Methods to improve the trajectory correction method

Abstract

The present invention relates to a method for improving a track alignment method, which involves determining the expected residual error of at least one track-moving track tamping machine that operates according to a track alignment method and stores the main input variables (Ei) and output variables (Fi) in a database after a tamping operation. This involves loading the stored input parameters (Ei) and stored output parameters (Fi) from the tamping operation into an unsupervised artificial intelligence program (US-KI) and training it with these parameters. According to the present invention, the trained unsupervised artificial intelligence program (US-KI-T) is used to predict the residual error (FKi) after the tamping operation before it is performed. This allows for prior preparation of the tamping operation for optimal results.

Description

[[Technical Field]]

[0001] The present invention relates to a method for improving a track leveling method applied to improve track position by a track tamping machine. [[Background Art]]

[0002] Most tracks for trains are constructed as ballast superstructures. Sleepers are laid in the ballast. Acting wheel forces from trains running on the track cause irregular settlement in the ballast and cause lateral displacement of the track position geometry. Due to the settlement of the ballast bed, errors occur in longitudinal height, superelevation (in curves), twist, gauge and alignment position. When certain comfort or safety limit values for these geometric variables are exceeded, maintenance work is carried out. Track tamping machines improve track geometry that has deteriorated due to train loading. For this purpose, the track is lifted and aligned to a target track position by an electrohydraulically controlled lifting and leveling device, and fixed in this position by compacting (tamping) the ballast under the sleepers.

[0003] For operating the leveling tools of superstructure machines, measurement and control systems based mainly on the three-point method are used. In addition to the three-point method, other methods such as the secant method, four-point method, or two-chord method are also known. Practical use shows that although the track position is improved, it is far from achieving the theoretically possible improvement. Track position errors are typically only reduced by between 30% and 50%. In this case, the shape and position of the track position error mostly remain unchanged, and only the amplitude of the error is reduced. In addition, phase shift (local displacement of the corrected track error to its original position) occurs. The correction characteristics of a track leveling method are wavelength-dependent and described by a transfer function, which indicates how the amplitude and phase position of an error of a specific wavelength change.

[0004] After such track geometry improvement work, the railway superstructure machinery is equipped with so-called inspection measuring devices and inspection recording devices so that the tracks can be released again for train operation. These inspection recording devices record any remaining errors. For release, the track position errors must be below a predetermined tolerance.

[0005] The smaller the residual error after maintenance work, the weaker the interaction force between the train's wheels and the rails, resulting in slower deterioration of the track geometry and longer track position durability. Therefore, it is desirable to bring the track geometry as close to the target position as possible, which can subsequently save considerable cost and effort.

[0006] The target geometry of the railway track is available as a track position plan and is used to guide the machine in the track geometry control computer of the tamping machine. Before work begins, the current track position is measured by various methods (mechanically or using hand-operated equipment). These are compared to the target position, from which track position errors are determined. To correct these, lift values ​​(German: Hebewerte) and alignment values ​​(German: Richtwerte) are determined and passed to the track geometry control computer of the tamping machine.

[0007] For example, a three-point mechanical orbit measurement system consists of three measurement vehicles. The front and rear measurement vehicles are strung with strings (steel strings or optical strings). The central measurement vehicle, located within the area of ​​the lifting and alignment unit (German: Hebe-Richt-Aggregates), measures the current rise height, longitudinal height, and cant. The orbit position computer pre-programs the target alignment values ​​(rise height), target lift (longitudinal height), and target cant at this location. Mechanical control adjusts the actuators of the lifting and alignment unit so that the difference between the target values ​​and the actual values ​​is zero. Correction values ​​for lifting and alignment guide the front end of the mechanical measurement device on the orbital target geometry, and the rear end on the already corrected orbit. The position of the tamping machine on the longitudinal axis of the orbit is determined using odometer or satellite measurement data. This method is called the three-point method.

[0008] The current shortcomings of the three-point method and other well-known methods lie in their inadequacy, as they only reduce track position errors by about 30–50%, in contrast to the theoretically predictable improvements in track position. This inadequate performance of the known methods prevents the full utilization of the potential for labor and cost savings that could be achieved with better applications. One reason for this inadequate performance is that (as in the case of the three-point system, for example) the rear end of the chord is not precisely guided to the track target geometry, resulting in residual errors that are fed back into the system. These errors are caused by irregular settlement of the track after lifting, springback of the track after alignment, and the feedback of these errors into the control circuit. The resulting settlement depends on the ballast height and condition, while the track springback depends on the alignment forces, the characteristics of the rail anchors, and the behavior of the track itself. Seamlessly welded tracks exhibit compressive stress at high rail temperatures (approximately T>20°C) and tensile stress at low temperatures (approximately T<20°C). After alignment, these internal stresses can cause the trajectory to spring back or bounce. Another reason is the transfer function of the three-point system, which introduces systematic errors that cannot be compensated for by conventional methods. Known solutions do not represent the overall behavior of the machine, but only the indirect effects of individual trajectory parameters (see Patent Document 1 (EP3743561A1)). Further factors affecting the outcome of the tamping machine's operation include the control behavior of the lifting and alignment unit, where lifting and alignment must be completed before tamping is finished, changes in lifting and alignment corrections during the work process, the wavelength of trajectory errors to be corrected, the magnitude of lifting and alignment values ​​(measured by the actuators of the lifting and alignment unit), the inaccuracy of the measurement system (mechanical, electronic, etc.), and the tamping mode. Tamping using eccentric shaft tamping units and fully hydraulic tamping drive units (Patent Document 2 (EP2770108A1)) is known. A fully hydraulic tamping unit measures the characteristics of the ballast bed and adjusts the tamping parameters accordingly to achieve the optimal tamping result (Patent Document 3 (AT520117B1)).

[0009] The remaining errors are recorded in the inspection and measurement report. When this is done using an inertial-based navigation measurement system (e.g., Patent Document 4 (EP3358079A1)), errors are known to be within a wavelength range of over 100m with an accuracy of less than 1mm.

[0010] The application of artificial intelligence (KI) methods is a cutting-edge technology. The KI models applied can be divided into different categories: artificial neural networks (ANNs), adaptive neural fuzzy inference systems (ANFIS), decision support systems (DSSs), and artificial learning models. AI models have the ability to represent complex orbital degradation behavior, or the type, location, magnification, and wavelength of orbital position errors, with high accuracy. KI models must be trained using training datasets. They are then tested using test datasets. Depending on the purpose of the KI application and the nature of the available data, supervised, unsupervised, reinforcement learning, or a combination of these methods may be used.

[0011] Today's tamping machines feature extensive data recording. The collected data is transferred to a database via mobile communication and stored on a server. The database server manages data from numerous machines. In this way, a constantly expanding database is continuously generated, representing the input data from multiple tamping machines, along with the results of their track work.

[0012] The quantity and nature of this data, generated over time, allows for the use of machine learning to generate a statistical model of the tamping machine's operation, which is then used in the machine's real-time operation. In this case, the following strategy is applied:

[0013] Unsupervised learning is characterized by the fact that data pre-interpreted by human experts is not available as training datasets. Algorithms in this category can autonomously reveal hidden patterns and conceptual ideas in large amounts of data without human intervention. Input data on tamping machines, tamping processes, and orbital infrastructure are grouped and divided into clusters. The resulting clusters reveal relationships in the data that, due to the large volume of data, are largely unrecognizable a priori by human experts. Correlation analysis, therefore, identifies the impact of specific input parameters, such as the compressive force and supply path of tamping units, on the quality of orbital position. For example, a generative KI system, such as a generative adversarial network (GAN), iteratively transforms this input data, thereby generating a corrected new image of the expected orbital position or each target parameter after processing. This iteration continues until a predetermined criterion (e.g., adherence to an upper limit of the residual error profile) is achieved.

[0014] Supervised learning begins with a database already interpreted by human experts. This data is used to train a model that will confront unknown data in a second stage. From this, the system calculates predicted values ​​for a target variable, which can be either discrete (classification) or numerical (regression). Distinguishing between ballast properties (hardness category, ballast height, material) at a given orbital position is resolved by classification. Regression predictions provide the remaining orbital error expected after processing in the tamping machine.

[0015] Reinforcement learning is based on an evaluation system that rewards desirable behavior and punishes undesirable behavior with respect to a specific goal (e.g., adherence to an upper limit of trajectory position error). This system operates autonomously and attempts to develop new solution strategies through changes in behavior to maximize evaluation results. The behavior of this system corresponds to state transitions at defined points in time, from which individual evaluations are derived. In this case, the system interacts with an environment that may or may not be known beforehand. The advantage of this method is that it works without sample data and can be simulated independently of the tamping machine, for example, in a laboratory. This technique allows for the optimization of the machine's tamping cycle, assists the machine in dealing with obstacles in the trajectory, and enables optimized entry and exit at a given trajectory position in the field. [Prior art documents] [Patent Documents]

[0016] [Patent Document 1] European Patent Application Publication No. 3743561 [Patent Document 2] European Patent Application Publication No. 2770108 [Patent Document 3] Austrian Patent No. 520117 [Patent Document 4] European Patent Application Publication No. 3358079 [Overview of the project] [Problems that the invention aims to solve]

[0017] The object of this invention is to identify a method for mapping the overall operation of a machine so that the results of a tamping operation can be predicted with respect to known correction values ​​and machine parameters. This invention can detect deviations in the overall machine behavior from a standard so that maintenance or repair measures can be taken early. If this transfer function of the entire machine is known, the quality of the operation can be assessed. [Means for solving the problem]

[0018] The present invention solves the problem by the features described in claims 1, 3, and 5. Advantageous developments of the present invention are shown in the dependent claims.

[0019] The system is trained using artificial intelligence with known measurement data from the machine's past operations. The purpose of this training is to predict the results of trajectory position correction based on pre-given input data.

[0020] The prediction of correction results by the tamping machine for known machine and trajectory parameters demonstrates the accuracy of the method. According to the present invention, following a tamping operation, the result is evaluated to determine whether it falls within the predicted results of a trained KI plan, including uncertainties. This tracks the machine's operation and provides guidance if the machine's operation results deteriorate over time. Reasons for this include, for example, errors or failures in hydraulic valves, failures in the machine measurement system, and changes in the control behavior of the lifting and alignment units. This enables overall monitoring of the machine's condition regarding its operation behavior. Even during trial tamping operations in the commissioning phase of a new machine, the trained model can be used to check whether it conforms to the quality of the machine's operation, and consequently, the control and adjustment mechanisms of that machine's operation. This allows for preparation of the tamping operation to be performed for optimal results.

[0021] The drawings provide an illustrative and schematic description of the subject matter of the present invention. [Brief explanation of the drawing]

[0022] [Figure 1] A schematic description of the operation of the three-point method is shown. [Figure 2] A transfer function of the three-point method is shown. [Figure 3] Input variables and output variables during the training stage of an unsupervised AI program are shown. [Figure 4] Input variables and predicted output variables of a trained unsupervised AI program are shown. [Figure 5] Orbital error before operation, and expected remaining orbital error after operation within the error range of the AI program are shown. [Figure 6] Input variables and output variables during the training stage of an unsupervised AI program for design lifting are shown. [Figure 7] Input variables and predicted output variables of a trained unsupervised AI program for design lifting after loading are shown. [Figure 8] Diagrams of track height error, lift value, settlement under load, and excess lift value are shown. [Figure 9] A schematic histogram of the residual error distribution of a target measured variable is shown. [Figure 10] Input variables and evaluation variables during the training stage of a supervised AI program are shown. [Figure 11] Input variables and predicted evaluation variables during the training stage of a supervised AI program are shown. [Figure 12] Specific embodiments are shown. DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION

[0023] Method for Carrying Out the Invention Figure 1 schematically illustrates the three-point method as a representative method for trajectory position correction. The actual geometry IG, including trajectory errors, is shown by a solid black line. A hypothetical ideal chord IS is marked by a dotted line. This is stretched between the forward and rear measuring vehicles, for example, via a steel chord. By theoretically correcting the trajectory error precisely, the ideal chord IS is guided forward and backward relative to the target geometry SG. The target value center SM is predetermined via the center measuring vehicle. Electronic equipment controls the lifting and alignment units to this target value, thereby correcting the trajectory error. However, the actual chord is located ahead of the actual trajectory with errors. To avoid this error, the deviation from the target position to the actual position must be determined before the tamping operation. The actual chord position RS is compensated to the ideal chord position IS by the correction value K thus determined. A and b indicate chord divisions, and L indicates chord length. Orbital kilometers s represent the arc length in the longitudinal direction of the orbit, and Y(m) represents the deviation of the orbit (in height) from Y=0.

[0024] Figure 2 shows an example of a transfer function for the three-point method. The orbital error is amplified by an amplification v depending on the wavelength (or reduced if v < 1). The x-axis represents the wavelength λ of the orbital error Fi, and the y-axis represents the amplification. The shape of the transfer function depends on the string divisions a and b and the string length L. In addition to the amplification process, there is also a phase process. The described transfer function shows, for example, that the error at a wavelength of 7 m is not substantially eliminated (v=1). The error at a wavelength of 30 m was reduced by 70%. This means that in actual errors where orbital errors of wavelengths from 3 to 150 m are mixed, residual errors remain in the orbit. These orbital errors occur at the position of the rear end of the string and are fed back to the central measurement point via the resulting string position (see Figure 1). This means that the correction method can not theoretically completely eliminate orbital errors, and residual errors will always occur depending on the wavelength of the original error.

[0025] Regarding the transfer function of the three-point method, see below:

number

[0026] The amplification v is obtained by v(λ) = abs(H(λ)), where abs represents the absolute value of the complex function.

[0027] The phase shift φ is obtained by φ(λ) = angle(H(λ)), where angle represents the calculation of the phase angle of the complex function H.

[0028] Figure 3 schematically illustrates where the unsupervised KI program US-KI is trained. Characteristic input variables Ei that affect or may affect the results of the tamping machine are supplied to the KI program US-KI. Pre-operation errors (e.g., height error profile 101, alignment error profile 102, and cant error profile 103) are usually measured by an independent measurement process. If the corrections to be implemented are too high, this is handled by the person in charge. Further input variables Ei are the planned lift correction value 104, the planned alignment correction value 105, the tamping mode 106, the tamping time 107, the lift / align control parameters 108, the type of measurement system (parameters) 109, the transfer function of the method for correcting the trajectory position 110, the actual lift during operation 111, the actual alignment during operation 112, and further characteristic input parameters 113. This also applies to the first and last ramps, because the trajectory should not be corrected abruptly. The lift and alignment values ​​are built up slowly through the ramps. The planned lift and alignment correction values ​​arise from this process. The tamping mode, such as laying new track or maintenance tamping, also affects the tamping results. Tamping time can also have an effect. The parameters of the lift and alignment controller used are also important. Fixation by tamping can only be confirmed if the lift and alignment process is completed before the end of tamping. The type of alignment method applied also has an effect. As already mentioned above, the transfer function is equally important. The actual lift and alignment values ​​performed by the actuators of the lift and alignment unit deviate from the pre-given correction values. That is, the track position changes during the lift and alignment process, and therefore the correction values ​​for the next work step also change. In addition, there are further characteristic input parameters113 that can affect the tamping results (e.g., ballast bed hardness, compressive force, etc.). Furthermore, the unsupervised KI receives results from individual construction sites as profiles of height error, alignment error, and cant error. The KI is trained using the results of multiple tamping operations.It learns which output variable Fi arises depending on the input variable. The output variable Fi includes the post-work height error 114, the post-work alignment error 115, and the post-work cant error 116.

[0029] Figure 4 schematically illustrates the functionality of unsupervised KI after training US-KI-T. When the trained KI model US-KI-T is applied to a known characteristic input variable Ei, it provides a prediction of the resulting FKi, a predicted height error 114V after operation, a predicted alignment error 115V after operation, and a predicted cant error 116V after operation. This has the following advantages and offers the following possibilities: After operation, the results can be compared with the predictions. For example, the difference can be calculated as a standard deviation. If the machine is functioning normally, the difference is always within a certain confidence area. If it exceeds this, it can be concluded that the machine's functionality is degraded. This may be due to, for example, a change in control parameters or deterioration of the measurement system. This necessitates performing machine maintenance or a general inspection of the machine. If the machine is operating for the first time, the quality of the tamping machine can be evaluated after only a few test runs. KI is trained on specific machine types.

[0030] The upper figure in Figure 5 according to the present invention shows the orbital error F of the orbital section to be worked on before the work is to be done. The lower figure shows the prediction of the residual error FKi (and planned input variables) of KI for this orbital section. In this case, KI provides the confidence region of the prediction as the tolerance bandwidth FTOLKi. Each change in the input parameters leads to a correction of the prediction.

[0031] Figure 6 shows one possible scenario for training the unsupervised KI program US-KI-DH when so-called design lift is applied. From practice and research, it is known that a lift of 15-20 millimeters will return the track to its previous error position under running load. This is because the ballast structure beneath the sleepers does not change substantially. Only with greater lifts are new stones delivered beneath the sleepers, which permanently eliminates the track error. During design lift, the track in question is lifted according to the error profile. Under subsequent tensile load, the track sinks accordingly, aiming for the desired height position with less error. The output variable FBi is, here, the residual error after loading, typically after a 2 million ton traffic load. After this load, most of the sinking occurs. Post-load height error 114B, post-load alignment error 115B, and post-load cant error 116B. The input variable excess lift value 117 characterizes which additional lift amount is used. These can be calculated, for example, as an additional percentage of the planned lift value.

[0032] Figure 7 shows the application of the trained KI program US-KI-DH-T from Figure 6 according to the present invention. The predictions provide information on the residual error FBKi after loading. Predicted height error after loading: 114VB, predicted alignment error after loading: 115VB, predicted cant error after loading: 116VB. Here again, the smallest possible residual error after orbit loading can be achieved by optimizing the excess lift value.

[0033] Figure 8 schematically illustrates the design lift method. Typically, the lift value H is calculated from the track error Fi by introducing a lift reference (the zero position of the track). After tamping, ideally, the track is laid without deviation from the reference line. Tamping compresses the ballast under the sleepers after lifting in order to fix this position. The ballast stones are not in the densest position after tamping, but are forced into place by the compaction force, so their position changes as the train runs. The denser placement results in settlement S, which depends on the lift value H. Thus, irregular settlement S occurs, and the track error reappears. Design lift theoretically includes this expected settlement S in the calculation. In the ideal case, the track theoretically assumes the ideal track zero position after sufficient load.

[0034] Figure 9 schematically shows a histogram (frequency FR) of residual errors Fi (height, alignment, cant). That is, the standard deviation of residual errors Fi for the measured variable [mm] is classified by the same cumulative frequency. Typically, the standard deviation of residual errors for a 200m track section is calculated. The higher the quality of the track geometry, the lower the standard deviation. Measurement results from the same machine are stored in a database. Empirically, these distributions can be represented by a log-normal distribution. This residual error distribution is divided into classifications (K1 to K10) of the same cumulative frequency. The lowest classification K1 corresponds to the best track geometry, and the highest classification K10 corresponds to the worst track geometry. This classification allows for objective evaluation after tamping by classifying a particular result into a specific classification.

[0035] Figure 10 shows the supervised KI program S-KI. In contrast to the unsupervised KI program, this program must provide evaluation FKLi of the results during training. Therefore, each dataset learned during training must be pre-evaluated. This can be done by specifying the quality classification (K1-K10) of the residual error Fi (evaluation parameter) for the interval in question. Quality class 114Q for height error, quality class 115Q for alignment error, and quality class 116Q for cant error.

[0036] Figure 11 schematically illustrates the application of the trained supervised KI program S-KI-T. Given the input parameter Ei of the trajectory section to be evaluated, it provides predictions for the assignment of residual errors to quality classification FKLKi. Predictions for the quality classification of height errors 114VQ, for the quality classification of alignment errors 115VQ, and for the quality classification of cant errors 116VQ. In the supervised KI program S-KI-T, the application of optimization methods is possible by adjusting the input variable Ei before starting the tamping operation to obtain the best possible results. If the results of the tamping machine deviate from the predictions, this can also be considered a malfunction, and the machine can be inspected or serviced.

[0037] Figure 12 shows a possible embodiment of an unsupervised KI program. This example is based on a normalized height error of a known orbital segment as the input signal, referred to as “pre-work” 1. The input signal is a superposition of periodic orbital error wavelength signals of an unknown, each representing an orbital height error with a characteristic wavelength (e.g., λ=20m).

[0038] Encoder network 2 analyzes the input signal and divides it into signal components that are weighted according to their relevance to the overall signal 3. Major wavelengths in the input signal (e.g., λ=15m) receive higher weighting than wavelengths with smaller contributions (e.g., λ=120m). Meanwhile, the signals encoded by multiple layers of the network are directly transferred to decoder network 4 and, in parallel, to a further deep neural network (DNN) 6.

[0039] The decoder network 4 reconstructs the individual signal components according to their weights and calculates the output signal. In step 5, a "pre-work" output is generated, which matches the "pre-work" input signal as closely as possible. The further this output deviates from the original input, the worse the encoder-decoder pipeline 3 performs. This comparison is represented by a cost function or evaluation function that forms the difference between the "pre-work" input signal and the "pre-work" output signal. This cost function, along with the previously unevaluated training data, is used to train the encoder network and the decoder network (unsupervised learning). In this case, the overall system quality depends heavily on the size of the available training dataset.

[0040] To obtain a prediction of the height error that should be expected for the selected track section "after the operation," the signal is transformed by DNN6 before being passed to the decoder network. In step 7, a modified representation of the data is created and then passed to decoder network 4. The decoder network calculates the prediction of the "after the operation" height error as output signal 8.

[0041] DNN6 must also be explicitly trained. In addition to the "before work" height errors, the training dataset consists of recorded results, i.e., the "after work" residual height errors. These residual height errors represent the underlying "expertise" that allows the DNN to be trained (supervised learning). A portion (5% to 10%) of the total available training dataset is omitted and used to validate the DNN and thus reflect the quality of predictions.

[0042] When this system is actually applied, it is possible to pre-calculate the expected results of tamping in the case of a predetermined height error position in the track section. Discrepancies between the predicted results and the actual processing results provide information about potential problems with the tamping machine.

Claims

1. A method for determining the expected residual error of the position where the trajectory should be corrected after correction has been made by at least one orbital-moving orbital tamping machine, A track tamping machine having a lifting and alignment unit and a tamping unit operates according to a track positioning method and stores the input variables (Ei) and output variables (Fi) assigned to the tamping operation in a database. First, the unsupervised artificial intelligence program (US-KI) for the orbital tamping machine is trained using the stored input variables (Ei) and stored output variables (Fi) from a preceding tamping operation. Next, the trained unsupervised artificial intelligence program (US-KI-T) is given the orbital position parameters of the trajectory to be corrected before performing the tamping operation, and the unsupervised artificial intelligence program (US-KI-T) pre-calculates the residual error (FKi) that will exist after the tamping operation is performed. method.

2. The unsupervised artificial intelligence program (US-KI-T) determines the confidence region of the residual error (FKi) prediction as the tolerance band (FTOLKi), The tolerance range (FTOLKi) is used as a measure of the working accuracy and wear quality of the tamping machine. The method according to feature 1.

3. A method for determining the expected residual error of the position where the trajectory should be corrected, after corrections have been made, using at least one orbital-moving orbital tamping machine, A track tamping machine having a lifting and alignment unit and a tamping unit operates according to a track positioning method and stores the input variables (Ei) and output variables (Fi) assigned to the tamping operation in a database. First, the supervised artificial intelligence program (S-KI) for the orbital tamping machine is trained using the stored input variables (Ei) and stored classified evaluation variables (FKLi) from a preceding tamping operation. Next, the trained supervised artificial intelligence program (S-KI-T) is given the orbital position parameters of the trajectory to be corrected before performing the tamping operation, and the supervised artificial intelligence program (S-KI-T) pre-calculates the evaluation variables (FKLi) that will exist after the tamping operation is performed, before performing the tamping operation. method.

4. The supervised artificial intelligence program (S-KI-T) determines the confidence region of the prediction for the evaluation variable (FKLi). The width of the confidence zone is used as a measure of the quality of the tamping machine in terms of operational accuracy and wear. The method according to feature 3.

5. A method for determining the expected residual error of the position where the trajectory should be corrected after correction has been made by at least one orbital-moving orbital tamping machine, A track tamping machine, having a lifting / aligning unit and a tamping unit, operates according to a track alignment method, and stores the input variables (Ei) and output variables (Fi) assigned to the tamping operation, along with the load from train operation after the tamping operation, in a database. First, the unsupervised artificial intelligence program for the orbital tamping machine (US-KI-DH) is trained using the stored input variables (Ei) and stored load output variables (FBi) from a preceding tamping operation. Next, the trained unsupervised artificial intelligence program (US-KI-DH-T) is given the orbital position parameters of the trajectory to be corrected before performing the tamping operation, and the unsupervised artificial intelligence program (US-KI-DH-T) pre-calculates the loaded residual error (FBKi) that will exist after the tamping operation is performed. method.

6. The unsupervised artificial intelligence program (US-KI-DH-T) determines the confidence region of the residual error (FBKi) prediction as the tolerance band (FTOLKi), The tolerance range (FTOLKi) is used as a measure of the working accuracy and wear quality of the tamping machine. The apparatus according to feature 5.

Images