial intelligence program (US-KI-DH) of a track tamping machine is trained with stored input variables (Ei) and stored loaded output variables (FBi) of previous tamping operations, after which the trained unsupervised artificial intelligence program (US-KI-DH-T) before carrying out a tamping operation,Track position parameters of a track to be corrected are read in and the unsupervised artificial intelligence program (US-KI-DH-T) calculates the residual errors (FBKi) present after tamping before carrying out a tamping work.
6. Method according to claim 5, characterized in that the unsupervised artificial intelligence program (US-KI-DH-T) Confidence interval of the prediction of the residual errors (FBKi) is determined as a tolerance band (FTOLKi), whereby the width of the tolerance band (FTOLKi) serves as a quality measure for the working accuracy and wear of the tamping machine.