Single-stage evaluation of an axle counter time characteristic

Abstract

Single-stage evaluation of an axle counter time characteristic Abstract A method for the evaluation of a data set (110) is described, the method including: i) recording of a one-dimensional time characteristic as a data set (110) by means of an axle counter (100); and ii) evaluation of the data set (110) by means of the determination of signals in the one-dimensional time characteristic and classification of the signals in the one- dimensional time characteristic in one step by means of a single-stage algorithm (150) based on deep learning. (Figure 1) Single-stage evaluation of an axle counter time characteristic Abstract A method for the evaluation of a data set (110) is described, the method including: i) recording of a one-dimensional time characteristic as a data set (110) by means of an axle counter (100) ; and ii) evaluation of the data set (110) by means of the determination of signals in the one-dimensional time characteristic and classification of the signals in the one- dimensional time characteristic in one step by means of a single-stage algorithm (150) based on deep learning. (Figure 1) 20 25 20 29 07 24 A pr 2 02 5 2 0 2 5 2 0 2 9 0 7 2 4 A p r 2 0 2 5 A b s t r a c t 2 0 2 5 2 0 2 9 0 7 2 4 A p r 2 0 2 5 A b s t r a c t 1 / 6 111 135 130112 110 131, 160 FIG 2 111 132, 160 FIG 3 110 151 152 150 120 112 111 FIG 1 1 / 6 FIG 1 120 150 110 151 152 111 112 FIG 2 131, 160 112 130 110 135 111 FIG 3 132, 160 111 20 25 20 29 07 24 A pr 2 02 5 2 0 2 5 2 0 2 9 0 7 2 4 A p r 2 0 2 5 1 1 0 1 5 2 1 1 0 1 3 5 2 0 2 5 2 0 2 9 0 7 2 4 A p r 2 0 2 5 1 1 0 1 5 2 1 1 0 1 3 5

Description

Single-stage evaluation of an axle counter time characteristic Related application

[0001] The present application claims the benefit of priority to German Application 10 2024 204 664.7, filed 21 May 2024, the entire contents of which are incorporated by reference herein for all purposes. Technical field

[0002] The invention relates to a method for data evaluation of axle counter data, wherein a one-dimensional time characteristic is recorded by means of the axle counter and is evaluated by means of determination / detection of signals in the onedimensional time characteristic and classification of the signals in the one-dimensional time characteristic in one step (in a single stage) by means of a single-stage algorithm based on deep learning. The invention further relates to an apparatus for data processing and to an axle counter system.

[0003] The invention can thus relate to the technical field of the evaluation of data sets of an axle counter, in particular about rail vehicles. Technical background

[0004] Axle counters are known in the field of rail-bound vehicles or rail infrastructure and form a significant part of the control, for example regarding whether a wheel or a wheel axle (of a train) passes a track and how many wheels have moved past. This can be of crucial importance for many applications, 2025202907   24 Apr 2025 for example track vacancy notification and in particular for safety within a rail infrastructure. In general, sections of a track are in each case equipped with two axle counter sensors (known as counting points), wherein one is located at the start of the section and the other at the end of the section. Whether a track section is ascertained as free or occupied then depends in particular upon the number of wheels that are captured by the two axle counter sensors. A higher-level evaluation unit (this can form an axle counter system together with the axle counter sensor, for example) evaluates the signals captured by the sensors and derives the status of the track section therefrom. Only if a section is also actually identified as free is a train permitted to pass this track section.

[0005] In the context of the present document, the term “axle counter” can comprise both the unit “sensor (or counting point) and evaluation unit” and also relate only to the evaluation unit (the actual axle counter). If reference is to be made specifically to the sensor, then the term “axle counter sensor” can be used, and if reference is to be made specifically only to the evaluation unit, then the term “evaluation unit” can be used. If reference is to be made to axle counter sensor and axle counter evaluation unit together, then it is possible to use the term axle counter system.

[0006] Particularly in the field of rail vehicles, however, particularly high requirements are placed on safety and reliability. Accordingly, it is only possible to use data sets of an axle counter evaluated in numerous application cases, in particular for automation of rail traffic, if particularly high reliability has been demonstrated.

[0007] Previously, an axle counter has been unable to achieve 100% counting accuracy in use. This is due on the one hand to complicated / complex signals that are captured by the axle 2025202907   24 Apr 2025 counter sensors, and on the other hand to the processing methods for signals that are not accurate enough to handle complex signals; among other reasons, also because the signals can potentially be subject to interference (noise) and because they potentially produce numerous curve forms due to the different velocities. In extreme cases, trains can stand and / or turn around at the sensor. This then would not be a safety-related problem, as the counts would always be on the safe side (in the sense that the safety buffer is increased), but rather an availability problem.

[0008] At present, the counting of the wheels (or evaluation of the data set) is performed in a very conservative manner, in order to ensure the safety of the rail vehicles. Thus, a section of a route is considered to be “occupied” and not “free” if the counting accuracy is not 100% certain.

[0009] A further limitation of the conventional axle counters consists in that, although they can identify the appearance of wheels with high accuracy, even highly modern products can only deliver such results at the expense of false positive identifications (which means that, although no wheel is passing the sensor, a signal is identified (for example, by metal masses that influence the sensor) and is counted as a passing wheel by the sensor).

[0010] Furthermore, a conventional axle counter is unable to further classify features of a passing rail vehicle (for example, magnetic track brakes, wheels, trucks, or even whether it involves a freight train, passenger train, etc.).

[0011] Figure 8 shows an example of a data set 200, which has been recorded by an axle counter sensor (or also counting point, here consisting of two sensor subsystems). In particular, the data set has a one-dimensional time characteristic (only one 2025202907   24 Apr 2025 parameter, the voltage, is measured over time). In this example, four trucks of a rail vehicle have passed the sensor. Since each truck in this example has two axles (each with two wheels, of which the one on the side of the sensor is always captured), two signals (or peaks, pulses) are always localized close together (peak cluster). In other words, the high signals correspond to the real wheels. Furthermore, in the second and third truck, lower signals are observed between the high signals. These are usually caused by magnetic track brakes.

[0012] Figure 9 shows a detailed view of a signal cluster (signal pattern) of a dual-axle truck, wherein a low signal of a magnetic track brake 204 is localized between two high signals of the wheels 203.

[0013] Conventionally, the “single thresholding” method is used for the processing of axle counter signals. This method is shown in an illustrative manner in Figure 10. For each axle counter, a predefined threshold value is configured. After recording the data set 200, the captured signal is truncated for a particular threshold value of an axle counter (reference character 210): Values above this fixed threshold value are mapped to “1” and values below this threshold value to “0”. The value “1” indicates a point in time at which a wheel passes an axle counter. After the truncation, data series 210 therefore ideally only shows wheel signals. The advantage of this method consists in that it is fast enough in order to provide real-time identification of the presence of wheels.

[0014] However, the conventional method can have a crucial disadvantage: The context is not taken into consideration. If a wheel signal does not reach the threshold value, then it cannot be captured (false negative). If the data set has noise or a non-wheel signal, caused by a magnetic track brake, for example, 2025202907   24 Apr 2025 then this signal can exceed the threshold value and an additional, non-present wheel is counted (false positive).

[0015] If the signals are close to the threshold value, it can therefore be extremely difficult to determine whether or not it involves a wheel. These are also the main reasons for “false positive” and “false negative” detection. Error cases of this kind can occur very frequently and lead to inaccurate evaluations of the axle counter data sets. This then, for example, leads to a track section being considered occupied and hindering train traffic until it is then resolved again manually.

[0016] Figure 11 shows an example of a “false positive” error in accordance with the conventional “single threshold” method. Within the box in the drawing, the first row shows the signal captured by the first sensor subsystem and the second row shows the signal captured by the second sensor subsystem. These are the two sensors in one axle counter. The third and fourth rows are the results of a conventional single threshold algorithm, which evaluates the digitized threshold value signals of the first and second sensor subsystem (corresponding to the assessment of the analog signals from the first and second row). The horizontal line in the first row indicates the specified threshold value (trigger threshold). In the example of the box, for a single threshold value, the algorithm incorrectly considers the result to be three wheels on a dual-axle truck due to the peak in the middle (caused by a magnetic track brake between the two wheels).

[0017] Figure 12 shows an example of a “false negative” error in accordance with the conventional “single threshold” method. As in Figure 11, the signals of the first sensor subsystem and the second sensor subsystem, as well as the threshold values thereof, are indicated in rows one to four. In the example of 2025202907   18 Jun 2026 the box in the drawing, in the second row, there are two high signals from wheels and a lower signal, which is caused by a magnetic track brake. The horizontal line is the specified threshold value. However, in the corresponding fourth row, which shows the threshold value signal of the second sensor subsystem, no wheel is identified. This is because the values of the signals in the second row are too low and do not exceed the threshold value.

[0018] The disadvantages explained above are not in line with efficient rail traffic. Furthermore, these disadvantages can be an obstacle to a desired advancing automation in rail traffic (with simultaneously high safety standards). Summary of the invention

[0019] It is an object of the present invention to overcome or ameliorate one or more of the disadvantages of the prior art, or to provide a useful alternative. [0019a] In one aspect of the invention, there is provided a method for the evaluation of a data set, the method having: recording of a one-dimensional time characteristic of an electrical measured variable as a data set by means of an axle counter; and evaluation of the data set by means of determination of signal features in the one-dimensional time characteristic, and classification of the signal features into signal classes in the one-dimensional time characteristic, in one step by means of a single-stage algorithm based on deep learning. [0019b] In another aspect of the invention, there is provided an apparatus for data processing, which has at least one processor, and which is configured to: obtain a one-dimensional time characteristic of an electrical measured variable recorded 2025202907   18 Jun 2026 by an axle counter as a data set; and evaluate the data set by means of determination of signal features in the onedimensional time characteristic, and classification of the signal features into signal classes in the one-dimensional time characteristic, in one step by means of a single-stage algorithm based on deep learning. [0019c] In a further aspect of the invention, there is provided an axle counter system, having: at least one axle counter for recording a one-dimensional time characteristic; and an apparatus for data processing as claimed in claim 14, which is coupled to the at least one axle counter.

[0020] There could be a need to evaluate a data set of an axle counter in a fast and reliable manner.

[0021] A method for the evaluation of a data set of an axle counter, an apparatus for data processing, an axle counter system, and a computer program product are described in the following.

[0022] In accordance with a first aspect of the invention, an (in particular computer-implemented) method for the evaluation of a data set (of an axle counter) is described, the method having: i) recording (or measurement / determination) of a one-dimensional time characteristic (only one measurement variable is measured / determined over time, for example, the voltage) as a 2025202907   24 Apr 2025 data set by means of an axle counter (or axle counter sensor), and ii) (computer-implemented) evaluation of the data set by means of a) determination / detection of signals (in particular determination of signal features, further in particular signal points, for example, position of peak tips) in the onedimensional time characteristic, and b) classification of the signals (in particular assignment of signal features and / or signal patterns to signal classes) in the one-dimensional time characteristic (for example, as wheel, truck, etc.) c) in one step (in a single stage) by means of a single-stage algorithm based on artificial intelligence (for example, deep learning) (or single-stage detector) (for example, YOLO or SSD).

[0023] In accordance with a second aspect of the invention, an apparatus for data processing (for example, a computer, a control system, an axle counter evaluation unit, etc.) is described, which has at least one processor, and which is configured to: i) obtain a one-dimensional time characteristic recorded by an axle counter (in particular wherein the data set is / has been converted from analog to digital), and ii) evaluate the data set by means of a) determining signals in the one-dimensional time characteristic, and b) classifying the signals in the one-dimensional time characteristic, c) in one step by means of a single-stage algorithm based on deep learning.

[0024] In accordance with a third aspect of the invention, an axle counter system (or arrangement) is described, which has: 2025202907   24 Apr 2025 i) at least one axle counter (in particular having at least one axle counter sensor) for recording a one-dimensional time characteristic, and ii) an apparatus for data processing as described above (in particular configured as an axle counter evaluation unit), which is coupled (in a communicative manner) to the at least one axle counter (for example, interconnected, remote operation, coupled in a wired or wireless manner, etc.).

[0025] In accordance with a fourth aspect of the invention, a computer program product is described, which has commands which, when the program is executed by a computer, prompt it to carry out the method described above.

[0026] In the context of this document, the term “artificial intelligence (AI)” in particular can relate to computer-based approaches for emulating cognitive functions of a human mind, in particular learning and problem-solving. A large number of different mathematical algorithms and calculation models have been developed, in order to implement AI functionalities, for example “machine learning”, in particular by means of neural networks. The main aim of said approaches can be considered to be developing and improving a model, by the algorithm being trained with training data, so that a learning effect occurs and the problem-solving capability of the algorithm is improved over time. This can take place with or without human intervention (for example, improvement).

[0027] In the context of this document, the term “deep learning” (multi-level / deep learning) in particular can relate to a specific form of machine learning (as a specific form of AI), in particular neural networks. In particular, in the case of deep learning algorithms, artificial neural networks with numerous intermediate layers (hidden layers) between input layer and output layer are used, whereby a comprehensive inner 2025202907   24 Apr 2025 structure can be developed. In one example, deep learning can be particularly well suited to an algorithm improving itself through training (in a self-adaptive manner).

[0028] In the context of this document, the term “convolutional neural network (CNN)” in particular can relate to a specific deep learning approach. In one example, the structure of a convolutional neural network can have one or more convolutional layers, followed by one or more pooling layers. After a few recurring units, finally a fully connected layer can be provided. Due to the reduction of the amount of data in the convolutional and pooling layers, CNNs can be particularly well suited to processing large amounts of data, for example, images with particularly high resolution.

[0029] In the context of this document, the term “singlestage algorithm” in particular can relate to an algorithm which is based on AI (in particular on deep learning, further in particular on CNN) and is configured to perform a determination and classification of signals in one step (or in a single stage). In an exemplary example, a single-stage algorithm of this kind can be implemented by means of an established algorithm such as YOLO or SSD (forms for single-stage algorithms). Conventional approaches for the evaluation of twodimensional data such as images (such as CNN) use a two-stage approach: First, a determination of the objects is performed and these objects are only classified in a following step. In contrast to this, the single-stage algorithm performs the determination and classification in one step, which makes it possible to save processing time (and thus also computer resources). For example, the single-stage algorithm can use a frame to break up an image into sections and analyze it in sections. Each section can then be considered once (“shot”) and analyzed in one step. 2025202907   24 Apr 2025

[0030] In the context of this document, the term “one dimensional time characteristic” (or time series) in particular can relate to a representation or analysis of data that only has one dimension, namely the time. By means of this representation, the development / change of a particular measurement variable can be considered over time (without spatial or other variables being included). In the case of an axle counter, for example, it is possible to consider an electrical parameter such as the voltage over time, in order to analyze the presence of axles (in particular wheels, trucks, etc.) in the time characteristic.

[0031] In the context of this document, the term “determination of signals” in particular can relate to signals being identified or detected within the data set, in particular on the basis of (unique) “signal features”. In one example, the detection of a peak (for example, a wheel axle) can be understood as a determination of the signal feature “local maximum”. In one exemplary embodiment, a signal feature can be localized by means of a signal point (here, for example, the peak tip). In one exemplary embodiment, on the basis of one or more signal features, it is possible to identify a “signal pattern”, for example, the detection of two or three signal features (in a signal cluster, for example, a truck) as the determination of the signal features “two / three local maxima”. In other words, the determination can relate to detecting whether and where a unique / unmistakable feature is located in a signal (characteristic) (localization and extraction; where is it?) (see also, for example, Figures 2, 3 and 4).

[0032] In one exemplary embodiment, the “determination” can be carried out in a more precise variant, which can be referred to as “segmentation”. In this context, a precise localization of multiple signal features can take place, for example, multiple peaks in a truck signal cluster (working out fine structures). 2025202907   24 Apr 2025

[0033] In the context of this document, the term “classification (assignment) of signals” in particular can relate to one or more signals (or signal features and / or signal patterns) being assigned to a particular signal class (signal type). In order to classify signals, these are detected / determined on the basis of signal features (conventionally, the steps “determination” and “classification” take place in two passes / stages). In one example, the assignment of an individual signal feature “peak” to the signal class “wheel” can be a classification. Furthermore, for example, the assignment of a signal pattern “high peak-low peak-high peak” to the signal class “truck with magnetic track brake” can be a classification.

[0034] In accordance with one exemplary embodiment, the invention can be based on the idea that a data set of an axle counter (the one-dimensional time characteristic) can be evaluated in a fast and reliable manner, if the determination and classification of the signals takes place in one step by means of a single-stage AI algorithm.

[0035] In contrast to this, conventional approaches for the evaluation of axle counter data sets (also by means of AI) are based on a detection taking place first and a classification of the signals taking place after this (in a second pass).

[0036] Algorithms for analyzing images (i.e. two-dimensional data sets) are known, such as convolutional neural networks, for example. These networks are conventionally used for demanding 2D and 3D representations, because they can be particularly advantageous for processing large amounts of data such as complex images with high resolution. Usually, these algorithms use the concept of “bounding boxes”, in order to be able to identify particular objects in the images. Further algorithms 2025202907   24 Apr 2025 are known (see YOLO, SSD), which can perform such a 2D / 3D image analysis in a single stage.

[0037] The inventors have now identified that such single stage algorithms for the analysis of two-dimensional images can be transferred to one-dimensional data sets such as the time characteristic of an axle counter in a surprisingly fast, effective and reliable manner. Such an approach has previously been unknown, because algorithms such as YOLO or SSD are just optimized for two-dimensional images and an advantageous implementation with regard to one-dimensional (measurement) data is therefore not realized.

[0038] According to the invention, the simultaneous classification and detection of the one-dimensional signal (wheels, brakes, etc.) by means of the single-stage (“singleshot”) technique can be used.

[0039] In one example, the method described can identify and classify (and segment) various kinds of train features, such as magnetic track brakes, wheels, trucks and entire cars, in contrast to conventional evaluation approaches which, for example, only identify wheels and cannot distinguish between other features (such as magnetic track brakes).

[0040] In one example, the described method can contribute to considerably increasing the robustness to signal noise, as greater intervals of the signal are used for detection and classification. Moreover, the method can calculate a “quality measure” of the underlying signal, which can be used in higher- level control systems to calculate reliability values.

[0041] In one example, the method described can enable considerably improved accuracy values compared to current methods according to the prior art. It is thereby possible to 2025202907   24 Apr 2025 obtain a particularly reliable evaluation in a short time, which in addition to the determination also enables a simultaneous classification of the captured signals. The method described can be performed in a particularly fast and cost-effective manner from a computing perspective, and can be implemented directly into existing systems. Exemplary embodiments

[0042] In accordance with one exemplary embodiment, the single-stage algorithm has a convolutional neural network (CNN). This can have the advantage that an implementation can be achieved by means of an efficient and established algorithm. CNNs, however, have been known to date from the analysis of images (two-dimensional data); an advantageous application to one-dimensional measurement data, in particular in terms of the specific application of axle counters, has not been used to date.

[0043] In accordance with one exemplary embodiment, the single-stage algorithm has a YOLO algorithm (or a derivative thereof) or an SSD algorithm (or a derivative thereof). This can have the advantage that it is possible to achieve a fast and reliable implementation by means of tested and reliable means.

[0044] YOLO (You Only Look Once) is a popular variant which is frequently used in computer vision applications. YOLO is an object identification algorithm which processes an entire image in a single pass and predicts bounding box coordinates and class probabilities directly. The algorithm can achieve real-time performance by the image being divided into a grid and the identification of objects being performed for each grid cell.

[0045] SSDs (single shot detectors) use multiple feature maps with different resolutions, in order to identify objects with 2025202907   24 Apr 2025 different sizes efficiently in real time. An SSD is known for its accuracy, particularly during the identification of small objects, which the algorithm uses for a large number of computer vision applications.

[0046] In accordance with one exemplary embodiment, the single-stage algorithm is performed in sections (section by section) along the one-dimensional time characteristic. In accordance with one exemplary embodiment, the sections have a particular section width (in other words, at least two sections with the same dimensioning are used). In accordance with one exemplary embodiment, the section width is partially variable (in other words, at least two sections with different dimensioning are used).

[0047] Single-stage algorithms such as YOLO or SSD divide a two-dimensional image into a grid and analyze the grid cells. In order to transfer this approach to one-dimensional data, according to the invention it is possible to use sections. This can be in a line (with or without a spacing between the sections), so that the algorithm can analyze the data set section by section. In one preferred example, the sections only require one pass of the algorithm (in a single-stage manner). An illustrative example of the sections is shown in Figures 5 and 6, for example. The section width can be (partially) varied, for example, wider for a slower train and narrower for a faster train.

[0048] In accordance with one exemplary embodiment, the determination includes: identification of signal features in the one-dimensional time characteristic. In one exemplary embodiment, the determination includes: identification of signal patterns (based on signal features) in the one-dimensional time characteristic. Object identification algorithms usually use the concept of bounding boxes, in order to localize objects (for 2025202907   24 Apr 2025 example, a car) in an image. The present case, however, works with one-dimensional data, so that this concept is not applicable.

[0049] It has been identified by the inventors, however, that it is possible to use signal features, particularly in the form of individual points, and / or signal patterns analogous to bounding boxes in an efficient and reliable manner. In one example, a local maximum (peak tip) can be identified as a signal feature and this point can be determined by means of localization (see Figures 2 and 3, for example).

[0050] In accordance with one exemplary embodiment, the determination of signal patterns includes: identification of a (truck-associated) signal pattern based on five signal points.

[0051] This variant can be efficient specifically for the identification of trucks. For example, the five signal points could be set as follows: i) start of the truck, ii) position of first wheel (for example, peak tip), iii) position of magnetic track brake (present or not), iv) position of second wheel (for example, peak tip), and v) end of the truck. If these five signal features are set, then the signal pattern “truck” can be determined. This can simultaneously result in the classification into the signal class “truck”.

[0052] In accordance with one exemplary embodiment, the determination of signal features includes: identification of a (wheel-associated) signal pattern based on at least one signal point (and the level / position thereof).

[0053] In this variant, each signal feature can correspond to one signal pattern, for example, “wheel” or “brake” (here, the 2025202907   24 Apr 2025 level of the signal point can be crucial). Through the identification of the signal pattern “wheel”, in one example, the classification into the signal class “wheel” is also produced.

[0054] In accordance with one exemplary embodiment, the method includes: assessing whether a section contains (relevant) information, and i) using sections which contain information and / or ii) eliminating sections which do not contain information. In this connection, the term “information” in particular can comprise at least one signal feature. The efficiency can be increased if only sections which contain information are analyzed (taken into consideration).

[0055] In accordance with one exemplary embodiment, the single-stage algorithm is performed in anchor regions along the one-dimensional time characteristic in addition to the sections. In other words, not only are the sections defined and taken into consideration, but anchor regions can additionally be defined and taken into consideration at particular positions.

[0056] In one example, the sections are used homogeneously according to a fixed pattern (analogously to a grid), while the anchor regions are used non-homogeneously at particular positions. By means of the anchor regions, it is possible for regions of interest (high density of signal features) to be analyzed more specifically. See the illustration in Figure 7 for more.

[0057] Single-stage algorithms (for example, YOLO) use “grid cells” to identify characteristic patterns (for example, a car). The grid cells can be derived by simply dividing the image into multiple smaller fields. Sometimes, this single-stage object identification algorithm additionally adds the concept of anchor boxes. In this case, different sizes of the image are extracted 2025202907   24 Apr 2025 at grid positions (for example, for training and inference). In one exemplary embodiment, the time characteristic is divided into fields with predefined fixed length / width (sections). As in many single-stage approaches, according to the invention it is also possible to apply the concept of the use of anchor regions.

[0058] In accordance with an exemplary embodiment, the anchor regions have at least two different dimensionings. In one example, the anchor regions can be provided in the form of frames / boxes. In this context, the sizes can differ, for example.

[0059] In accordance with one exemplary embodiment, at least one anchor region is smaller / larger in at least one dimension than a corresponding section. There can be a particular level of flexibility as a result.

[0060] A fine grid (compared to the signals to be detected) (or a narrower section width) and larger anchor regions can result in redundant detections. In such cases, it can be necessary to suppress identifications with lower reliability (non-maximum suppression).

[0061] In accordance with one exemplary embodiment, the method includes: providing an output, which has the signal class (classification or classification probability; what is it?) and at least one signal feature (determination /

[0062] detection) (or the position of at least one signal feature; where is it?), in particular in vector format. In other words, the output of the single-stage algorithm has both the determination and the classification. Conventionally, this is carried out in two stages. In the single-stage algorithm, however, the information of the determination and the 2025202907   24 Apr 2025 classification can be interlinked with one another, for example as connected layers (see also Figure 1).

[0063] In accordance with one exemplary embodiment, the classification includes a categorization into / assignment to a signal class (in particular based on at least one signal feature and / or a signal pattern). These classes can be defined according to the respective application and the underlying AI algorithm can be trained accordingly. A signal class can have one or more distinguishing features, which are indicative of this class. A signal class can relate to individual signals and / or signal clusters (two or more signals, which are associated with one another).

[0064] In accordance with one exemplary embodiment, the signal classes include at least one of the following: a wheel, a magnetic track brake, two wheels, two wheels and one magnetic track brake, two wheels and no magnetic track brake, single-axle truck, dual-axle truck. This is only an exemplary selection of signal classes, which can be particularly frequent or relevant (to safety) in the field of rail traffic.

[0065] In accordance with one further exemplary embodiment, the method further includes: training the single-stage algorithm with multiple one-dimensional time characteristics of an axle counter (determined in an experimental and / or theoretical manner). In accordance with one further exemplary embodiment, the training further includes: learning signal patterns and / or signal features, which are indicative of particular signal classes.

[0066] In this context, the training data can comprise measured (real) and / or simulated data sets. Depending on the application, particular training sets can be particularly suitable. 2025202907   24 Apr 2025

[0067] In accordance with one further exemplary embodiment, the method further includes: determination of at least one of the following, based on the evaluation: a velocity, a length, a vehicle type, an anomaly, in particular a defect, necessary maintenance.

[0068] As a result, the method described can enable new possibilities for the overall control system of rail vehicles, in the following few exemplary embodiments: i) The identification and classification of trucks could be used to calculate the train velocity (or the train type and the dimensions are known, meaning that it is possible to calculate the velocity); ii) The length and the type of the train (for example, freight train, passenger train) could be guessed. The number of cars could be counted; iii) Anomalies (for example, defects on the wheels and / or magnetic track brakes) could be identified and reported; iv) Predictive maintenance could be based on the signal recorded by the axle counter.

[0069] In accordance with one further exemplary embodiment, the method further includes: preprocessing the data set to remove noise, in particular by means of smoothing and / or downsampling. This can have the advantage that signals with low resolution can also be used.

[0070] In accordance with one further exemplary embodiment, the method is used in connection with / in the context of railbound vehicles and / or rail infrastructure.

[0071] The axle counter application in rail traffic has already been discussed in detail above. Furthermore, the field of rail vehicles has particularly high requirements for safety and reliability and corresponding standards. For this reason, 2025202907   24 Apr 2025 the use of the method for validating evaluations can be particularly important and advantageous in this connection.

[0072] In accordance with one further exemplary embodiment, the single-stage algorithm / classifier is particularly fast (faster than other methods based on deep learning, for example Unet). In one example, the algorithm is based on single-shot classifiers (such as the YOLO model, which is used for object identification in computer vision).

[0073] In accordance with one further exemplary embodiment, a true neural end-to-end network is described, which predicts bounding boxes (signal features) and class probabilities at once. In one example, it is not possible to post-process the data (i.e. to calculate classes or coordinates based on the prediction of the neural network), which can make the overall algorithmic system more robust.

[0074] It should be noted that embodiments of the invention have been described while making reference to different subject matters. In particular, some embodiments have been described while making reference to method claims, while other embodiments have been described while making reference to apparatus claims. A person skilled in the art, however, will understand from the above and the following description that, unless indicated otherwise, in addition to any combination of features that belong to one type of subject matter, any combination of features that relate to different subject matters is considered to be disclosed by this document. In particular, this also includes between features of the method claims and features of the apparatus claims.

[0075] The aspects defined above and further aspects of the present invention result from the examples of the embodiments described in the following and are explained with reference to 2025202907   24 Apr 2025 the examples of the embodiments. In the following, the invention is described in further detail with reference to embodiments, to which the invention is not restricted, however. Brief description of the drawings

[0076] Figure 1 schematically shows a method in accordance with an exemplary embodiment of the invention.

[0077] Figure 2 shows a determination of signal patterns based on five signal points, in accordance with an exemplary embodiment of the invention.

[0078] Figure 3 shows a determination of signal patterns based on five signal points, in accordance with a further exemplary embodiment of the invention.

[0079] Figure 4 shows a determination of signal patterns based on individual signal points, in accordance with an exemplary embodiment of the invention.

[0080] Figure 5 shows a performance of the single-stage algorithm in sections along the one-dimensional time characteristic with five signal points, in accordance with an exemplary embodiment of the invention.

[0081] Figure 6 shows a performance of the single-stage algorithm in sections along the one-dimensional time characteristic with an individual signal point, in accordance with an exemplary embodiment of the invention.

[0082] Figure 7 shows a performance of the single-stage algorithm along the one-dimensional time characteristic in sections and anchor regions, in accordance with an exemplary embodiment of the invention. 2025202907   24 Apr 2025

[0083] Figures 8 and 9 each show a data set of an axle counter.

[0084] Figures 10 to 12 illustrate a conventional threshold value-based method and the disadvantages thereof.

[0085] Figure 13 shows an axle counter system in accordance with an exemplary embodiment of the invention. Detailed description of the drawings

[0086] The representations in the drawings are schematic. It is noted that similar or identical elements or features are provided with the same reference characters or with reference characters in different images, the reference characters only differing from the corresponding reference characters within the first digit. In order to avoid unnecessary repetitions, elements or features that have already been explained with regard to a previously described embodiment are not explained once again at a later point in the description.

[0087] Moreover, spatially relative terms such as “front” and “back”, “top” and “bottom”, “left” and “right”, etc. are used to describe the relationship between an element and another element, as represented in the images. Thus, the spatially relative terms can pertain to orientations that are used, which deviate from the orientation represented in the images. It is apparent that these spatially relative terms merely relate to a simplification of the description and to the orientation shown in the images and are not necessarily restrictive, as an apparatus in accordance with one embodiment of the invention can assume other orientations than those shown in the images, particularly when they are used. 2025202907   24 Apr 2025

[0088] Figure 1 schematically shows a method in accordance with an exemplary embodiment of the invention. Firstly, a data set 110 (input) is provided, which has been recorded / measured by an axle counter (see Figure 13, for example). The data set has a one-dimensional time characteristic (see Figures 8 and 9, for example). The evaluation of this data set 110 is performed by means of a single-stage AI algorithm (single-stage / shot algorithm / detector) 150.

[0089] The AI algorithm 150 can be constructed as a CNN with up to twenty layers 151. At the end of the CNN, for example, there are two fully connected layers 152, which provide the following output: the class probability and the position of the signal features.

[0090] In the present context, signal features can be used to localize a particular signal position. While a bounding box is used to define the position and size of an object in the twodimensional region, according to the invention this can take place point-by-point in the one-dimensional region or signal features can be used in the form of signal points. One or more signal points can be indicative of a particular signal pattern. The signal pattern can be used in an analogous manner to the bounding box. In one exemplary embodiment, the determination of signals can relate to “where is it?”, while the classification of signals can relate to “what is it?”.

[0091] In the example shown, a local maximum has been identified as a peak tip and a signal feature 112 has been set by means of a signal point (determination of signals). Based on the signal feature, the signal pattern 111 “wheel” has been identified (here as an analogy to a bounding box), from which the assignment to the signal class “wheel” is achieved (classification of signals). The evaluation of the data set is thus achieved by means of the determination of signals in the 2025202907   24 Apr 2025 one-dimensional time characteristic and classification of the signals in the one-dimensional time characteristic in one step by means of a single-stage algorithm based on deep learning.

[0092] Figure 2 shows a determination of signal features based on five signal points, in accordance with an exemplary embodiment of the invention. By means of the five signal features, it is possible to identify a particular signal pattern, here a truck-associated signal pattern. The signal cluster shown is located within a single section 160, for simpler representation.

[0093] The following signal points 112 can be set: i) start of the truck, ii) position of first wheel (signal feature high peak 130), iii) position of magnetic track brake 135 (signal feature low peak 135), iv) position of second wheel (signal feature high peak), and v) end of the truck.

[0094] All five signal points can be set and the signal pattern “truck with magnetic track brake” 131 can be identified, which simultaneously results in the associated classification into the signal class “truck with magnetic track brake”.

[0095] Figure 3 shows a determination of signal features based on five signal points, in accordance with a further exemplary embodiment of the invention. This example is similar to that in Figure 2. The difference consists in there being no peak present for the magnetic track brake at position iii). For this reason, the five signal points result in the signal pattern “truck without magnetic track brake” 132, which simultaneously 2025202907   24 Apr 2025 results in the associated classification (signal class “truck without magnetic track brake”).

[0096] Figure 4 shows a determination of signal features based on one signal point in each case and the level / position thereof, in accordance with an exemplary embodiment of the invention. By means of one signal feature 112 in each case, it is possible to identify a particular signal pattern 111. For simpler representation, each signal pattern 111 is located in its own section 160-162.

[0097] In the first section 160, a local maximum is identified and the signal feature “high peak” 130 can be set. This leads to the signal pattern “wheel”. In the second section 161, a local maximum is identified and the signal feature “low peak” 135 can be set. This leads to the signal pattern “magnetic track brake”. In the third section 162, a local maximum is identified and the signal feature “high peak” 130 can be set. This leads to the signal pattern “wheel”. The signal pattern sequence “wheel-magnetic track brake-wheel” in this context leads to the assignment to the signal class “truck with magnetic track brake”.

[0098] While five signal features 112 are assigned a common signal pattern 111 in the examples in Figures 2 and 3, each signal feature 112 is assigned an individual signal pattern 111 in Figure 4. In both cases, however, only one signal class is present.

[0099] Figure 5 shows a performance of the single-stage algorithm along the one-dimensional time characteristic through three sections 160-162 with the application of the five signal point identification, in accordance with an exemplary embodiment of the invention. The five signal points can be set as described for Figures 2 and 3. 2025202907   24 Apr 2025

[00100] In the case of the second section 161, a correct classification of the signal curve and the setting of the five signal points (“regression box”) were possible, despite only part of the signal curve being present in the second section 161. In the first section 160 and in the second section 162, however, a correct classification is not possible: the selection would be ambiguous from the given data.

[0100] Figure 6 shows a performance of the single-stage algorithm along the one-dimensional time characteristic through three sections 160-162 (these would be the “shots”) with the application of the one signal point identification, in accordance with an exemplary embodiment of the invention.

[0101] In the first three sections 160-162, a classification (with the “regression windows” that have been calculated by the neural network) of the signal feature or signal pattern 111 is possible (wheel or brake). The fourth section 163, however, does not contain a feature (signal feature) and is therefore considered to be a section which does not contain information, in particular is discarded.

[0102] Figure 7 shows a performance of the single-stage algorithm along the one-dimensional time characteristic in sections 160 and anchor regions 170, in accordance with an exemplary embodiment of the invention. In contrast to the other examples, a very fine sampling grid (small section width) is used. Anchor regions / boxes 170 with greater dimensions are additionally used. This can lead to multiple identifications, which can be handled with a non-maximum suppression algorithm, in order to use only the identifications with the highest confidence interval. It is possible to use anchor regions with different dimensions (see 170-172), in order to cover different 2025202907   24 Apr 2025 signal levels. Anchor regions 175 with non-ideal placement are unable to identify any signal feature in this example.

[0103] Figure 13 shows an axle counter system 100 in accordance with an exemplary embodiment of the invention. This has an axle counter sensor apparatus 102 (for example, one, two or more sensors, for example, dual sensor system), which is fastened to a rail 101 and can record / measure the data sets (one-dimensional time series) described above. In this example, the sensor apparatus 102 is connected to an axle counter evaluation unit 103 by cable (alternatively without cables). This evaluation apparatus 103 is configured to obtain the measured data set and to evaluate or validate it according to the method described above. Alternatively, the evaluation can also be performed in remote operation.

[0104] It should be noted that the term “having” does not exclude other elements or steps and the use of the article “a” does not exclude a plurality. Elements that are described in connection with different embodiments can also be combined. It should also be noted that reference characters in the claims should not be interpreted such that they restrict the scope of the claims.

[0105] Independent of the grammatical term usage, individuals with male, female or other gender identities are included within the term. 2025202907   24 Apr 2025 Reference characters 100 axle counter system 101 rail 102 axle counter sensor 103 axle counter evaluation unit 110 data set, time series 111 signal pattern, signal class 112 signal feature, signal point 130 signal feature, high peak 131 signal pattern, truck with magnetic track brake 132 signal pattern, truck without magnetic track brake 135 signal feature, low peak 150 single-stage algorithm 151 layers 152 connected layers 160-163 section 170-172 anchor region 175 further anchor region 200 data set 201 signal cluster, truck without magnetic track brake 202 signal cluster, truck with magnetic track brake 203 signal, wheel axle 204 signal, magnetic track brake

Claims

1. A method for the evaluation of a data set, the method having:recording of a one-dimensional time characteristic of anelectrical measured variable as a data set by means of an axlecounter; andevaluation of the data set by means ofdetermination of signal features in the one-dimensional time characteristic, andclassification of the signal features into signal classes in theone-dimensional time characteristic, in one step by means of a single-stage algorithm based on deeplearning.

2. The method as claimed in claim 1,wherein the single-stage algorithm has a convolutionalneural network, CNN.

3. The method as claimed in claim 1 or 2,wherein the single-stage algorithm has a YOLO algorithm ora derivative thereof; orwherein the single-stage algorithm has an SSD algorithm ora derivative thereof.

4. The method as claimed in any one of the preceding claims,wherein the single-stage algorithm is performed in sectionsalong the one-dimensional time characteristic.

5. The method as claimed in any one of the preceding claims,wherein the determination includes:identification of a signal pattern based on a plurality of signal features.2025202907   18 Jun 20266. The method as claimed in claim 5, further having: identification of a truck-associated signal pattern basedon five signal points; oridentification of a wheel-associated signal pattern basedon at least one signal point.

7. The method as claimed in any one of the preceding claims, wherein the single-stage algorithm is performed in at least one anchor region along the one-dimensional time characteristic in addition to the sections, in particular wherein the anchor regions have at least two differentdimensionings; and / or wherein at least one anchor region is larger in at least onedimension than a corresponding section.

8. The method as claimed in any one of the preceding claims, wherein the classification includes a categorization into asignal class, in particular based on an identified signalpattern.

9. The method as claimed in claim 8, wherein the signal classes include at least one of thefollowing: a wheel, a magnetic track brake, two wheels, twowheels and one magnetic track brake, two wheels and no magnetictrack brake, single-axle truck, dual-axle truck.

10. The method as claimed in one of the preceding claims, including:providing an output, which has at least one signal featureand an associated signal class, in particular in vector format.

11. The method as claimed in any one of the preceding claims,2025202907   18 Jun 2026further including:training the single-stage algorithm with multiple onedimensional time characteristics of an axle counter, in particular wherein the training further includes: learning signal features and / or signal patterns, which are indicative of particular signal classes.

12. The method as claimed in any one of the preceding claims, further including:determination of at least one of the following, based on the evaluation: a velocity, a length, a vehicle type, an anomaly, in particular a defect, necessary maintenance.

13. The method as claimed in any one of the preceding claims, wherein the method is used in rail-bound vehicles and / or rail infrastructure.

14. An apparatus for data processing, which has at least one processor, and which is configured to:obtain a one-dimensional time characteristic of an electrical measured variable recorded by an axle counter as a data set; andevaluate the data set by means ofdetermination of signal features in the one-dimensional time characteristic, andclassification of the signal features into signal classes in theone-dimensional time characteristic, in one step by means of a single-stage algorithm based on deeplearning.

15. An axle counter system, having:at least one axle counter for recording a one-dimensional time characteristic; and2025202907   18 Jun 2026an apparatus for data processing as claimed in claim 14, which is coupled to the at least one axle counter.Siemens Mobility GmbHPatent Attorneys for the Applicant / Nominated Person SPRUSON & FERGUSON

Images