One-step evaluation of axle counter time curve

Abstract

A method for evaluating a data set (110) is described, comprising: i) recording a one-dimensional time series as a data set (110) using an axis counter (100); and ii) evaluating the data set (110) by identifying signals in the one-dimensional time series and classifying the signals in the one-dimensional time series in one step using a one-stage algorithm (150) based on deep learning.

Description

Technical field

[0001] The invention relates to a method for evaluating axle counter data, wherein a one-dimensional time history is recorded by the axle counter and evaluated by identifying / detecting signals in the one-dimensional time history and classifying the signals in the one-dimensional time history in a single step (single-stage) using a single-stage deep learning-based algorithm. The invention further relates to a data processing device and an axle counter system.

[0002] The invention can therefore relate to the technical field of evaluating data sets from an axle counter, particularly with regard to rail vehicles. Technical background

[0003] Axle counters are a common feature in the field of rail vehicles and rail infrastructure, forming an essential part of the monitoring system. They track, for example, whether a wheel or axle (of a train) passes a track and how many wheels have passed. This can be crucial for many applications, such as track occupancy detection and, in particular, for safety within a rail infrastructure. Typically, sections of track are equipped with two axle counter sensors (so-called counting points), one at the beginning and the other at the end of the section. Whether a track section is determined to be clear or occupied depends primarily on the number of wheels detected by the two axle counter sensors. A higher-level evaluation unit (this can be, for example, a control unit) then processes the data.The system (which, together with the axle counter sensors, forms an axle counter system) evaluates the signals detected by the sensors and derives the status of the track section from them. Only if a section is actually recognized as clear may a train pass through that track section.

[0004] In the context of this document, the term "axle counter" can encompass both the unit "sensor (or counting point) and evaluation unit" and refer only to the evaluation unit (the actual axle counter). If the term "axle counter sensor" is used to refer specifically to the sensor, and "evaluation unit" is used to refer specifically to the evaluation unit. If the axle counter sensor and the axle counter evaluation unit are to be referred to together, the term "axle counter system" can be used.

[0005] However, particularly high demands are placed on safety and reliability in the rail vehicle sector. Accordingly, in numerous applications, evaluated data sets from an axle counter, especially for the automation of rail transport, can only be used if exceptionally high reliability has been demonstrated.

[0006] Currently, axle counters cannot achieve 100% counting accuracy in practical applications. This is due, firstly, to the complex signals captured by the axle counter sensors, and secondly, to signal processing methods that are not precise enough to handle complex signals. This is partly because the signals may be affected by interference (noise) and because varying speeds can result in numerous waveforms. In extreme cases, trains may stop at the sensor and / or reverse direction. This would not pose a safety issue, as the counts are always conservative (in the sense that the safety margin is increased), but rather an availability problem.

[0007] Currently, the counting of wheels (or the evaluation of the data set) is carried out very conservatively to ensure the safety of the rail vehicles. Thus, a section of track is considered "occupied" and not "free" if the counting accuracy is not 100% certain.

[0008] Another limitation of conventional axle counters is that, although they can detect the occurrence of wheels with a high degree of certainty, even state-of-the-art products can only deliver such results at the cost of false positives (meaning that although no wheel passes the sensor, a signal is detected (e.g., by masses of metal affecting the sensor) and is counted by the sensor as if a passing wheel were passing).

[0009] Furthermore, a conventional axle counter cannot further classify characteristics of a passing rail vehicle (e.g., magnetic track brakes, wheels, bogies, or whether it is a freight train, passenger train, etc.).

[0010] Figure 8Figure 200 shows an example of a data set recorded by an axle counter sensor (or counting point, here consisting of two sensor subsystems). The data set exhibits a one-dimensional time history (only one parameter, voltage, is measured over time). In this example, four bogies of a rail vehicle have passed the sensor. Since each bogie in this example has two axles (each with two wheels, of which the one on the sensor side is always recorded), two signals (or peaks, pulses) are always located close together (peak cluster). In other words, the high signals correspond to the actual wheels. Furthermore, lower signals can be observed between the high signals in the second and third bogies. These are usually caused by magnetic track brakes.

[0011] Figure 9shows a detailed view of a signal cluster (signal pattern) of a two-axle bogie, with a low signal of a magnetic track brake 204 located between two high signals of the wheels 203.

[0012] Conventionally, the "single-thresholding" method is used to process axle counter signals. This method is used in Figure 10This is clearly illustrated. A predefined threshold is configured for each axle counter. After recording data set 200, the captured signals are truncated at a specific axle counter threshold (reference 210): Values ​​above this fixed threshold are mapped to "1" and values ​​below the threshold to "0". The value "1" indicates the point in time at which a wheel passes an axle counter. Ideally, data set 210 thus shows only wheel signals after truncation. The advantage of this method is that it is fast enough to provide real-time detection of the presence of wheels.

[0013] However, the conventional method has a crucial disadvantage: it fails to consider the context. If a wheel signal does not reach the threshold, it cannot be detected (false negative). If the data set contains noise or a non-wheel signal, for example, caused by a magnetic track brake, this signal can exceed the threshold, and an additional, non-existent wheel will be counted (false positive).

[0014] When the signals are close to the threshold, it can be extremely difficult to determine whether they indicate a wheel or not. This is also the main reason for false positives and false negatives. Such errors can occur quite frequently and lead to inaccurate evaluations of the axle counter data. This can then result, for example, in a track section being considered occupied and obstructing train traffic until it is manually cleared.

[0015] Figure 11This shows an example of a false positive error according to the conventional single-threshold method. Within the drawn box, the first line shows the signal captured by the first sensor subsystem, and the second line shows the signal captured by the second sensor subsystem. These are the two sensors in an axle counter. The third and fourth lines are the results of a conventional single-threshold algorithm that evaluates the digitized threshold signals of the first and second sensor subsystems, respectively (corresponding to the evaluation of the analog signals from the first and second lines). The horizontal line in the first line indicates the set threshold (trigger threshold). In the example shown in the box, due to the peak in the middle (caused by a magnetic track brake between the two wheels), the algorithm incorrectly interprets the result as three wheels on a two-axle bogie.

[0016] Figure 12 This shows an example of a "false negative" error according to the conventional "single threshold" method. As in Figure 11 Rows one through four show the signals from the first and second sensor subsystems, along with their threshold values. In the example shown in the box, the second row contains two high signals from wheels and one lower signal caused by a magnetic track brake. The horizontal line represents the defined threshold. However, in the corresponding fourth row, which shows the threshold signal from the second sensor subsystem, no wheel is detected. This is because the signal values ​​in the second row are too low and do not exceed the threshold.

[0017] The disadvantages outlined above are incompatible with efficient rail transport. Furthermore, these disadvantages can hinder the desired progressive automation of rail transport (while maintaining high safety standards). Summary of the invention

[0018] There might be a need to quickly and reliably evaluate a data set from an axle counter.

[0019] A method for evaluating a data set from an axle counter, a device for data processing, an axle counter system, and a computer program product are described below.

[0020] According to a first aspect of the invention, a (particularly computer-implemented) method for evaluating a data set (of an axle counter) is described, comprising the method: i) Recording (or measuring / determining) a one-dimensional time series (only one measurand is measured / determined over time, e.g., voltage) as a data set using an axle counter (or axle counter sensor), and ii) (computer-implemented) evaluation of the data set by a) determining / detecting signals (in particular, determining signal characteristics, further, in particular signal points, e.g., the position of peak peaks) in the one-dimensional time series and b) classifying the signals (in particular, assigning signal characteristics and / or signal patterns to signal classes) in the one-dimensional time series (e.g., as wheel, bogie, etc.) c) in one step (single-stage) using a single-stage algorithm (or single-stage detector) based on artificial intelligence (e.g., deep learning) (e.g., YOLO or SSD).

[0021] According to a second aspect of the invention, a device for data processing (e.g. a computer, a control system, an axle counter evaluation unit, etc.) is described, which has at least one processor and is configured for: i) Obtaining a one-dimensional time series recorded by an axle counter (in particular, where the data set is / was converted from analog to digital), and ii) evaluating the data set by a) identifying signals in the one-dimensional time series and b) classifying the signals in the one-dimensional time series, c) in one step using a single-stage algorithm based on deep learning.

[0022] According to a third aspect of the invention, an axle counter system (or arrangement) is described which comprises: i) at least one axle counter (in particular comprising at least one axle counter sensor) for recording a one-dimensional time series, and ii) a device for data processing as described above (in particular configured as an axle counter evaluation unit) which is (communicatively) coupled with the at least one axle counter (e.g. built together, remote operation, wired or wirelessly coupled, etc.).

[0023] According to a fourth aspect of the invention, a computer program product is described which has instructions which, when the program is executed by a computer, cause it to execute the method described above.

[0024] In the context of this document, the term "artificial intelligence (AI)" can refer specifically to computer-based approaches to mimicking the cognitive functions of the human mind, particularly learning and problem-solving. A variety of mathematical algorithms and computational models have been developed to implement AI functionalities, such as "machine learning," especially through neural networks. The main purpose of these approaches can be seen as developing and improving a model by training the algorithm with training data, thus resulting in a learning effect and an improvement in the algorithm's problem-solving ability over time. This can occur with or without human intervention (e.g., refinement).

[0025] In the context of this document, the term "deep learning" (multi-layered / deep learning) can refer specifically to a particular form of machine learning (as a special form of AI), especially neural networks. Deep learning algorithms, in particular, employ artificial neural networks with numerous hidden layers between the input and output layers, allowing for the development of a complex internal structure. Deep learning can be particularly well-suited, for example, for an algorithm to improve itself through training (self-adaptive).

[0026] In the context of this document, the term "Convolutional Neural Network (CNN)" may refer specifically to a particular deep learning approach. The structure of a Convolutional Neural Network might, for example, include one or more convolutional layers, followed by one or more pooling layers. After several iterative layers, a fully connected layer may be included. Due to the reduction in data volume in the convolutional and pooling layers, CNNs are particularly well-suited for processing large datasets, such as high-resolution images.

[0027] In the context of this document, the term "single-stage algorithm" refers specifically to an algorithm based on AI (particularly deep learning, and further specifically CNNs) that is designed to identify and classify signals in a single step (i.e., in one stage). As an example, such a single-stage algorithm can be implemented using an established algorithm like YOLO or SSD (examples of single-stage algorithms). Conventional approaches to evaluating two-dimensional data such as images (like CNNs) use a two-stage approach: first, the objects are identified, and only in a subsequent step are these objects classified. In contrast, the single-stage algorithm performs identification and classification in one step, thereby saving processing time (and thus computing resources).For example, the single-stage algorithm can divide an image into sections using a grid and analyze it section by section. Each section can then be viewed ("shot") once and analyzed in a single step.

[0028] In the context of this document, the term "one-dimensional time series" refers specifically to a representation or analysis of data that has only one dimension: time. This representation allows the development / change of a specific measurement over time to be observed (without including spatial or other variables). In the case of an axle counter, for example, an electrical parameter such as voltage can be viewed over time to analyze the presence of axles (especially wheels, bogies, etc.) over time.

[0029] In the context of this document, the term "determining signals" can refer specifically to the recognition or detection of signals within the dataset, particularly based on (unique) "signal features." For example, detecting a peak (e.g., on a wheel axle) can be understood as determining the signal feature "local maximum." In another embodiment, a signal feature can be located using a signal point (e.g., the peak tip). In yet another embodiment, a "signal pattern" can be recognized based on one or more signal features; for example, detecting two or three signal features (in a signal cluster, e.g., of a bogie) is considered determining the signal features "two / three consecutive local maxima."In other words, determination can refer to detecting whether and where a unique / unmistakable feature is located in a signal (progression) (localizing and extracting; where is it?) (see e.g. also . Figures 2, 3 , and 4 ).

[0030] In one embodiment, the "determination" can be performed in a more precise variant, which can be described as "segmenting." This allows for the precise localization of multiple signal features, e.g., several peaks in a bogie signal cluster (extracting finer structures).

[0031] In the context of this document, the term "classifying (assigning) signals" can refer specifically to the assignment of one or more signals (or signal characteristics and / or signal patterns) to a specific signal class (signal type). To classify signals, they are detected / identified based on their signal characteristics (conventionally, the steps of "identifying" and "classifying" are performed in two passes / stages). For example, assigning a single signal characteristic "peak" to the signal class "wheel" can be considered classification. Furthermore, assigning a signal pattern "high peak-low peak-high peak" to the signal class "bogie with magnetic track brake" can also be considered classification.

[0032] According to an exemplary embodiment, the invention can be based on the idea that a data set from an axle counter (the one-dimensional time history) can be evaluated quickly and reliably if the determination and classification of the signals is carried out in one step using a single-stage AI algorithm.

[0033] In contrast, conventional approaches to evaluating axle counter data sets (including using AI) are based on first detecting and then (in a second pass) classifying the signals.

[0034] Algorithms for analyzing images (i.e., two-dimensional datasets) are well-known, such as convolutional neural networks. These networks are conventionally used for demanding 2D and 3D representations, as they can be particularly advantageous for processing large amounts of data, such as complex, high-resolution images. These algorithms typically use the concept of "bounding boxes" to recognize specific objects within the images. Furthermore, algorithms are known (see YOLO, SSD) that can perform such 2D / 3D image analysis in a single step.

[0035] The inventors have now realized that such single-stage algorithms for analyzing two-dimensional images can be transferred to one-dimensional datasets, such as the time history of an axle counter, in a surprisingly fast, effective, and reliable manner. Such an approach is previously unknown because algorithms like YOLO or SSD are optimized for two-dimensional images, and therefore an advantageous implementation for one-dimensional (measurement) data has not been realized.

[0036] According to the invention, the simultaneous classification and detection of the one-dimensional signal (wheels, brakes, etc.) can be used by means of the single-stage ("single shot") technique.

[0037] The described method can, in an example, detect and classify (and segment) various types of train features, such as magnetic track brakes, wheels, bogies and entire wagons, in contrast to conventional evaluation approaches that, for example, can only detect wheels and cannot distinguish between other features (such as magnetic track brakes).

[0038] The described method can, in one example, significantly increase robustness against signal noise, as larger intervals of the signal are used for detection and classification. Furthermore, the method can calculate a "quality measure" of the underlying signal, which can be used in higher-level control systems to calculate reliability values.

[0039] The described method can, in one example, achieve significantly improved accuracy compared to current state-of-the-art methods. This allows for particularly reliable evaluation in a short time, enabling not only the determination but also the simultaneous classification of the acquired signals. The described method can be implemented computationally quickly and cost-effectively and can be directly integrated into existing systems. Exemplary implementation examples

[0040] According to one embodiment, the single-stage algorithm incorporates a Convolutional Neural Network (CNN). This can have the advantage of enabling implementation using an efficient and established algorithm. However, CNNs are currently known primarily from the analysis of images (two-dimensional data); their advantageous application to one-dimensional measurement data, particularly with regard to the specific application of axle counters, has not yet been explored.

[0041] According to one embodiment, the single-stage algorithm comprises a YOLO algorithm (or a derivative thereof) or an SSD algorithm (or a derivative thereof). This can have the advantage that a fast and reliable implementation using proven and reliable methods is possible.

[0042] YOLO (You Only Look Once) is a popular variant frequently used in computer vision applications. YOLO is an object detection algorithm that processes an entire image in a single pass and directly predicts bounding-box coordinates and class probabilities. The algorithm can achieve real-time performance by dividing the image into a grid and performing object detection for each grid cell.

[0043] SSD (Single-Shot Detector) uses multiple feature maps with different resolutions to efficiently detect objects of varying sizes in real time. SSD is known for its accuracy, especially in detecting small objects, making the algorithm useful for a wide range of computer vision applications.

[0044] In one embodiment, the single-stage algorithm is executed section by section along the one-dimensional time course. In one embodiment, the sections have a specific section width (in other words, at least two sections with the same dimensions are used). In another embodiment, the section width is partially variable (in other words, at least two sections with different dimensions are used).

[0045] Single-stage algorithms like YOLO or SSD divide a two-dimensional image into a grid and analyze the grid cells. To transfer this approach to one-dimensional data, sections can be used according to the invention. These can be linked together (with or without a gap between the sections) so that the algorithm can analyze the data set section by section. In a preferred example, only one pass (single-stage) of the algorithm through the sections is required. An illustrative example of the sections is, for example, in the Figures 5 and 6 shown. The section width can be (partially) varied, e.g. wider for a slow train and narrower for a faster train.

[0046] According to one embodiment, the determination involves: recognizing signal features in the one-dimensional time course. In another embodiment, the determination involves: recognizing signal patterns (based on signal features) in the one-dimensional time course. Object recognition algorithms usually use the concept of bounding boxes to locate objects (e.g., a car) in an image. However, in this case, one-dimensional data is used, so this concept is not applicable.

[0047] However, the inventors recognized that signal features, especially in the form of individual points, and / or signal patterns can be used efficiently and reliably, analogous to bounding boxes. For example, a local maximum (peak) can be recognized as a signal feature and determined by locating this point (see, e.g., [reference]). Figures 2 and 3 ).

[0048] According to one embodiment, determining signal patterns involves: recognizing a (bogie-associated) signal pattern based on five signal points.

[0049] This variant can be particularly efficient for detecting bogies. For example, the five signal points can be set as follows: i) Start of the bogie, ii) Position of first wheel (e.g. peak), iii) Position of magnetic track brake (present or not), iv) Position of second wheel (e.g. peak), and v) End of the bogie.

[0050] If these five signal characteristics can be set, the signal pattern "Bogie" can be determined. This can simultaneously result in the classification into the signal class "Bogie".

[0051] According to one embodiment, determining signal characteristics involves: recognizing a (wheel-associated) signal pattern based on at least one signal point (and its height / position).

[0052] In this variant, each signal characteristic can correspond to a signal pattern, e.g., "wheel" or "brake" (here, the height of the signal point can be crucial). Recognizing the signal measure "wheel" also results, in one example, in the classification into the signal class "wheel".

[0053] According to one embodiment, the method comprises: determining whether a section contains (relevant) information, and i) utilizing information-containing sections and / or ii) discarding information-free sections. In this context, the term "information" can, in particular, include at least one signal feature. Efficiency can be increased if only information-containing sections are analyzed (considered).

[0054] According to one embodiment, the single-stage algorithm is executed in anchor regions along the one-dimensional time history in addition to the segments. In other words, not only are segments defined and considered, but anchor regions at specific positions can also be defined and considered.

[0055] In one example, the sections are used homogeneously according to a fixed pattern (analogous to a grid), while the anchor regions are used inhomogeneously at specific positions. The anchor regions allow for more specific analysis of areas of interest (high density of signal features). See also the illustration in Figure 7 .

[0056] Single-stage algorithms (e.g., YOLO) use "grid cells" to recognize characteristic patterns (e.g., a car). The grid cells can be derived by simply dividing the image into several smaller fields. Sometimes, this single-stage object recognition algorithm additionally incorporates the concept of anchor boxes. In this case, different sizes of the image (e.g., for training and inference) are extracted at grid positions. In one embodiment, the time sequence is divided into fields with a predetermined fixed length / width (sections). As with many single-stage approaches, the concept of using anchor regions can also be applied according to the invention.

[0057] According to one embodiment, the anchor areas have at least two different dimensions. In one example, the anchor areas can be provided in the form of frames / boxes. The sizes of these can differ, for example.

[0058] According to one embodiment, at least one anchor area is smaller / larger than a corresponding section in at least one dimension. This can provide a particular degree of flexibility.

[0059] With a fine grid (compared to the signals to be detected) (or narrow section width) and large anchor areas, redundant detections can occur. In such cases, it may be necessary to suppress detections with lower reliability (non-maxima suppression).

[0060] According to one embodiment, the method includes: providing an output that includes the signal class (classification or classification probability; what is it?) and at least one signal feature (identify / detect) (or the position of at least one signal feature; where is it?), particularly in vector form. In other words, the output of the single-stage algorithm includes both the identification and the classification. Conventionally, this is performed in two stages. However, in the single-stage algorithm, the identification and classification information can be concatenated, e.g., as linked layers (see also Figure 1 ).

[0061] According to one embodiment, classification involves assigning signals to a class (particularly based on at least one signal feature and / or a signal pattern). These classes can be defined according to the specific application, and the underlying AI algorithm can be trained accordingly. A signal class can have one or more distinguishing features that are indicative of that class. A signal class can refer to individual signals and / or signal clusters (two or more signals that are associated with each other).

[0062] According to one embodiment, the signal classes have at least one of the following: one wheel, one magnetic track brake, two wheels, two wheels with one magnetic track brake, two wheels with no magnetic track brake, single-axle bogie, two-axle bogie. This is only an exemplary selection of signal classes that may be particularly common or (safety-)relevant in the field of rail transport.

[0063] According to a further embodiment, the method further comprises: training the single-stage algorithm with a plurality of one-dimensional time profiles of an axle counter (determined experimentally and / or theoretically). According to a further embodiment, the training further comprises: learning signal patterns and / or signal characteristics that are indicative for certain signal classes.

[0064] The training data can include measured (real) and / or simulated datasets. Depending on the application, certain training approaches may be particularly suitable.

[0065] According to a further embodiment, the method also includes: determining at least one of the following based on the evaluation: a speed, a length, a vehicle type, an anomaly, in particular a defect, a required maintenance.

[0066] This allows the described method to open up new possibilities for the entire control system of rail vehicles; some examples of implementation are given below: i) The detection and classification of bogies could be used to calculate the train speed (or the train type and dimensions are known, so the speed can be calculated); ii) The length and type of train (e.g., freight train, passenger train) could be estimated. The number of wagons could be counted; iii) Anomalies (e.g., defects in the wheels and / or magnetic track brakes) could be detected and reported; iv) Predictive maintenance could also be based on the signal recorded by the axle counter.

[0067] According to a further embodiment, the method also includes: preprocessing the data set to remove noise, in particular by smoothing and / or downsampling. This can have the advantage that even low-resolution signals can be used.

[0068] According to another embodiment, the method is used in connection / context with rail-bound vehicles and / or rail infrastructure.

[0069] The application of axle counters in rail transport has already been discussed in detail above. Furthermore, the rail vehicle sector has particularly high demands regarding safety and reliability, as well as corresponding standards. Therefore, the use of this method for validating evaluations can be especially important and advantageous in this context.

[0070] According to another embodiment, the single-stage algorithm / classifier is very fast (faster than other deep-learning-based methods, e.g., Unet). In one example, the algorithm is based on single-shot classifiers (such as the YOLO model used in computer vision for object recognition).

[0071] In another embodiment, a true end-to-end neural network is described that makes predictions of bounding boxes (signal features) and class probabilities simultaneously. In this example, it is not necessary to post-process the data (i.e., calculate classes or coordinates based on the neural network's prediction), which can make the entire algorithmic system more robust.

[0072] It should be noted that embodiments of the invention have been described with reference to different subject matter. In particular, some embodiments have been described with reference to method claims, while other embodiments have been described with reference to apparatus claims. However, a person skilled in the art will understand from the foregoing and the following description that, unless otherwise stated, in addition to any combination of features belonging to one type of subject matter, any combination of features relating to different subject matter is also deemed to be disclosed by this document. This applies in particular to features of the method claims and features of the apparatus claims.

[0073] The aspects defined above and further aspects of the present invention will become apparent from the examples of embodiments described below and will be explained with reference to these examples. The invention will be described in more detail below with reference to embodiments to which, however, the invention is not limited. Brief description of the drawings

[0074] Figure 1 Figure 1 schematically shows a method according to an exemplary embodiment of the invention. Figure 2 shows a determination of signal patterns based on five signal points, according to an exemplary embodiment of the invention. Figure 3 shows a determination of signal patterns based on five signal points, according to a further exemplary embodiment of the invention. Figure 4shows a determination of signal patterns based on individual signal points, according to an exemplary embodiment of the invention. Figure 5 Figure 1 shows a step-by-step execution of the single-stage algorithm along the one-dimensional time course with five signal points, according to an exemplary embodiment of the invention. Figure 6 shows a step-by-step execution of the single-stage algorithm along the one-dimensional time course with a single signal point, according to an exemplary embodiment of the invention. Figure 7 shows an execution of the single-stage algorithm along the one-dimensional time course in sections and anchor regions, according to an exemplary embodiment of the invention.

[0075] The Figures 8 and 9 Each shows a data record from an axle counter.

[0076] The Figures 10 to 12 illustrate a conventional threshold-based method and its disadvantages.

[0077] Figure 13 shows an axle counter system according to an exemplary embodiment of the invention. Detailed description of the drawings

[0078] The representations in the drawings are schematic. It should be noted that in different illustrations, similar or identical elements or features are designated with the same reference numerals or with reference numerals that differ from the corresponding reference numerals only in the first digit. To avoid unnecessary repetition, elements or features that have already been explained in relation to a previously described embodiment will not be explained again later in this description.

[0079] Furthermore, spatially relative terms such as "front" and "back," "top" and "bottom," "left" and "right," etc., are used to describe the relationship of one element to another, as illustrated in the figures. Thus, these spatially relative terms may apply to orientations used that differ from the orientation shown in the figures. Obviously, these spatially relative terms merely serve to simplify the description and the orientation shown in the figures and are not necessarily restrictive, since a device according to an embodiment of the invention may assume orientations other than those shown in the figures, particularly when in use.

[0080] Figure 1Figure 1 schematically shows a method according to an exemplary embodiment of the invention. First, a data set 110 (input) is provided, which is read by an axle counter (see e.g. Figure 13 ) was recorded / measured. The dataset exhibits a one-dimensional time course (see e.g. Figures 8 and 9 ). The evaluation of this data set 110 is carried out using a single-stage AI algorithm (single stage / shot algorithm / detector) 150.

[0081] The Kl algorithm 150 can, for example, be structured as a CNN with up to twenty layers 151. At the end of the CNN, there are, for example, two fully connected layers 152 that provide the following output: the class probability and the position of the signal features.

[0082] Signal features can be used in this context to locate a specific signal position. While a bounding box is used to define the position and size of an object in a two-dimensional domain, according to the invention, this can be done point by point in a one-dimensional domain, or signal features can be set in the form of signal points. One or more signal points can be indicative of a specific signal pattern. The signal pattern can be used analogously to the bounding box. In one embodiment, determining signals can mean "where is it?", while classifying signals can mean "what is it?".

[0083] In the example shown, a local maximum was identified as a peak, and a signal feature 112 was set using a signal point (signal identification). Based on this signal feature, the signal pattern 111 "Rad" (here as a bounding-box analogue) was identified, leading to its assignment to the signal class "Rad" (signal classification). The evaluation of the data set is thus achieved by identifying signals in the one-dimensional time course and classifying these signals in a single step using a single-stage deep learning algorithm.

[0084] Figure 2Figure 1 shows a method for determining signal characteristics based on five signal points, according to an exemplary embodiment of the invention. Using these five signal characteristics, a specific signal pattern can be recognized, in this case, a bogie-associated signal pattern. For ease of representation, the signal cluster shown is contained within a single section 160.

[0085] The following signal points 112 can be set: i) Start of the bogie, ii) Position of the first wheel (signal feature high peak 130), iii) Position of the magnetic track brake 135 (signal feature low peak 135), iv) Position of the second wheel (signal feature high peak), and v) End of the bogie.

[0086] All five signal points can be set and the signal pattern "Bogie with magnetic track brake" 131 can be recognized, which also results in the associated classification in the signal class "Bogie with magnetic track brake".

[0087] Figure 3 Figure 1 shows a determination of signal characteristics based on five signal points, according to a further exemplary embodiment of the invention. This example is that of the Figure 2 Similarly. The difference is that there is no peak for the magnetic track brake at position iii). Therefore, the five signal points result in the signal pattern "Bogie without magnetic track brake" 132, which also results in the associated classification (signal class "Bogie without magnetic track brake").

[0088] Figure 4 Figure 1 shows a method for determining signal characteristics based on a signal point and its height / position, according to an exemplary embodiment of the invention. A specific signal pattern 111 can be recognized by means of each signal characteristic 112. For clarity, each signal pattern 111 is shown in its own section 160-162.

[0089] In the first section 160, a local maximum is detected, and the signal characteristic "high peak" 130 can be set. This results in the signal pattern "wheel". In the second section 161, a local maximum is detected, and the signal characteristic "low peak" 135 can be set. This results in the signal pattern "magnetic track brake". In the third section 162, a local maximum is detected, and the signal characteristic "high peak" 130 can be set. This results in the signal pattern "wheel". The signal pattern sequence "wheel-magnetic track brake-wheel" leads to its assignment to the signal class "bogie with magnetic track brake".

[0090] While in the examples of the Figures 2 and 3 five signal characteristics 112 are assigned a common signal pattern 111, is at Figure 4 Each signal characteristic 112 is assigned a single signal pattern 111. In both cases, however, only one signal class is present.

[0091] Figure 5Figure 1 shows the execution of the one-stage algorithm along the one-dimensional time course through three sections 160-162 with the application of the five signal-point detection, according to an exemplary embodiment of the invention. The five signal points can be used as for the Figures 2 and 3 described and set.

[0092] 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") was possible, even though only a portion of the signal curve is present in the second section 161. However, in the first section 160 and the second section 162, a correct classification is not possible: the choice would be ambiguous based on the given data.

[0093] Figure 6Figure 1 shows the execution of the single-stage algorithm along the one-dimensional time course through three sections 160-162 (these would be the "shots") using single-signal-point detection, according to an exemplary embodiment of the invention.

[0094] In the first three sections, 160-162, a classification (using the "regression windows" calculated by the neural network) of the signal feature or signal pattern 111 (wheel or brake) is possible. However, the fourth section, 163, contains no feature (signal characteristic) and is therefore considered an information-free section and, in particular, discarded.

[0095] Figure 7Figure 1 shows the execution of the single-stage algorithm along the one-dimensional time course in sections 160 and anchor regions 170, according to an exemplary embodiment of the invention. In contrast to the other examples, a very fine sampling grid (small section width) is used. Larger anchor regions / boxes 170 are also used. This can lead to multiple detections, which can be processed with a non-maxima suppression algorithm to use only those detections with the highest confidence interval. Anchor regions of different dimensions can be used (see Figures 170-172) to cover different signal levels. In this example, poorly positioned anchor regions 175 cannot detect a signal feature.

[0096] Figure 13Figure 1 shows an axle counter system 100 according to an exemplary embodiment of the invention. This system comprises an axle counter sensor device 102 (e.g., one, two, or more sensors, e.g., a dual-sensor system) which is attached to a rail 101 and can record / measure the data sets (one-dimensional time series) described above. In this example, the sensor device 102 is connected via cable (alternatively wirelessly) to an axle counter evaluation unit 103. This evaluation unit 103 is configured to receive the measured data set and to evaluate or validate it according to the method described above. The evaluation can alternatively also be performed remotely.

[0097] It should be noted that the term "comprising" does not exclude other elements or steps, and the use of the article "a" does not exclude a plurality. Elements described in connection with different embodiments may also be combined. It should also be noted that reference numerals in the claims should not be interpreted as limiting the scope of the claims.

[0098] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included. Reference sign

[0099] 100 Axle counter system 101 Rail 102 Axle counter sensor 103 Axle counter evaluation unit 110 Data record, time series 111 Signal pattern, signal class 112 Signal feature, signal point 130 Signal feature high peak 131 Signal pattern bogie with magnetic track brake 132 Signal pattern bogie without magnetic track brake 135 Signal feature low peak 150 Single-stage algorithm 151 Layers 152 Linked layers 160-163 Section 170-172 Anchor area 175 Further anchor area 200Data set 201Signal Cluster bogie-without-magnetic track brake 202Signal Cluster bogie-with-magnetic track brake 203Signal wheel axle 204Signal magnetic track brake

Claims

1. A method for evaluating a data set (110) comprising: recording a one-dimensional time series as a data set (110) using an axis counter (100); and evaluating the data set (110) by identifying signals in the one-dimensional time series and classifying the signals in the one-dimensional time series in one step using a one-stage deep learning-based algorithm (150).

2. The method according to claim 1, wherein the single-stage algorithm (150) comprises a Convolutional Neural Network, CNN.

3. The method according to claim 1 or 2, wherein the single-stage algorithm (150) comprises a YOLO algorithm or a derivative thereof; and / or wherein the single-stage algorithm (150) comprises an SSD algorithm or a derivative thereof.

4. The method according to one of the preceding claims, wherein the single-stage algorithm (150) is carried out section by section (160) along the one-dimensional time course (110), in particular in sections (160) with a certain section width, further in particular wherein the section width (160) is partially variable.

5. The method according to one of the preceding claims, wherein the determining comprises: determining signal features (112) in the one-dimensional time course, in particular recognizing a signal pattern (111) based on at least one signal feature (112).

6. The method according to claim 5, further comprising: detecting a bogie-associated signal pattern (131, 132) based on five signal points; and / or detecting a wheel-associated signal pattern based on at least one signal point.

7. The method according to one of the preceding claims, wherein the single-stage algorithm (150) is carried out in at least one anchor region (170) along the one-dimensional time history (110) in addition to the sections (160), in particular wherein the anchor regions (170) have at least two different dimensions; and / or wherein at least one anchor region (170) is larger than a corresponding section (160) in at least one dimension.

8. The method according to one of the preceding claims, wherein the classification comprises assigning a signal to a class, in particular based on a detected signal pattern (111).

9. The method according to claim 8, wherein the signal classes comprise at least one of the following: one wheel, one magnetic track brake, two wheels (201), two wheels-one magnetic track brake (202), two wheels-no magnetic machine brake, single-axle bogie, two-axle bogie.

10. The method according to one of the preceding claims, comprising: providing an output (120) which has at least one signal feature (112) and an associated signal class (111), in particular in vector form.

11. The method according to one of the preceding claims, further comprising: training the single-stage algorithm (150) with a plurality of one-dimensional time histories of an axle counter, in particular wherein the training further comprises: learning signal features (112) and / or signal patterns (111) which are indicative for certain signal classes.

12. The method according to one of the preceding claims, further comprising: determining at least one of the following based on the evaluation: a speed, a length, a vehicle type, an anomaly, in particular a defect, a required maintenance.

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

14. A data processing device comprising at least one processor, and configured for: receiving a one-dimensional time series recorded by an axle counter (100) as a data set (110); and evaluating the data set (110) by identifying signals in the one-dimensional time series and classifying the signals in the one-dimensional time series in one step using a one-stage deep learning-based algorithm (150).

15. An axle counter system (100) comprising: at least one axle counter (102) for recording a one-dimensional time sequence (110); and a data processing device (103) according to claim 14, which is coupled to the at least one axle counter (102).

Images