System and method for detecting anomalies in railway radar measurements

Abstract

A train control system for detecting anomalies in radar sensor measurements includes a radar sensor system (200) with microwave transceivers (210) configured to transmit microwave beams and receive reflected signals, wherein the radar sensor system (200) is configured to provide speed measurements of an object based on the reflected signals, and a module (450) operably coupled to the radar sensor system (200), wherein the module (450) is configured, via computer executable instructions and a processor (460), to execute a prediction model (452) to determine anomalies in the speed measurements in real time.

Description

[0001]202312687 1 SYSTEM AND METHOD FOR DETECTING ANOMALIES IN RAILWAY RADAR MEASUREMENTS Technical Field Aspects of the present disclosure generally relate to radar sensor systems and more particularly to systems and methods for detecting anomalies in radar sensor measurements, for example in connection with railway applications. Background Art Radar sensor systems include a transmitter producing electromagnetic waves in the radio or microwaves domain, a transmitting / receiving antenna, a receiver, and a processor to determineproperties of objects, such as motion and velocity. By detecting an object’s Doppler shift, orchange in wave frequency, a radar sensor can compute the object’s speed along with its direction. Radar technology is found in various systems but is most prominently used in CW (continuous wave) Doppler, P (pulse radar) and FMCW (frequency modulated CW) radar sensors. Doppler radar sensors are specialized radars that use the Doppler effect to measure a target’s speed at different distances. FMCW radar sensors and pulse sensors are primarily used to measure a target’s distance. These methods can also use the Doppler effect to obtain additional target velocity information. However, determining the Doppler shift is complex or only possible at certain times. The CW method, on the other hand, is optimized for the evaluation of the Doppler shift and allows a continuous and robust measurement of the speed. An application of the radar sensor is in the field of trains and railway technology, herein referred to as railway radar sensor (RR sensor) or railway radar sensor system (RR sensor system). In anexample, the RR sensor emits two independent microwave beams directed toward a track andreceives reflected signals. Various sensor algorithms process raw signals and calculate their Doppler shift to derive vehicle speed, e. g. train speed. The vehicle speed is then transmitted to a train control system for further processing and downstream tasks. 202312687 2 Summary Briefly described, one or more embodiments of the present disclosure provide a system and methods for detecting anomalies in radar measurements, performed by a radar sensor system, for example a RR sensor system. The present disclosure provides embodiments to use machine learning methods to detect incorrect behavior of the radar sensor system, based on sensor measurements. A first aspect of the present disclosure provides a train control system, which is configured as or incorporated in an on-board system of a train, for detecting anomalies in radar sensor measurements, the system comprising a radar sensor system comprising microwave transceivers configured to transmit microwave beams and receive reflected signals, wherein the radar sensor system is configured to provide speed measurements of an object based on the reflected signals, and a module operably coupled to the radar sensor system, wherein the module is configured, via computer executable instructions and at least one processor, to execute a prediction model to determine anomalies in the speed measurements. A second aspect of the present disclosure provides a method for determining anomalies in radar sensor measurements, the method comprising receiving reflected signals including raw signal data, wherein the reflected signals are in response to emitted microwave beams, applying multiple algorithms to the raw signal data to provide speed measurements of an object along with a quality level of the speed measurements, and applying a prediction model configured to determine anomalies in the speed measurements in real time. A third aspect of the present disclosure provides a method for training a supervised machine learning algorithm to determine anomalies in speed measurements, the method comprising receiving training data including speed measurement data and speed reference data, the speed measurement data comprising measured speeds along with a quality level provided by a first device, and the speed reference data comprising reference speeds provided by a second device, and labeling the training data comprising assigning a tolerance interval to the quality level of each measured speed, wherein each quality level has a specific tolerance interval, associating a reference speed with each measured speed, wherein the measured speed is labeled as positive 202312687 3 when the associated reference speed falls within the tolerance interval of the quality level of the measured speed. Brief Description of the Drawings FIG.1 illustrates an example train control system with radar sensor system in accordance with an exemplary embodiment of the present disclosure. FIG.2 illustrates a block diagram of an example railway radar sensor system in accordance with an exemplary embodiment of the present disclosure. FIG. 3 illustrates a block diagram of a processing unit, including signal processing flow, of a railway radar sensor system in accordance with an exemplary embodiment of the present disclosure. FIG.4 illustrates a block diagram of a processing unit, including signal processing flow, with a prediction model based on a machine learning algorithm for detecting anomalies in measurements of a railway radar sensor system, in accordance with an exemplary embodiment of the present disclosure. FIG.5 illustrates a flow chart of a method for training the prediction model with supervised learning to determine anomalies in speed measurements in accordance with an exemplary embodiment of the present disclosure. Detailed Description To facilitate an understanding of embodiments, principles, and features of the present disclosure, they are explained hereinafter with reference to implementation in illustrative embodiments. In particular, they are described in the context of systems and methods for detecting anomalies in measurements of radar sensor system, for example railway radar sensor systems. The components and materials described hereinafter as making up the various embodiments are intended to be illustrative and not restrictive. Many suitable components and materials that would 202312687 4 perform the same or a similar function as the materials described herein are intended to be embraced within the scope of embodiments of the present disclosure. FIG.1 illustrates an example train control system 100 with radar sensor system in accordance with an exemplary embodiment of the present disclosure. Train control system 100 comprises a train control computer 110, configured as an on-board unit (OBU) in a train, specifically a locomotive. The train control computer 110 is operably coupled to a display 120 and a diagnostic interface 130, via a bus system 140, for example a multifunction vehicle bus (MVB). The MVB is a serial communication bus for railway vehicles and is used to connect devices, sensors, and actuators etc. The system 100 further comprises a communication module 150, for example GSM (Global System for Mobile Communications) module 150 and antenna 160, for example a GSM-R antenna. GSM-R stands for Global System for Mobile Communications-Railway and is an international wireless communications standard for railway communication and applications. The train control computer 110 is operably coupled to a valve 170 for a main air pipe, wherein the train control computer 110 is configured to send a command to open the valve 170 to release pressure in the main air pipe to activate the train’s air brakes to stop the train in an emergency. The system 100 further comprises one or more RR sensor systems 200 which are positioned on the train, specifically mounted to the locomotive so that the radar sensor systems 200 can measure the train speed over ground by applying the Doppler effect. The RR sensor systems 200 are connected to the train control computer 110. The train control computer is further coupled to odometer 180, designed for example as an odometer pulse generator that measures a distance by counting pulses derived from wheel rotation. The system 100 may further comprise balise antenna 190. Balise antenna 190 in conjunction with balises may be used for intermittent track- to-train communications. The balise system uses a transmission technique that is based on inductive coupling and data transmission with frequency shift keying. FIG.2 illustrates a block diagram of an example railway radar sensor system 200 in accordance with an exemplary embodiment of the present disclosure. 202312687 5It should be noted that FIG.2 illustrates a simplified diagram of RR sensor system 200, includinga signal path. RR sensor system 200 comprises microwave transceivers 210, processing unit 220 and other components 230. The RR sensor system 200 includes at least two microwave transceivers 210. In other examples, there may be more than two transceivers 210. The processing unit 220 can be configured as a digital signal processor board (DSP board). The other components 230 may include for example memory, power supply, etc. The RR system 200 is configured as a Doppler radar system which processes microwave signals reflected in track bed 240 and forwards measurements, calculations and / or determinations, such as speed measurements, variances, and other data to odometry system 250. More specifically, the microwave transceivers 210 emit two independent microwave beams directed toward the track bed 240 and receives reflected signals. Various algorithm(s) in the processing unit 220 process the reflected raw signals and calculate their Doppler shift to derive the speed, e. g. train speed. The speed is then transmitted to the odometry system 250 and / or train control system 210 for further processing and downstream tasks. FIG.3 illustrates a block diagram of a processing unit 300, including signal processing flow, of the RR radar sensor system in accordance with an exemplary embodiment of the present disclosure. The processing unit 300 can be for example the processing unit 220 and the railwayradar sensor system can be RR sensor system 200 as described in connection with FIG.2.The processing unit 300 includes algorithm(s) 310 that receive reflected microwave signals 340 as raw data and provide determinations and measurements, for example train speed. The RR sensor system 200 uses different traditional deterministic algorithms to determine the determinations, measurements, or values, such as speed and distance, based on the microwave signals 340. The algorithms 310 work in parallel and report their results, such as speed measurements, to a quality voter algorithm 320. The quality voter algorithm 320 is configured to compare results of the algorithms 310 under several criteria, to choose a delivered speed information and to perform self-estimation on the delivered results / values. The voter algorithm 320 provides its results via interface 330, wherein the results include measurement values including quality level, variance estimates, and other information. 202312687 6 If a degraded signal is detected, the quality voter algorithm 320 decreases the reported quality, e. g. a high-quality level corresponds to an accurate measurement and a low-quality level corresponds to a less accurate measurement. However, there may be instances when the RR sensor system 200 reports inaccurate speed without reducing the reported signal quality level, herein also referred to as '’incorrect behavior''. For example, weather and track conditions can adversely affect the accuracy of the measurement values derived by the RR system 200. For example, the track bed 204 can become icy after snow, which significantly degrades the sensor signal performance. This problem cannot be corrected by sensor calibration. FIG.4 illustrates a block diagram of a processing unit 400, including signal processing flow, with a prediction model for detecting anomalies in measurements of a railway radar sensor system, in accordance with an exemplary embodiment of the present disclosure. As noted above, there may be instances when the RR sensor system reports inaccurate measurement values, such as speed without reducing the reported signal quality, which are herein referred to as anomalies or incorrect behavior of the RR sensor system.The processing unit 400 corresponds to the previously described processing unit 300 of FIG.3 incertain components. The processing unit 400 includes at least one processor 460 and at least one memory 470 and further includes algorithm(s) 410 that receive reflected microwave signals 440 as raw data and provide determinations and measurements, for example train speed. The RR sensor system uses different traditional deterministic algorithms to determine the determinations, measurements, or values, such as speed and distance, based on the microwave signals 440. The algorithms 410 work in parallel and report their results (for example speed) to a quality voter algorithm 420. The quality voter algorithm 420 is configured to compare results of the algorithms 410 under several criteria, choose a delivered speed information and perform self- estimation on the delivered results / values. The voter algorithm 420 provides its results via interface 430, wherein the results include measurement values including quality level, variance estimates, and other information. In accordance with an exemplary embodiment of the present disclosure, systems and methods for detecting anomalies in radar sensor measurements, in connection with railway applications, are provided. More specifically, machine learning methods are utilized to detect incorrect behavior 202312687 7 of the RR sensor system 200 based on the measurements by the RR sensor system 200. More specifically, a module 450 is provided and configured, via computer executable instructions and at least one processor 460, to execute a prediction model 452 to determine anomalies in the speed measurements. The prediction model 452 is built and based on machine learning (ML) algorithm(s). In an embodiment, a train control system with RR sensor system 200 and module 450 is configured as or incorporated into an on-board unit of a train. The module 450 may be embodied as software or a combination of software and hardware. The module 450 may be a separate module or may be an existing module programmed to perform a method as described herein. For example, the module 450 may be incorporated, for example programmed, into an existing processing unit component or RR sensor system component, by means of software. In another example, the module 450 may be a firmware plugin in the processing unit 400.In the example of FIG.4, the module 450 comprising the prediction model 452 is operably coupledand incorporated into the processing unit 400 of a RR sensor system. In another example, the module 450 with prediction model 452 may be external to the processing unit 400 and may be included in another component of the RR sensor system. In yet another example, the prediction model 452 may be external to the RR sensor system and can be included in a train controlcomputer 110, as for example illustrated in FIG.1, which can be an on-board unit of a locomotive.The module 450 is operably coupled, directly or indirectly, to the RR sensor system to receive and analyze the measurement values of the RR sensor system. In an embodiment, the module 450 is configured to classify the speed measurements into first speed measurements and second speed measurements, the first speed measurements including a correct behavior and second speed measurements comprising an incorrect behavior of the RR sensor system. The module 450 is configured to determine and / or output incorrect behavior of the RR sensor system, the incorrect behavior including incorrect speed measurements, more specifically, speed measurements with an incorrect quality level. For example, a speed measurement 50 km / h may have an indicator that the signal quality is high, which means that the speed measurement is correct. However, due to an icy track bed, this measurement is likely incorrect. The provided prediction model 452 is trained and configured to identify and detect such incorrect behavior of the RR sensor system. 202312687 8 In an embodiment, the prediction model 452 is trained with supervised learning and is based on a binary classification supervised learning problem. The learning and training are described laterwith reference to FIG. 5. Once trained, the prediction model 452 can perform real-time detectionor near-real-time detection of the incorrect behavior of the RR sensor system. A real-time detection refers to data processing within milliseconds so that an output / result it is available virtually immediately as feedback. A near real-time detection refers to data processing that is slightly slower than rea-time processing, for example within seconds or minutes. The prediction model (non-deterministic) 452 provides feedback to the deterministic model implemented by the voter algorithm 420, to improve performance. This is a key aspect of the present disclosure, because many safety-critical systems in the railway industry refuse to trust statistical models. If the information generated by the prediction model 452 is checked for suitability with "classical" algorithms, the effects of "false predictions" can be dealt-with and bounded. The prediction model 452 therefore cannot make its own decisions, but it provides additional information. Further, an odometry system 250 is configured to receive the speed measurements from the RR sensor system, for example from the processing unit 400, for further processing and downstream tasks. Specifically, the measurements that are correct are forwarded to the odometry system 250, after the module 450 with model 452 has evaluated the measurements. FIG.5 illustrates a flow chart of a method 500 for training the prediction model 452 with supervised learning to determine anomalies in speed measurements in accordance with an exemplary embodiment of the present disclosure. While the method is described as a series of acts or steps that are performed in a sequence, it is to be understood that the method may not be limited by the order of the sequence. For instance, unless stated otherwise, some acts may occur in a different order than what is described herein. In addition, in some cases, an act may occur concurrently with another act. Furthermore, in some instances, not all acts may be required to implement a methodology described herein. In accordance with an exemplary embodiment of the present disclosure, the method 500 includes receiving training data 510 including speed measurement data and speed reference data. The 202312687 9 speed measurement data and reference data can be actual RR sensor system data or can be simulated data. The speed measurement data comprise measured speeds along with a quality level provided by a first device, which is the RR sensor system (see for example RR sensorsystem 200 in FIG. 1), and the speed reference data comprise reference speeds provided by asecond device, which is for example a wheel tachometer. A wheel tachometer is mounted on a wheel, specifically a train / locomotive wheel, and measures or determines different characteristics including speed of the train. Act 520 comprising data pre-processing including processing anomalous reference data and resampling the reference data. The reference data is measured from a train wheel, which might suffer from slip and slide and result in anomalous reference data. Domain knowledge is utilized to eliminate part of the anomalous reference data. The rest of the reference data are treated as valid and can be used for labeling. As the SRR data and reference data are measured at different frequencies, it is essential to resample the reference data so that we have a reference speed for each measured speed provided by the RR sensor. This is done using piecewise linear interpolation of the reference data and sampling the RR sensor time points from the interpolated function. Act 530 includes data labeling comprising associating a reference speed with each measured speed, assigning a tolerance interval to the quality level of each measured speed, wherein each quality level has a specific tolerance interval, wherein the measured speed is labeled as positive when the associated reference speed falls within the tolerance interval of the quality level of the measured speed. In an embodiment, there are four quality levels in the speed measurement data: 0 - extremely low quality –> high tolerance interval, 1 - low quality, 2 - medium quality, and 3 - high quality –> small tolerance interval. With different quality levels of the reported speed measurement value / signal, different tolerance intervals are set or associated with the speed measurement values. A reported speed with a quality level 3 has a small tolerance interval. The lower the quality level is, the greater the 202312687 10 tolerance interval is. If the respective reference speed falls within the tolerance interval, the speed measurement is labeled as positive (correct behavior); otherwise, it is labeled as negative (incorrect behavior or an anomaly). In a simple example, a speed measurement is reported as 50 km / h with a quality level of 3. Quality level 3 has a small tolerance interval, for example + / - 1 km / h. The respective reference speed, provided by the wheel tachometer is 49 km / h, which falls in the tolerance interval. This means that the speed measurement is labeled as positive, associated with a correct behavior of the RR sensor system. If the reference speed is 48 km / h, which does not fall within the tolerance interval, the speed measurement is labeled as negative (incorrect / anomalous behavior of the RR sensor system). Following the data labeling, feature extraction or feature engineering (act 54) is performed, which comprises selecting, manipulating, and transforming the labeled data into features usable by the ML algorithm 550. Feature engineering refers to the process of selecting, manipulating, and transforming the raw data into features that can be used in the machine learning pipeline. The following are the primary considerations in processing the measurement data into features: 1) Remove columns that are intuitively not an indicator of whether the measurement value has incorrect behavior, such as timestamp, sequence number, distance, and total distance. 2) There are columns of type “8-bit”. Each column of this type is split into eight binary features. 3) Normalize the range of each feature into the interval [-1, 1]. Optional step in feature engineering or feature extraction: In the discussions above, an incorrect behavior in the RR sensor is predicted at a certain time point by only leveraging RR sensor information at that specific time point. However, this information might not be enough to make such a decision. RR sensor information in previous time points might be helpful as well. This can be addressed by adopting a ‘sliding window approach to consider past information’. For example, this information can be included by concatenating features of consecutive time points and treating the concatenation as one sample. The features are then fed into the machine learning algorithm 550 to build a prediction model 560, which eventually is deployed as a prediction model to the RR sensor system, see act 580. 202312687 11 The binary classification problem is solved using a machine learning algorithm 550, selected from the following: 1) Decision Tree algorithm: A decision tree is a non-parametric supervised learning method. It learns simple decision rules inferred from the data features to predict the value of a target variable. The decision tree reports a balanced accuracy of 88.3% on the RR sensor system dataset. 2) AdaBoost algorithm: AdaBoost stands for adaptive boosting and is a popular ensemble method that combines the predictions of several base estimators to improve generalizability over a single estimator. In experiments, the base estimators are small decision trees. AdaBoost reports a balanced accuracy of 58.8% on the RR sensor dataset. 3) Support Vector Machine (SVM) algorithm: SVM is a robust prediction method. The original SVM is a linear classifier, but SVM can efficiently perform non-linear classification using the kernel trick. The downside of kernel SVM is that the algorithm becomes impractical when the number of samples is large. Several techniques like clustering and online kernel SVM are developed to deal with large datasets. When applying linear SVM to the RR sensor dataset, a balanced accuracy of 70.3% is reported. 4) Neural Network (Multi-layer perceptron): A multi-layer perceptron (MLP) or a feed-forward network is the simplest neural network architecture. Our MLP reports a balanced accuracy of 93.2%. Note: Balanced accuracy is a metric that evaluates binary classification results by averaging the accuracy (recall) of the positive and negative classes. Balanced accuracy is a better metric than accuracy when the number of positive and negative samples is unbalanced. After feeding the features into the ML algorithm 550, which can be for example a decision tree algorithm, the prediction model 560 is built. The prediction model 560 is then evaluated and tuned, to make the prediction model 560 as accurate as possible. The prediction model 560 is thendeployed to the RR sensor system, for example to the processing unit 400, as illustrated in FIG.4. Thus, the prediction model 560 corresponds to the prediction model 452 of FIG.4.As described, unlabeled data sets have been turned into a labeled dataset that can be readily used by supervised learning algorithms. The reference speed is only used for labeling purposes. The reference speed is not a feature in the dataset used for training the machine learning 202312687 12 algorithm. The algorithm needs to learn to infer whether the speed measured by the RR sensor system along with a reported quality level is an ‘incorrect behavior’ only using information from the RR sensor system. Once the algorithm is trained, the algorithm / prediction model 560 is deployed on the RR sensor system for (near-)real-time detection of incorrect behavior of the SRR sensor system.

Claims

202312687 13 Claims 1. A train control system for detecting anomalies in radar sensor measurements, thesystem comprising: a radar sensor system (200) comprising microwave transceivers (210) configured to transmit microwave beams and receive reflected signals, wherein the radar sensor system (200) is configured to provide speed measurements of an object based on the reflected signals, and a module (450) operably coupled to the radar sensor system (200), wherein the module (450) is configured, via computer executable instructions and at least one processor (460), to execute a prediction model (452) to determine anomalies in the speed measurements in real time.

2. The train control system of claim 1,wherein the module (450) is configured to classify the speed measurements into first speed measurements and second speed measurements, the first speed measurements including correct speed measurements and second speed measurements including incorrect measurements.

3. The train control system of any of the preceding claims 1 or 2,wherein the prediction model (452) is based on machine learning and a binary classification supervised learning problem.

4. The train control system of any of the preceding claims 1, 2 or 3,wherein the radar sensor system (200) comprises a processing unit (400) storing the module (450) and the prediction model (452).

5. The train control system of any of the preceding claims 1 to 4,wherein the module (450) is operably coupled to the radar sensor system (200), or the module (450) is incorporated into the radar sensor system (200).

6. The train control system of any of the preceding claims 1 to 5, further comprising:an odometry system (250) configured to receive the speed measurements from the radar sensor system (200) for further processing, wherein speed measurements including a correct behavior are forwarded to the odometry system (250).202312687 14 7. The train control system of claim 1,wherein the train control system with radar sensor system (200) and module (450) is configured as or incorporated into an on-board unit of a train.

8. A method (500) for training a supervised machine learning algorithm to determineanomalies in speed measurements, the method comprising: receiving training data (510) including speed measurement data and speed reference data, the speed measurement data comprising measured speeds along with a quality level provided by a first device, and the speed reference data comprising reference speeds provided by a second device, and labeling (530) the training data comprising assigning a tolerance interval to the quality level of each measured speed, wherein each quality level has a specific tolerance interval, associating a reference speed with each measured speed, wherein the measured speed is labeled as positive when the associated reference speed falls within the tolerance interval of the quality level of the measured speed.

9. The method (500) of claim 8, further comprising:performing feature engineering (540) including selecting, manipulating, and transforming the labeled training data into features usable by the machine learning algorithm.

10. The method (500) of claim 9, further comprising: feeding the features into the machine learning algorithm (550) and building a prediction model (452), wherein the machine learning algorithm (550) is selected from a decision tree algorithm, adaptive boosting (AdaBoost) algorithm, a support vector machine (SVM) algorithm, and a neural network.202312687 15 11. A method for determining anomalies in radar sensor measurements, the method comprising: receiving reflected signals including raw signal data, wherein the reflected signals are in response to emitted microwave beams, applying multiple algorithms (410) to the raw signal data to provide speed measurements of an object along with a quality level of the speed measurements, and applying a prediction model (452) configured to determine anomalies in the speed measurements.

12. The method of claim 11, wherein the prediction model (452) is configured to classify the speed measurements into first speed measurements and second speed measurements, the first speed measurements including a correct measurement and the second speed measurements including an incorrect measurement.

13. The method of claim 12, further comprising: forwarding the first speed measurements including a correct behavior to an odometry system (250).

14. The method of claim 11, wherein the method is performed by a radar sensor system (220) mounted on a train, and wherein the emitted microwave beams are directed toward a train track.

15. A non-transitory computer readable medium storing executable instructions, which, when executed by a computer, perform a method for determining anomalies in speed measurements as claimed in claim 11.

Images