SYSTEM FOR CLASSIFYING RADAR MARKS INTO FALSE ITCH MARKS AND OTHER MARKS THAT REFER TO TARGETS OF INTEREST SUCH AS AIRCRAFT OR FLYING BODY

Description

[0001] The present invention relates to a system for classifying radar tracks into false tracks, corresponding to radar angels, and other tracks, corresponding to targets of interest such as aircraft or missiles.

[0002] The invention relates to radar angels, which are radar echoes that cannot be attributed to reflections from targets of interest, such as aircraft in the field of air surveillance, but can be attributed to atmospheric phenomena such as variations in the refractive index of air or the presence of airborne particles. When atmospheric conditions are right, radar angels form sufficiently coherent sets for radars to initiate and confirm tracks. By extension, these false leads, i.e., tracks that are not on targets of interest, are called false angels. These false angels are added to the tracks, are propagated throughout the surveillance chain, and disrupt the air surveillance mission.

[0003] One aim of the invention is to prevent these false leads from appearing in the situation presented to the operator or being identified as such.

[0004] Existing solutions are often based on minimum speed criteria for track confirmation because false angel tracks often have much lower speeds than aircraft. Increasing the minimum speed threshold has the drawback of also eliminating tracks on slow-moving drones, a particularly interesting type of target. Sometimes, increasing the minimum speed threshold is limited to a restricted area around the radar and within a specific altitude range. However, the area in which false angel tracks are observed extends over large distances and very wide altitude ranges. In practice, this is impractical because the area in which drone detection is affected is too large. A minimum radial speed threshold is sometimes defined per angular sector.In practice, this is ineffective and very difficult to adjust because false angel tracks rarely follow a radial trajectory relative to the radar. Finally, these workarounds are adjusted manually, but since the presence and characteristics of angel tracks depend on atmospheric conditions, they require frequent adjustments.

[0005] When atmospheric conditions are right, particularly when angel phenomena originate from wind-borne particles, radar angels form sufficiently coherent sets for radars to initiate and confirm tracks. These tracks generally share common characteristics. First, they move at relatively low speeds (less than 50 m / s) and are similar in heading and velocity. They are distributed over a disk centered on the radar, and when observed over a long period, their straight trajectories appear to have been traced by a toothless comb passing through the sand.

[0006] Document EP3167305B1 discloses a radar track classification system, into false tracks corresponding to abnormal atmospheric propagation phenomena and tracks of interest corresponding to aircraft.

[0007] There [ Fig. 1 ] schematically represents, in thicker lines, runways representative of aircraft, and in thinner lines, radar angels.

[0008] The phenomenon of radar angels is not permanent, as it depends on atmospheric conditions. The duration of the phenomenon ranges from a few tens of minutes to a few hours. When it is active, the stationary period of the phenomenon exceeds ten minutes.

[0009] These false leads, like angels, add to the other leads, are propagated throughout the surveillance chain and disrupt the surveillance mission.

[0010] It is known that signal processing and computer processing are used to handle angels. The techniques used are: Elimination by tracking only moving echoes; Echo filtering using their Doppler velocity; Correlation with secondary surveillance radars; Constant false alarm rate; Counting tracks exhibiting predefined characteristics of angel phenomena and suppressing these tracks when their number exceeds a threshold in a given sector.

[0011] These types of known solutions have the following limitations, respectively: Eliminating targets by tracking only moving echoes requires determining a minimum threshold beyond which movement is considered present. The massive emergence of slow-moving targets (such as slow-moving drones) makes it impossible to apply thresholds that eliminate angels without also eliminating drones. Defining this threshold is, at best, manual; at worst, a factory setting that cannot be adjusted by a radar operator, compromising timely adaptation to current conditions. Filtering echoes using their Doppler velocity also requires setting a minimum threshold that applies to the radial velocity relative to the radar—that is, the velocity of an object measured in the direction of the beam (or line of sight) to or from the radar. Applying this threshold also eliminates truly slow-moving targets. Furthermore, the distribution of angels around the radar with non-radial trajectories makes setting a relevant threshold difficult.The threshold definition is at best manual, at worst a factory setting that cannot be adjusted by a radar operator, which compromises timely adaptation to current conditions. Correlation with secondary surveillance radars is not possible to distinguish between angels and non-cooperative targets, which, by definition, do not respond to secondary radars. While the constant false alarm rate method is adaptive, it leads to more stringent detection confirmation and therefore affects the detection of targets with a small radar cross-section, including drones.Counting tracks with predefined "angel" characteristics and deleting these tracks when their number exceeds a threshold in a given sector presents two major drawbacks: the presumed characteristics of the angels are assumed to be known, which is poorly suited to evolving phenomena; counting by sector instead of non-sectorized counting is poorly suited to a non-sectorized phenomenon and leads either to delaying the application of the deletion process or to triggering it untimely.

[0012] A primary radar is a radar equipped with a primary antenna that emits a signal and receives the signal reflected by a target, while a secondary radar is a radar equipped with a secondary antenna that receives an interrogation signal emitted by a target in response to the reception of a signal emitted by that antenna. A primary / secondary radar is a radar equipped with both a primary and a secondary antenna.

[0013] A track is represented by a set of parameter values ​​that represent the fusion of successive radar detections of a target. When a track is maintained by primary detections, it has a primary source. When a track is maintained by secondary detections, it has a secondary source. A track can have both a primary and a secondary source.

[0014] One aim of the invention is to enable these false leads (angels) to be automatically identified so as to either eliminate them or allow them to be filtered out of the airspace situation presented to a surveillance operator. This filtering must not be at the expense of genuine leads on slow-moving targets with a small radar cross-section, and must adapt automatically to the current situation.

[0015] According to one aspect of the invention, a system for classifying radar tracks of aircraft-type targets of interest is proposed, into false tracks (angels), corresponding to radar angels, and other tracks, corresponding to targets of interest: a first source classification stage, receiving a track update as input, configured to determine if the updated track has a secondary source and output a classification as another track if the updated track has a secondary source, or pass the track update to a second stage; the second stage for detecting and characterizing an angel phenomenon, configured to detect an angel phenomenon and output a status of detection or not of an angel phenomenon, to a first status test module, and, in case of detection of an angel phenomenon, output a characterization of the detected angel phenomenon, to a third stage;the first status test module, receiving as input the track update from the first stage and the detection status or not of the angel phenomenon from the second stage, configured to test the value of the detection status or not of an angel phenomenon by the second stage, and, in case of non-detection status of an angel phenomenon, deliver as output a classification as another track of the updated track, and in case of detection status of angel track phenomenon, transmit the track update to the third stage;the second stage of detection and characterization of an angel phenomenon includes: a fourth test module for whether or not the end of a repetitive time interval has been reached, configured to, after storing the track update in a database of the track update history of the current time interval, if the end of the current time interval is reached, extract the contents of the database of the track update history of the current time interval, reset the database of the track update history of the current time interval, and transmit track updates of the current time interval to a partitioning module;The partitioning module is configured to perform partitioning of track updates within the elapsed time interval by studying the similarity of track update attributes, including heading and speed, detecting the presence of an angle phenomenon, outputting a status indicating whether or not an angle phenomenon has been detected, and, if an angle phenomenon is detected, outputting a characterization of the detected angle phenomenon, at the third stage; and the third stage is configured to perform classification of the updated track as a false angle track or other track, using two pre-trained classifiers arranged in series, each employing a random forest algorithm.

[0016] In one embodiment, the first source classification stage includes: a second test module configured to test whether an identifier of the updated track corresponds to a track identifier having a secondary radar source already listed as a track with a secondary radar source, by accessing a database of track histories having a secondary source, to output a classification as another track of the updated track if the identifier corresponds to a track having a secondary radar source in the database of track histories having a secondary radar source, and to transmit the track update to a third test module otherwise;The third test module is configured to test whether the track update originates from a secondary radar source, to store the updated track in memory in the track history database for tracks with a secondary radar source if the updated track is considered to originate from a secondary radar source and output a classification as another track for the updated track, and to transmit the track update to the second stage otherwise.

[0017] In one embodiment, the third tier of trail update classification, for a time interval, such as false lead, angel, or other trail, includes: a first pre-trained classifier using a random forest type algorithm to deliver a current track classification as output when the current time interval has not elapsed; a second pre-trained classifier using a random forest type algorithm; a classification likelihood calculation module for the second classifier; and a rule base for determining the final classification from the likelihood vector, configured to be applied as output from the likelihood calculation module and to deliver a classification of the track as a false track, angel track, or other track at the end of the current time interval.

[0018] According to one embodiment, the system includes a track update database configured to train the first and second classifiers of the third floor by learning.

[0019] Another aspect of the invention also proposes an air surveillance center, equipped with a false trail detection system, as previously described.

[0020] The invention will be better understood upon examination of some embodiments described by way of non-limiting examples and illustrated by the accompanying drawings, in which: [ Fig.1 ] schematically illustrates an angelic phenomenon, according to the state of the art; and [ Fig.2 ] schematically illustrates a system for detecting angular tracks comprising an angular phenomenon detection device and a radar track classification device for determining whether a track is a false angular track or another track, according to one aspect of the invention; Fig.3 ] schematically illustrates a first stage of a system of the [ Fig.2 ], according to one aspect of the invention; [ Fig.4 ] schematically illustrates a second stage of a system of the [ Fig.2 ], according to one aspect of the invention; [ Fig.5 ] schematically illustrates a third stage of a system of the [ Fig.2 ], according to one aspect of the invention; and [ Fig.6 ] schematically illustrates a chronology of the system's operation of the [ Fig.1 ], according to one aspect of the invention.

[0021] Across all figures, elements with identical references are similar.

[0022] There [ Fig.2 ] schematically illustrates an angle track detection system comprising an angle phenomenon detection device and a radar track classification device for false angle track or other track, according to one aspect of the invention.

[0023] A lead is a history of successive updates to the state, according to a set of characteristics, of the tracking of a presumed target. A lead is designated by a unique identifier, called a lead identifier.

[0024] The false lead detection system takes a lead update as input and provides as output a classification "false lead angel" or "other lead".

[0025] The radar track classification system, into false angel tracks, corresponding to radar angels, and other tracks, corresponding to targets of interest, comprises three levels: a first stage 1 of source classification, receiving as input a track update, configured to determine if a track has a secondary radar source and output a classification as another track if an updated track has a secondary radar source, or transmit the track update to a second stage 2; the second stage 2 of detection and characterization of an angel phenomenon, configured to detect an angel phenomenon and output a status of detection or not of an angel phenomenon, to a first status test module 3, and, in case of detection of an angel phenomenon, output a characterization of the detected angel phenomenon, to a third stage 4;the first status test module 3, receiving as input the track update from the first stage 1 and the detection status of an angel phenomenon from the second stage 2, configured to test the value of the detection status of an angel phenomenon by the second stage 2, and, in the case of a non-detection status of an angel phenomenon, output a classification as another track of the updated track, and in the case of a detection status of an angel phenomenon, transmit the track update to the third stage 4; and the third stage 4, configured to perform a classification of the track update as a false angel track or another track, by two pre-trained classifiers, arranged in series, each using a random forest type algorithm.

[0026] The characterization of the detected angel phenomenon includes parameter values ​​including the direction and speed of the detected angel phenomenon.

[0027] The present invention has the following advantages: automatic detection of the angel phenomenon; automatic extraction of the characteristics of the phenomenon; automatic activation of the pre-trained false lead angels vs other lead classifier; automatic consideration by the classification stage of the parameters of the phenomenon; periodic re-evaluation of the false lead classification "angels" vs other lead; significant filtering of false leads angels; and a low risk of deleting drone tracks.

[0028] Two phases can be distinguished, classic to a "Machine Learning" algorithm, in the operation of the system: a learning phase which is to be carried out prior to any exploitation but only needs to be executed once by providing a set of labeled track updates (i.e. whose actual classification is known) as input, the different classifiers of the system based on supervised Machine Learning self-parameterize; and an inference phase which takes track updates as input, and returns a classification indicating whether it is likely a false lead or another lead.

[0029] There [ Fig3 [This schematically illustrates the first stage 1 of source classification. The first stage 1 includes:] a second test module 5 configured to test whether an identifier of the updated track corresponds to a track having an already identified secondary radar source, by accessing 6 a database 7 of tracks having a secondary radar source, to deliver as output a classification as another track of the updated track if the identifier corresponds to a track identifier having a secondary radar source in the database 7 of track histories having a secondary radar source, and to transmit the track update to a third test module 8 otherwise;the third test module 8 configured to test whether the track update has a secondary radar source, to store 9 in memory the updated track in the database 6 of tracks having a secondary radar source if the track update is considered to have a secondary radar source and to output a characterization of the detected false track angle, and to transmit the track update to the second stage 2 otherwise. ;

[0030] There [ Fig4 [This diagram schematically illustrates the second stage 2 for the detection and characterization of an angel phenomenon. The second stage 2 comprises:] a fourth test module 10 of the attainment or not of the end of a repetitive time interval ΔT configured to, after storage 11 of the track update in a database 12 of the history of track updates of the current time interval ΔT, if the end of the current time interval ΔT is attained to extract 13 the contents of the database of the history of track updates of the current time interval, reset 13 the database 12 of the history of track updates of the current time interval ΔT, and transmit track updates of the current time interval ΔT to a partitioning module 14;the partitioning module 14 configured to perform a partitioning of the track updates of the elapsed time interval ΔT, by studying the similarity of attributes of the track update including the heading and speed of the track, detect 15 the presence of an angle phenomenon, deliver as output a status of detection or not of an angle phenomenon, and, in case of detection 15 of an angle phenomenon, deliver as output a characterization of the detected angle phenomenon, at the third stage 4. ;

[0031] There [ Fig5 ] schematically illustrates the third stage 4 of the track update classification, for a time interval Δt, as a false lead, angel, or other lead. Time is divided into successive time intervals Δt, as illustrated on the [ Fig.6 The third floor, number 4, comprises: a first pre-trained classifier 18 using a random forest type algorithm to deliver a current classification of the track as output when the current time interval Δt has not elapsed; a second pre-trained classifier 19 using a random forest type algorithm; a classification likelihood calculation module 20 of the second classifier 19; and a rule base 21 for determining the final classification from the likelihood vector, configured to be applied as output of the likelihood calculation module 20 and to deliver a classification of the track as a false track, angular track, or other track at the end of the current time interval ΔT.

[0032] From the track update and the characterization of the detected angle phenomenon, a module 22 for extracting and calculating features of interest performs a data extraction of the track update and the characterization of the detected angle phenomenon to provide the first classifier 18 with features of the track update, i.e. parameter values ​​representative of the track update including its kinematic parameters and its radar classification.

[0033] The first classifier outputs a classified track update which is stored in memory by a memory storage module 23 in a track update history database 24 of the segment (i.e., all updates of a track that took place during the current time interval ΔT)

[0034] A fifth test module 25 tests whether the current time interval Δt has elapsed, and if not, as illustrated in the [ Fig.6 ], without being at the end of the current time interval Δt, a memory extraction module 6 extracts from a database 27 of track classification histories, a current classification of the output track.

[0035] If the fifth test module 25 determines that the current time interval Δt has elapsed, a data extraction and memory reset module 28 outputs data from a segment track update history database 29 to a feature of interest extraction and calculation module 30, and resets the latter 29.

[0036] Module 30 for extracting and calculating features of interest also receives as input the characterization of a false track detected by the second stage 2, and delivers as output features of the track segment, i.e. values ​​of parameters representative of the track segment, including statistics of the kinematic characteristics and radar classification of track updates of the segment, for the second classifier 19.

[0037] The second classifier 19 delivers as output a classification of the track segment as false track, angel or other track, to the classification likelihood calculation module 20 which also receives as input a memory extraction performed by a memory extraction module 31 which extracts data from a likelihood vector history database 32.

[0038] The classification likelihood calculation module 20 calculates, for example, by matrix calculation involving the classification of the second classifier 19, the memory extraction of the last likelihood vector of the track delivered by module 31 and a matrix of fixed parameters then a likelihood vector, which it delivers as output to a module 33 for storing said likelihood vector in memory in the database 32 of likelihood vector history.

[0039] Based on the likelihood vector, the 21-point determination rule allows the likelihood vector values ​​to be evaluated against fixed thresholds and outputs a classification of the track as a false lead (radar angels) or another track. A memory storage module stores this information in a historical database. Track classification

[0040] The system according to the invention also includes a track update database configured to train by learning the first and second classifiers of the third stage 4.

[0041] According to one aspect of the invention, an air surveillance center can be equipped with a radar track classification system for false tracks, corresponding to radar angels, and other tracks, as described above.

[0042] The specificity of the present invention, compared to a classic Machine Learning algorithm, lies in several aspects which follow.

[0043] First, the operation is adaptive; the first stage 1 determines whether the track contains updates from a secondary source and, if not, attempts to detect an angle phenomenon by extracting its kinematic characteristics. This detection uses a partitioning or "clustering" algorithm on specific attributes of the track updates. This requires storing the track updates in memory over a time interval ΔT. The present invention is innovative and better suited than a radar sector study because the angle phenomenon is more readily formalized in a Cartesian or geographic coordinate system than in the polar coordinate system traditionally used by radar operators. Depending on the response to secondary source detection and angle phenomenon detection, the track updates are either passed or not to the third classification stage 4.This reduces the risk of misclassifying true leads as angels.

[0044] Furthermore, the structure of the third stage 4 is novel, comprising two classifiers 18, 19 arranged in series, each using a random forest algorithm. The first classifier 18 processes track updates individually according to predetermined attributes, and the second classifier 19 processes the tracks as a whole, based on calculated statistical characteristics, to which are added the predictions of the first classifier 18. The invention makes it possible both to capture local variations and to take into account the known overall characteristics of the track over a time interval.Furthermore, random forests are far less resource-intensive and faster to run than neural networks and are based on a decision tree structure, similar to a human decision-making process, thus improving explainability (the ability to relate and make understandable the elements considered by the AI ​​system to produce a result). The classification is given for a track segment and therefore requires storing track updates in memory over a time interval.

[0045] Furthermore, the invention presents another novel aspect: the classification given to each track and the detection of angel phenomena is reversible because the track is not systematically deleted when it is classified as a "false angel track," as in some existing processes, but is marked as such. This allows for periodic re-evaluation at fixed time intervals. This is also made possible by the speed of execution of the classifiers seen previously. Angel phenomenon detection is re-evaluated over wide time intervals, making it possible to monitor the local evolution of the atmospheric situation in real time and to adapt / inhibit the angel classification if the meteorological phenomenon evolves or ceases. In practice, a unit time interval Δt is determined and used for the short interval, while the long interval ΔT corresponds to a multiple of this unit interval, as in the example of the [ Fig.6 ].

[0046] Finally, another novel aspect of the solution is the construction of a likelihood vector that takes into account the history of past predictions. This vector is unique to each track and indicates, for each possible label, the associated probability at a given time. It is therefore updated at each interval, taking into account the values ​​of the previous vector, the new predictions from the classifiers, as well as a transition matrix from one state to another (adapted to each scenario), weighting the contribution of the new values.

[0047] The main advantages of this solution include: Automatic detection of the angel phenomenon, allowing for automatic adaptation of processing to its intermittent nature; automatic extraction of the phenomenon's characteristics, essential for angel processing; automatic activation of the false lead vs. other lead classification stage. The false lead / other lead classification is only activated when necessary; automatic consideration of the phenomenon's parameters by the classification stage; periodic re-evaluation of the false lead vs. other lead classification, taking into account the lead's history; and a low risk of suppressing leads on slow-moving tracks with a small radar cross-section.There is no risk of incorrect angel classification when angel phenomenon detection is not in progress; the risk remains low when angel classification is activated; the characteristics of the angel phenomenon are expressed in a universal reference frame, adapted to the physical phenomenon and understandable to an operator; it does not require human intervention to determine the treatments adapted to the intermittent phenomenon; the classification stage is based on acquired learning and does not require relearning.

Claims

1. System for classifying radar tracks, as false angel tracks, corresponding to radar angels, and as other tracks, corresponding to targets of interest, including three stages: - a first stage (1) of source classification, receiving as input a track update, configured to determine whether a track has a secondary radar source and issue as output a classification as other track if an updated track has a secondary radar source, or transmit the track update to a second stage (2); - the second stage (2) for detecting and characterizing an angel phenomenon, configured to detect an angel phenomenon and deliver as output a detection status or not of an angel phenomenon, to a first test module (3) of the status, and, in case of detection of an angel phenomenon, issue as output a characterization of the angel phenomenon detected, to a third stage (4); - the first test module (3) of the status, receiving as input the track update from the first stage and the detection status or not of an angel phenomenon from the second stage (2), configured to test the value of the detection status or not of an angel phenomenon by the second stage (2), and, in case of a status of non-detection of an angel phenomenon, issue as output a classification as other track of the updated track, and in the case of a detection status of angel phenomenon, transmit the track update to the third stage (4); - the second stage (2) for detecting and characterizing an angel phenomenon comprises: - a fourth module (10) for testing the reaching or not of the end of a repetitive time interval configured to, after storage (11) of the update of tracks in a database (12) of the history of the updates of tracks of the current time interval, if the end of the current time interval is reached to extract (13) the content from the database of the history of the updates of tracks of the current time interval, reinitialize (13) the database (12) of the history of the updates of tracks of the current time interval, and transmit updates of tracks of the current time interval to a partitioning module (14); - the partitioning module (14) configured to carry out a partitioning of the updates of the elapsed time interval, by studying the similarity of attributes of the track update comprising the direction and the speed of the track, detect (15) the presence of an angel phenomenon, issue as output a detection status or not of an angel phenomenon, and, in case of detection (15) of an angel phenomenon, issue as output a characterization of the angel phenomenon detected, to the third stage (4); and - the third stage (4), configured to carry out a classification of the updated track as false angel track or other track, by two pre-trained classifiers, arranged in series, each using an algorithm of the random forest type.

2. The system according to claim 1, wherein the first source classification stage comprises: - a second test module test (5) configured to test whether an identifier of the updated track corresponds to a track identifier having a second radar source already listed as a track with a secondary radar source, by access to a database (7) of secondary radar track histories, to issue as output a classification as other track of the updated track if the identifier corresponds to a track identifier having a secondary radar source in the database (7) of track histories having a secondary radar source, and to transmit the track update to a third test module (8) otherwise; - the third test module (8) configured to test whether the track update comes from a secondary radar source, to store (9) in memory the updated track in the database (6) of track histories having a secondary radar source if the updated track is considered to be a track coming from a secondary radar source and issue as output a classification as other track of the updated track, and to transmit the updated track to the second stage (2) otherwise.

3. System according to one of the preceding claims, wherein the third stage (4) for classifying the track update, for a time interval (ΔT), as false angel track or other track comprises: - a first pre-trained classifier (18) using an algorithm of the random forest type making it possible to issue as output a current classification of the track, when the current time interval (ΔT ) has not elapsed; - a second pre-trained classifier (19) using an algorithm of the random forest type; - a module for calculating the classification likelihood (20) of the second classifier (19); and - a rule base (21) for determining the final classification using the likelihood vector, configured to be applied as output of the module (20) for calculating the likelihood and issue as output a classification of the track as false angel track or other track, at the end of the current time interval (ΔT).

4. System according to one of the preceding claims, comprising a database (24) of updates of tracks configured to train by learning the first and second classifiers of the third stage (4).

5. Aerial surveillance center, provided with a system for classifying radar tracks as false angel tracks, corresponding to radar angels, and as other tracks, according to one of the preceding claims.