EP3053804A2 - Method and device for optimizing rail superstructure maintenance by single error classification - Google Patents

Abstract

The invention relates to a method for optimizing the track superstructure maintenance with the steps of identifying (501) single fault types occurring in the track and emerging in track surveys, determining (501) reference measurements for each of these single fault types, applying (501) data sets each consisting of a reference measurement in the form a measurement data vector and its associated single error type in a database (502) for known single error types, training a classifier (505) with the created data sets, classifying new measurement data (503, 504) of a single error based on the created data sets by the classifier (505) and deciding (506) via maintenance measures related to the single error based at least in part on the classification of the measurement data (503, 504) of the individual error. The invention further relates to a corresponding database (502) for optimizing track superstructure maintenance.

Description

The present invention relates to a method for optimizing track superstructure maintenance, particularly with regard to single errors occurring in the tracks. Furthermore, the invention relates to a database used in this context.

If a track superstructure is mentioned, this means the structure of the rails, the sleepers, the fortifications between these two, the bedding (often gravel, concrete or steel) and the underlying underground railway infrastructure infrastructure. This track superstructure, like all other parts of the railway infrastructure, must be maintained, as it is exposed to operating loads and other weather conditions during operation.

In order to gain a first overall view of the current track condition, railway infrastructure operators nowadays rely less on visual inspections of the track section, but instead use almost exclusively measuring vehicles which perform measurements fully automatically. Such measuring vehicles detect directly or indirectly the vertical deviation of the track from its desired position as a function of the worn route kilometers of the track section. In this case, the indirect measurement takes place via the detour of considering the effects of the track deviations on the measuring vehicle reaction. More specifically, the vertical axlebox accelerations are detected and by a double integration converted to the vertical track deviation. In the measurement results of the vertical track deviations often occurs the fault pattern of a significant, punctual deformation of the track in the vertical direction. For such a vertical single fault, the limit values in Germany are within a range of 9 - 21 mm vertical deviation, depending on the maximum permitted line speed.

In the following, the three most important causes of vertical single defects, namely transition zones, bad or poorly processed rail components as well as mud spots and hollow layers, are explained for a better illustration.

With regard to the transition zones, it should be mentioned that the rails deform on sections where the rigidity of the bedding changes. This is the case, for example, for tunnel entrances and tunnel exits, because in the tunnel the sleepers and rails lie on concrete, but on the outside on gravel. The gravel is due to its slightly lower rigidity under the load of the rising train something down to the concrete. This means that the track outside the tunnel slopes down slightly due to the natural setting of the ballast, but not or only imperceptibly within the tunnel because of the concrete floor. The result is a small ramp at the tunnel entrance and a sink at the tunnel exit. Road sections where such a change in ballast stiffness exists (tunnels → gravel on concrete, railway crossings → gravel on asphalt, bridges → gravel on steel, etc.) are called transition zones. These vertical single errors are relatively long-wave geometric deviations of the desired track position, because they have an extension of about 2-5 meters along the track. With regard to poorly or poorly processed rail components, it should be mentioned that the rails are generally assembled from rail sections with a respective length of about 100 m. These rail elements must therefore be welded together every 100m. If, due to poor welding, the contiguous rail sections do not lie exactly flat against one another, a welding edge is created which strikes the rolling wheel accordingly when passing the spot. Depending on the direction of travel, this single error leads to a blow up or down. Furthermore, a phenomenon often recognized as a single error is the heart of a switch. Due to the design, the track-carrying rail has a small gap in the switch (the center piece gap). Depending on the driving dynamics characteristic of the vehicle passing over the wheel bends down in this gap and then hits the massive Weichenherzstück. With ideal processing of the switch, this effect is minimal. At high wear, however, such a vertical single error often reaches the specified limits. These vertical single errors are in contrast to the vertical individual errors of the transition zones relatively short-wave geometric deviations of the desired track position, because they have an extension of a few centimeters along the track.

Lastly, we will briefly discuss vertical single defects at mud spots and hollow sites. This is an inhomogeneity of the ballast bed. If the gravel is particularly stressed at one point, the ballast stones break off at the edges. If this process repeats itself often enough, individual gravel stones can be crushed in such a way that their grain size becomes small enough to be flushed out by the precipitation. If this happens, small chambers are created in the ballast bed, in which the number of stones and thus their tightness is reduced so that an upright vehicle, the rail by its weight in presses in this hollow layer. This type of vertical single error can be recognized as a sink in the measured data.

From a certain amplitude of a vertical single error, not least for reasons of driving safety and driving comfort, a repair of the corresponding track section is required. When assessing a single error, the amplitude of the individual error measured values was and still is often still considered today, as described, for example, in the Group guideline 821 of Deutsche Bahn. In this case, when a certain limiting value for the amplitude is exceeded, one starts automatically from a critical single error, but without knowing the true error type, that is to say, for example, a crossing of the center piece or a rail joint. With regard to an exacerbation of the single error with regard to one of its properties, the so-called degradation, one nowadays uses a simple, linear regression model when making a prediction of when the single error exceeds a critical limit. Deviating prognosis models are currently not suggested in a pure amplitude analysis, since there is no information about the single error type.

However, in addition to the amplitude, other individual features of a single error measurement curve are increasingly being taken into account in the evaluation of the individual error. Such as in the EP 1 977 950 B1 In particular, slopes and surface integrals of the vertical track deviation curve are considered. Furthermore, from the KR 102011134547 A and the European standard EN13848, to assess a single error also by its wavelength, ie its extension along the track. Here it is mainly assumed that a single-error amplitude of a certain size is more critical at a smaller wavelength than at a larger wavelength. So far, the assessment is happening the criticality of a single error on the basis of a few features of an often quite complex single error measurement curve. Since the entire form of the single error measurement curve is hardly taken into consideration, no consideration is also given to a possible systematic of the curve with regard to curves of already known individual error sources. This is questionable because the shape of the single error measurement curve and thus also the single error type significantly affect the effects of the individual error on a rail vehicle with the same amplitude.

Knowing the exact type of individual error to be assessed would therefore be extremely helpful, in particular in order to be able to make optimal decisions about maintenance measures related to the single fault. Because with such decisions, it often depends on empirical values of previous individual errors of the same type. For example, there are empirical values about how a particular single-fault type develops over time. A systematic assignment of a single error to a specific single error type on the basis of existing single error original measurement data is neither in the DE 102011101226 A1 where it is only an attempt to describe the form of a single error by means of predetermined functions, nor in the EP 1 271 364 A2 described. In the EP 1 271 364 A2 Namely, it is merely described that a designation such as "good state" or "bad state" is assigned to a single defect, a direct or indirect reference to the existing single defect type is not established.

Against this background, it is an object of the present invention to provide a method for optimizing track superstructure maintenance, in which the single-fault type is determined on the basis of already existing individual-error-original measured data for each newly measured individual error can be. Furthermore, a database used in this context should be specified.

• Identifizieren von im Gleis vorkommenden und bei Gleisvermessungen hervortretenden Einzelfehlertypen;
• Ermitteln von Referenzmessungen für jeden dieser Einzelfehlertypen;
• Anlegen von Datensätzen bestehend aus jeweils einer Referenzmessung in Form eines Messdatenvektors sowie deren zugeordneter Einzelfehlertyp in einer Datenbank für bekannte Einzelfehlertypen;
• Trainieren eines Klassifikators mit den angelegten Datensätzen;
• Klassifizieren von neuen Messdaten eines Einzelfehlers anhand der angelegten Datensätze durch den Klassifikator; und
• Entscheiden über Instandhaltungsmaßnahmen bezogen auf den Einzelfehler zumindest teilweise basierend auf der Klassifizierung der Messdaten des Einzelfehlers.

The object is achieved according to the invention by a method for optimizing track superstructure maintenance, which has the following steps:

• Identification of single-fault types occurring in the track and emerging during track surveys;
• Determining reference measurements for each of these single error types;
• Creation of data sets consisting in each case of a reference measurement in the form of a measurement data vector as well as their associated single error type in a database for known single error types;
• Training a classifier with the created data records;
• Classifying new measurement data of a single error based on the created data records by the classifier; and
• Decide on maintenance measures related to the single error based at least in part on the classification of the measurement data of the individual error.

The method according to the invention begins with the so-called initialization phase, which comprises the method steps of the identification of individual error types, the determination of reference measurements for the individual error types and the creation of the data records. Subsequent to this initialization phase, the use phase follows, in which first the classifier is trained with the data records created in the initialization phase, and thereafter, as soon as newly measured single error measurement values are available, they are classified by the classifier. Expediently, as many as possible occurring in the track single fault types such as cavities, rail joints, frogs, transition zones, mud spots and hitherto unknown or unnamed individual error types are identified. This offers the advantage that more data will then be available during the later usage phase. With reference to the determination of reference measurements, it should be mentioned that these are advantageously stored on a storage medium after they have been determined. The data sets created in the next method step preferably comprise two fields. The first field is an array of measurement data. In this field, the individual data points of the individual errors are stored as vectors of different lengths. The second field is used to store the single error type. With the aid of the measured data vectors in the data sets, existing original measured data can be taken into account as such directly in the determination of the single error form of a single error. Since the data sets contain the original measured values and not just approximation functions resulting from a curve fitting, a maximum information content of the data records is guaranteed. Particularly preferably, the exact individual error type of the individual error is determined directly by the classification of the measurement data of the individual error on the basis of the applied data records and assigned to the individual error. This assignment of a single fault type can be done very quickly after the measurement, without it being necessary to send a construction crew to assess the individual fault and to determine the single fault type on the track.

Since some individual errors are difficult to detect without loading the track, such a manual assessment of the individual error would also possibly incomplete and operations would have to be repeated. Furthermore, such an assessment could also be wrong, so that construction equipment be requested, which are not needed to eliminate the special single fault type. In the worst case, the machines still work and generate new bugs, so that in addition to the old, not eliminated problem now a new thing would come. This problem does not exist in the present method, since a first, already very reliable assignment of the single error type is done automatically.

By training the classifier with the created datasets using methods of data mining, such as the method of supervised learning, the single error type can also be determined very reliably. In this case, in the aforementioned methods, in particular by analyzing the measurement data provided and recognizing certain trends or patterns which deviate from the other structureless noise, a model is created on the basis of which a classification of later newly measured individual errors can take place. The classifier can be, for example, an algorithm, a program object or a function. It is extremely important to be able to make a decision about maintenance measures related to the single error based on a reliable classification of the measurement data of the individual error and thus on a reliably determined single fault type, because different maintenance types have to be used for different types of faults. Likewise, the preparation of a forecast on the further temporal development of the individual error depends strongly on its single-fault type.

• Validieren der Korrektheit der Klassifizierung der Messdaten des Einzelfehlers; und
• wenn die Klassifizierung korrekt ist, Anlegen eines neuen Datensatzes bestehend aus der Einzelfehlermessung in Form eines Messdatenvektors sowie deren als korrekt validierten Einzelfehlertyp in der Datenbank für bekannte Einzelfehlertypen; oder
• wenn die Klassifizierung nicht korrekt ist, Anlegen eines Datensatzes in einer Datenbank für unbekannte Einzelfehlertypen.

According to the invention, the method has the following further steps:

• Validating the correctness of the classification of the measurement data of the individual error; and
• if the classification is correct, create a new data set consisting of the single error measurement in the form of a measurement data vector and its correctly validated single error type in the database for known single error types; or
• if the classification is incorrect, create a record in a database for unknown single-error types.

This has the great advantage that possible errors in the classification of the measurement data of the individual error can be detected. In addition, erroneous records can be prevented from being added to the database for known single-error types. The process step of the validation can be carried out both in time before a decision on maintenance measures or an initiation of maintenance measures and in time thereafter. However, it is particularly advantageous first to initiate construction measures, for example, and then to use the presence of a construction crew at the location of the individual fault in order to have it validated whether the assignment of the single fault type was correct. The risk of incorrect allocation of resources is manageable due to the very reliable classification. With a correct classification, the creation of new data records in the database for known single-fault types makes it possible to take into account the corresponding single-fault original measurement data when determining the single-fault type of a later-measured individual fault. Thus, over time, a solid "single error-> single error type repository" arises and the method can be continuously improved over the long term, because with the new single error measurement data and thus an increasing quantity of reference measurements, the training of the classifier is steadily improved. Similarly, those in the database for unknown Individual error types created records for later further analysis.

According to the invention, it is further provided that, in particular after a data record has been created in the database for unknown single error types, an attempt is made to find new types of errors with the aid of a cluster method. As soon as a new error type is found, a new record is created in the database for known single error types. This is advantageous since, as a result, the number of individual classes available and listed in the database for known individual error types is thereby increased for individual errors measured in the future. Thus, individual errors measured in the future have an increased probability of having to determine their individual single-fault type at first attempt without having to make a detour via a cluster method. Also, the subsequent inclusion of new single error types makes sense, since it can be assumed that with constant technical development of the track superstructure technology always new, previously unknown sources of error for individual errors in the track are created. However, the new single error type is expediently validated before the creation of a new data record in order to ensure the reliability and correctness of the database of known single error types. In the event that no new error type can be determined, the first or the corresponding data records initially remain in the database for unknown individual error types. For more details on the clustering method, reference is made to the following discussion regarding the identification of single-fault types in the track survey measurement data in the case of unknown sources of error.

According to one embodiment of the invention, the validation of the correctness of the classification of the measurement data of the individual error by an assessment of the individual error by one or more experts, preferably a construction crew, carried out on site. This offers the advantage that, in the case of a manual assessment of the individual error, the experts, based on their experience, their specialist knowledge and the examination of the track on site, can quickly determine which source of error is involved. As a rule, such manual validation makes it much more reliable to determine whether the classification was correct or not via computer-implemented validation procedures.

• für bekannte Fehlerquellen, wie insbesondere gleisbauliche Elemente, Identifizieren von Einzelfehlertypen in den Messdaten der Gleisvermessungen anhand bekannter Wirkzusammenhänge und baulicher Kenntnisse der vermessenen Gleisstrecke; oder
• für unbekannte Fehlerquellen, Identifizieren von Einzelfehlertypen in den Messdaten der Gleisvermessungen durch:
  • Isolieren von n Einzelfehlern mit jeweils einem Messdatenbereich ε links und rechts einer erhöhten Einzelfehleramplitude basierend auf einer Amplituden-Grenzwert-Betrachtung aus den Messdaten der gesamten vermessenen Gleisstrecke, wodurch n Einzelfehler-Messdatenvektoren entstehen;
  • Definieren einer Anzahl von m Clustern;
  • Anwenden von Clustering Algorithmen auf die n isolierten Einzelfehler unbekannten Einzelfehlertyps für m Cluster;
  • Bestimmen der Qualität des Clustering-Resultats mit Qualitätsmaßen wie der Interklassenheterogenität und/oder der Innerklassenhomogenität;
  • gegebenenfalls Erhöhen der Cluster-Anzahl m um eins und wiederholte Anwendung der Clustering Algorithmen sowie der Bestimmung der Qualität, wobei die bestimmte Qualität des Clustering-Resultats für m Cluster eine optimale Klassenanzahl k ergibt, was darauf hindeutet, dass es eine Anzahl von k verschiedenen, unbekannten Einzelfehlertypen gibt.

According to a further embodiment of the invention, the method step of identifying individual error types comprises the following steps:

• for known sources of error, such as in particular track construction elements, identifying individual error types in the measurement data of the track surveys based on known causal relationships and structural knowledge of the measured track section; or
• for unknown sources of error, identification of individual error types in the measurement data of the track surveys by:
  • Isolating n single errors, each with a measured data range ε left and right of an increased single error amplitude based on an amplitude-limit value analysis from the measurement data of the entire measured track distance, whereby n single-error measurement data arise;
  • Defining a number of m clusters;
  • Applying clustering algorithms to the n isolated single errors of unknown single error type for m clusters;
  • Determining the quality of the clustering result with quality measures such as inter-class heterogeneity and / or inner-class homogeneity;
  • optionally increasing the cluster number m by one and repeated application of the clustering algorithms and the determination of the quality, wherein the determined quality of the clustering result for m clusters gives an optimal class number k, which indicates that there are a number of k different, unknown single fault types.

The known sources of error are single-fault types in which an operative connection is already known and which are often due to track-laying elements. The individual faults occurring in the track include track-building elements such as switch frogs or rail joints, which produce measured data which are errors according to the current standard (if the amplitude becomes large enough). Of these elements, the position is known. Thus, reference curves of these locations can easily be determined if they are defective in the sense of the amplitude consideration, by the position of a single error identified in the measured data matches the projected position of one of these components. If this is the case, a measurement data set of a defective track-laying element is available. The same applies to sections of the route at transition zones. Since the cause of this single fault type is the change in the rigidity of the ground, it is usually found - but not exclusively - directly in front of or behind engineering elements (such as tunnels, bridges, etc.) along the track. Since changes in the ballast stiffness are also known from the route configuration out, the identification of these individual errors and the determination of the perform typical reference curves analogous to the track construction elements.

With regard to the unknown error sources, it should be mentioned that all individual error types not yet described are now to be further identified by means of a cluster method. Here, for example, you proceed as follows. First, one isolates from the measured data sets of an entire track section all those measured data which, in the sense of the amplitude consideration, are sheepish with a certain epsilon of eg 15 meters left and right of the increased amplitude. This produces n-many vectors, each of which represents an error in the sense of the amplitude limit value analysis. Then one defines m-many clusters (eg two) and uses clustering algorithms such as the K-means algorithm or self-organizing maps. Then you measure the quality of the result from the previous cluster method with cluster analysis quality measures such as inter-class heterogeneity or inner-class homogeneity. Then you increase m by one and repeat the process. The previously determined quality of the clustering for m-many clusters now results in an optimal number of classes. This identified optimal number of classes k now indicates that there are obviously still k-many different unclassified single-error types. Thus, certain individual error types can be determined for initially unknown single errors. It is advantageous if this is verified by expert assessment on site. If it then turns out that these individual errors in the field actually represent a separate class, they must be added to the list of error types and the determination of the corresponding reference measurements can continue.

• Isolieren der Referenzmessung, bei bekannter Position einer Fehlerquelle, wie einem gleisbaulichen Element, aus vorhandenen Messdaten der vermessenen Gleisstrecke an der bekannten Position mit einem Messdatenbereich ε links und rechts der bekannten Position.

According to a further embodiment of the invention, the method step of determining a reference measurement for a single-fault type comprises the following step:

• Isolating the reference measurement, in the case of a known position of an error source, such as a track construction element, from existing measurement data of the measured track section at the known position with a measurement data area ε left and right of the known position.

If different types of single faults have previously been identified, the required reference measurements of elements where the position is known can be cut from the position based on existing vertical track deviation measurements along the track. For example, the positions of the switch frogs can be read out using the switch directory. If you now isolate the measured values at these positions with an epsilon of, for example, 15 meters to the right and left of the switch frog piece, you will get very comfortable and uncomplicated reference measurements for switch frogs. The choice of an appropriate Epsilons is advantageous in this case, since it can ensure that one selects and stores a maximum information content of the single error curve. Similarly, you can proceed with all other track construction elements (eg tunnel entrances and tunnel exits). For all other errors, these are first identified by a known amplitude comparison method and their measured values are cut out with an epsilon of, for example, 15 meters left and right of the increased amplitude from the total data set. The reference measurement or the measurement data characteristic of this single error is now available, but its type has not yet been confirmed. This can be done, for example, by expert personnel on site so that a definite type can be assigned to this reference curve. For example, when evaluating the single error, it turns out that the single error is due to a rinse in the ballast bed caused by rain, the now known error type could be referred to as "rinsing with rain".

A further embodiment of the invention is that the measurements are measurements of the vertical deflection of the tracks from their desired position as a function of the line kilometers, preferably under load of the tracks by a measuring vehicle driving thereon, and in the case of the single error by corresponding vertical single errors is. As already mentioned above, the possible fault patterns in the track superstructure are manifold. However, vertical individual errors of these varied fault patterns occur particularly frequently and, due to their physical orientation, have a particularly noticeable influence on the vehicle reaction of a rail vehicle and thus on ride comfort and driving safety.

According to a further embodiment of the invention, methods of data mining, methods of multivariate data analysis, methods of classification and of supervised and unsupervised learning, such as decision trees, neural networks, support vector, are used to train the classifier and / or to classify the measurement data of the individual error Machines, Genetic Algorithms, as well as k-nearest-neighbors, kNN, algorithms applied. In this case, it is particularly advantageous that the previously mentioned methods make it possible to train the classifier better with each new reference measurement, and thus also the quality of the results of the classification carried out by the classifier constantly improves with time. However, the aforementioned methods are only to be seen as examples.

A further embodiment of the invention is that regression methods and / or within the framework of the classification of the measurement data of the individual error Correlation method used to assess the degree of similarity or correspondence between the single error measurement curve to be classified and the reference measurement curves. Depending on the order of the respective regression method, a very good functional description of the measurement data of the individual error can be achieved. If the result of the regression method is included in the further evaluation of the individual error, one can speak of an extensive consideration of the single-error form. The use of correlation methods is also particularly advantageous, as these enable the relationship between the measurement curve to be classified and the reference measurement curves to be illuminated.

A further embodiment of the invention is that one or more features of the measurement curve of the individual error, such as amplitudes, frequencies, slopes, integrals, or the entire shape of the measurement curve of the individual error is taken into account within the classification of the measurement data of the individual error. In general, the probability of a correct classification of the individual error can be increased with an increasing number of considered features. It is therefore particularly advantageous to compare the measurement curve of the individual error with the respective reference measurement curves as holistically as possible. However, which features are taken into account and to what extent depends to a large extent on the algorithm used.

• Entscheiden über notwendige Ressourcen, insbesondere Personal-, Werkzeug-, Maschinen-, und Zeitressourcen, zur zielgerichteten Beseitigung des Einzelfehlertyps;
• Entscheiden über notwendige Sicherungsmaßnahmen einer einzurichtenden Baustelle im Hinblick auf den zu beseitigenden Fehlertyp; und/oder
• Treffen einer fehlertypspezifischen Prognose des voraussichtlichen Degradationsverhaltens des Einzelfehlers.

According to a further embodiment of the invention, the step of deciding on maintenance measures comprises the following steps:

• Deciding on necessary resources, in particular personnel, tool, machine, and time resources, for the purposeful elimination of the single fault type;
• Deciding on necessary safety measures of a construction site to be set up with regard to the type of defect to be eliminated; and or
• Making an error-type-specific forecast of the expected degradation behavior of the individual error.

Since the decision on maintenance measures can be made based on the classification of the individual fault to be eliminated and thus also on the basis of available information about the exact single fault type, the necessary resources can be allocated in advance of possible construction measures. It is crucial that each individual fault type must be maintained differently. In hollow layers, for example, the ballast must be stuffed, including a hand tamping machine or even a stuffing box is necessary, which compresses the underlying ballast again. In the case of center piece individual defects, the steel is ground into the appropriate shape, for which corresponding grinding machines are required. Furthermore, specially trained experts such as switch mechanics or point masters must be present when working on switches. Since different resources (technical as well as personnel or time resources) are required for each individual error type, it is particularly advantageous that in the present method, by early knowledge of the single error type, the right decisions are made about the necessary resources and the associated safeguards already in advance of a construction measure can be. A time-delayed decision about the necessary resources only after a manual inspection and in view of the single fault location is thus eliminated.

The information about the single fault type to be expected also allows a better scheduling of track closures, whereby a loss of availability of the route can be kept low. This is particularly the case because it is known how long it takes to repair a single fault of a particular type. In connection with the requirement of track closures, it must be taken into account in particular that construction crews in the track must be operationally secured due to occupational safety. This means that in a maintenance measure, depending on the size of the process, neighboring tracks must be locked or slow driving must be set up. If now too large machines are ordered due to ignorance of the exact single fault type, the corresponding safeguards can also be too large, which leads to loss of availability and the generation of delay minutes. All these problems do not occur in the present process.

Another advantage is that knowledge of the single-fault type and the possible error-type-specific prognosis of the expected degradation behavior of the individual error no longer limits one to a linear regression model in the prognosis. Such a linear relationship has hitherto often been achieved by merely looking at the course of the amplitudes of the same single error over a few consecutive days. So far, this has been very inaccurate, since different types of single defects can also differ in their degradation behavior. After all, a faulty rail joint can also develop logarithmically over time, as it is driven off the trains over time. A hollow layer, however, is getting worse and destroys the adjacent gravel under operating load, where This single error can also behave exponentially. Such error-type-specific forecasts of the development of individual errors thus offer the advantage that, for example, with two different individual error types, the decision can be made to eliminate one single-fault type due to its probable development over elimination of the other single-fault type despite the same single error amplitude of the two individual errors. Such an assessment of how long it will be possible to delay the elimination of a single fault is particularly advantageous with regard to minimizing the timetable failures and with regard to the annual financial budget of the railway undertakings.

Advantageously, the decision on maintenance measures may also include considerations and decisions with regard to ride comfort and passenger safety. The present method also allows a maintenance manager to be informed early on the exact single fault type. Maintenance managers must make decisions based on certain information, such as when, where, and how a track track is renewed or maintained. Through the present process, we will expand the information available and optimize the decision-making process of the maintenance manager. For example, the maintenance manager can optimally dimension the construction safety at an early stage. This results in big positive effects in terms of cost savings during the maintenance process.

The object of the present invention is achieved not only by the method described above, but also by a database used in this context. The database for optimizing the track superstructure maintenance according to the invention comprises data sets that each consist of reference error measurement values of a single error that occurs in the track and emerges in track surveying, in the form of a measured data vector and an associated single error type, with the data records a classifier is trained to classify later using the data sets new measurement data of a single error and being at least partially based on this classification on maintenance measures based on the single error. With regard to the advantages and further developments of such a database, reference is made to the above statements in the description of the method. In particular, however, it should be mentioned that the systematic construction of the database with reference error measurement values and associated single error type together with a possibility to constantly expand the database by adding new data sets enables a continuous improvement of the training of the classifier and thus an increasingly reliable determination of single error types. This is related to the increasing quantity of original measurement data from reference measurements taken into account in the training and classification procedures. It should be noted that the database is a database for known single-error types. A corresponding database for unknown individual error types has similar data records with reference error measurement values in the form of a measurement data vector whose individual error type, however, initially remains indefinite.

Thus, for the first time, a method for optimizing track superstructure maintenance is provided by means of the method described above, in which the single-fault type can be determined on the basis of already existing individual-error-original measured data for each newly measured individual error. Furthermore, a database that can be used in this context will be provided for the first time.

Figur 1

ein beispielhaftes Messergebnis der vertikalen Gleisauslenkung entlang einer vermessenen Gleisstrecke;

Figur 2

eine beispielhafte Auflistung von möglichen Datensätzen in einer Datenbank für bekannte Einzelfehlertypen;

Figur 3

ein erstes Ausführungsbeispiel des erfindungsgemäßen Verfahrens;

Figur 4

ein zweites Ausführungsbeispiel des erfindungsgemäßen Verfahrens; und

Figur 5

eine schematische Übersicht der Ausführungsbeispiele aus den Figuren 3 und 4 mit weiteren Aspekten des erfindungsgemäßen Verfahrens.

In the following drawings, the present invention is illustrated in more detail with reference to schematically illustrated embodiments. Show in it

FIG. 1

an exemplary measurement result of the vertical track deflection along a measured track section;

FIG. 2

an exemplary listing of possible records in a database for known single error types;

FIG. 3

a first embodiment of the method according to the invention;

FIG. 4

a second embodiment of the method according to the invention; and

FIG. 5

a schematic overview of the embodiments of the Figures 3 and 4 with further aspects of the method according to the invention.

The FIG. 1 shows an exemplary measurement result of the vertical track deflection in mm along a measured track distance, that is, depending on the worn track kilometer of the track section. Looking at the course of the trace, you will find some significant increases and decreases in the vertical track deflection. For a large number of these increases and decreases, an operative connection, that is to say the respective type, the respective cause or the respective source of error, is known. Viewed from the left, the first increase W occurs due to a known position of a switch frog, the second increase B and the first decrease T due to the known positions of tunnel entrances and exits and the last two increases Ü (or combinations of increase and decrease) due to a known position of a railroad crossing. After the amplitude consideration in particular the increases or decreases are to be regarded by definition as a single error exceeding the applicable limit G. In the exemplary measurement result of FIG. 1 the amplitude limit G is ± 11 mm. The measurement data of these individual errors must now be cut out in order to be examined more closely. Noticeable here is above all the third increase Z seen from the left, since this is a single fault on a free route with an unknown single fault type. Since its cause is not known, an assessment must be made accordingly.

The FIG. 2 shows an exemplary listing of possible records in a database for known single error types. In the left column, the first field, the data points of the individual errors are stored as vectors of different lengths. In the middle column, the second field, the individual error types are stored, whereas in the right column the data records are numbered consecutively for the sake of clarity. The single fault types listed here (tunnel, railway crossing, switch frog) are purely exemplary.

The FIG. 3 shows a first embodiment of the method according to the invention. Here it is assumed that an initialization phase has already taken place and thus already a database for known single error types is available. The method begins with the identification of a single-fault type by a classical amplitude analysis (step 1). Here, only the exceeding of a certain amplitude limit value is offset by an increase in the vertical track deflection measurement curve. In the next step (step 2), the measuring range relevant for the single error is extracted from the overall measurement. Then the extracted measurement range of the single error is classified based on the data records in the database (step 3.). According to the classification, this embodiment is the single error type B. In the next step, the correctness of the classification is validated (step 4). Since in this embodiment the classification was validated as correct, in the last step (step 5) a corresponding new data record is created in the database for known single error types.

The FIG. 4 shows a second embodiment of the method according to the invention. Here it is also assumed that an initialization phase has already taken place and thus already a database 1 for known single error types is available. The procedure begins with a test drive (see step I.) in which track deflections along the test section are detected. Subsequently, single errors from the total measurement (see step II.). In a next step, the measurement data of the individual errors are classified on the basis of the data records in the database 1 (see step III.). Subsequently, the correctness of the classification is validated (see step IV.). If the classification is validated as correct, a new record or records will be created in database 1 (see step V.). However, if the classification is not validated correctly, a clustering process (step VI.) Occurs to identify new single-fault types. In the diagram, each x represents a reclassified single error n to an n-dimensional feature space. First a number of clusters are defined and then a cluster algorithm is applied to the unclassified single errors for the defined number of clusters (see step VI.1.). Finally, the quality of the result of the applied cluster algorithm with quality measures such as the inner-class homogeneity and the inter-class heterogeneity is determined (see step VI.2.). The determined quality now yields an optimal number of 3 classes, which in turn indicates that 3 are still unknown Single error types exist (see step VI.3.). In a last step, the new error type is validated in the field before it is added to the database 1 (see step VI.4.).

The FIG. 5 shows a schematic overview of the embodiments of the Figures 3 and 4 with further aspects of the method according to the invention. The schematic overview begins with an initialization phase 501 identified within the individual error types and corresponding reference measurements are determined. A conclusion of the initialization phase is the creation of data records in a database 502 with reference measurement data for single errors whose type is known. The identification of individual errors and the extraction of measured values of the individual errors from a total measuring data set of a measuring train (see 503, 504) has already been described in more detail in FIGS. 3 and 4. As also related to the Figures 3 and 4 described in more detail, a classifier 505 classifies the measurement data of each individual error based on the data records in the database 502. Based on this classification, an optimal repair of the single error is then initiated (see 506). Subsequently, the correctness of the classification is validated (see 507). If the classification is correct, another reference record is added to the database 502 (see 508) to further improve the classification for the future. If the classification was incorrect, a reference record is also added (see 509), but to a database 510 with single measure reference metrics whose type is not known. In a next step, the attempt is made by means of a clustering method, which in connection with the FIG. 4 to find new types of errors (see 511). After validating whether a new error type has actually been found (see 512), a corresponding single error is either added to the single error reference measurement database 502 whose type is known (see 513) or it remains in the database 510 with reference measurement data for single errors whose type is not known (see 514).

Claims (12)

Method for optimizing track superstructure maintenance, comprising the steps: Identifying (501) single-fault types occurring in the track and emerging in track surveys; Determining (501) reference measurements for each of these single error types; Creation (501) of data sets each consisting of a reference measurement in the form of a measurement data vector and their associated single error type in a database (1, 502) for known single error types; Training a classifier (505) with the created records; Classifying, by the classifier (505), new measurement data (503, 504) of a single error based on the applied data records; and deciding (506) maintenance actions related to the single fault based at least in part on the classification of the measurement data (503, 504) of the single fault. The method of claim 1, further comprising the steps of: Validating (507) the correctness of the classification of the measurement data (503, 504) of the single error; and if the classification is correct, creating (508) a new data set consisting of the single error measurement in the form of a measurement data vector and its correctly validated single error type in the database (1, 502) for known single error types; or if the classification is incorrect, creating (509) a record in a database (510) for unknown single error types. Method according to claim 2, characterized in that after the creation (509) of a data record in the database (510) for unknown single error types the attempt (511) is made to find new types of errors with the aid of a cluster method and, as soon as a new error type is found, a new record is created in the database (1, 502) for known single error types (513). Method according to one of claims 2 or 3, characterized in that the validation (507) of the correctness of the classification of the measured data (503, 504) of the individual error by an assessment of the individual error by one or more experts, preferably a construction crew, carried out on site , Method according to one of the preceding claims, characterized in that the identification (501) of individual error types comprises the steps: for known sources of error, such as in particular track construction elements, identifying individual error types in the measurement data of the track surveys based on known causal relationships and structural knowledge of the measured track section; or for unknown sources of error, identification of individual error types in the measurement data of the track surveys by: Isolating n single errors, each with a measured data range ε left and right of an increased single error amplitude based on an amplitude-limit value analysis from the measurement data of the entire measured track distance, whereby n single-error measurement data arise; Defining a number of m clusters; Applying clustering algorithms to the n isolated single errors of unknown single error type for m clusters; Determining the quality of the clustering result with quality measures such as inter-class heterogeneity and / or inner-class homogeneity; possibly increasing cluster number m by one and repeated application of the clustering algorithms and the determination of the quality, whereby the determined quality of the clustering result for m clusters gives an optimal class number k, which suggests that there are a number of k different unknown single error types. Method according to one of the preceding claims, characterized in that the determining (501) of a reference measurement for a single error type comprises the step: Isolating the reference measurement, in the case of a known position of an error source, such as a track construction element, from existing measurement data of the measured track section at the known position with a measurement data area ε left and right of the known position. Method according to one of the preceding claims, characterized in that the measurements are measurements of the vertical deflection of the tracks from their desired position as a function of the route kilometers, preferably under load of the tracks by a measuring vehicle traveling thereon, and in the case of the single error corresponding vertical single error is. Method according to one of the preceding claims, characterized in that for training the classifier (505) and / or for classifying the measurement data (503, 504) of the individual error methods of data mining, methods of multivariate data analysis, methods of classification and the monitored, and unsupervised learning, such as decision trees, Neural networks, support vector machines, genetic algorithms, as well as k-nearest-neighbors, kNN, algorithms are applied. Method according to one of the preceding claims, characterized in that as part of the classification of the measured data (503, 504) of the individual error regression and / or correlation method for assessing the degree of similarity or the correspondence between the to be classified measurement curve of the single error and the curves of the reference measurements are applied , Method according to one of the preceding claims, characterized in that in the context of the classification of the measured data (503, 504) of the individual error one or more features of the trace of the individual error, such as amplitudes, frequencies, slopes, integrals, or the entire shape of the trace of the single error is taken into account. Method according to one of the preceding claims, characterized in that decision-making (506) via maintenance measures comprises the steps: Deciding on necessary resources, in particular personnel, tool, machine, and time resources, for the purposeful elimination of the single fault type; Deciding on necessary safety measures of a construction site to be set up with regard to the type of defect to be eliminated; and or Making an error-type-specific forecast of the expected degradation behavior of the individual error. Database (1, 502) for the optimization of track superstructure maintenance with data sets each consisting of reference error measurement values of a Single error that occurs in the track and emerges in track surveying, in the form of a measured data vector and an associated single error type, with the data sets a classifier (505) is trained to later classify new data based on the data sets (503, 504) of a single error and based at least in part on this classification, on maintenance measures related to the single error (506).

Images