DE102019100320A1 - DF method and DF system - Google Patents

Abstract

A direction finding method has the following steps: a) Receiving a signal by means of at least two antenna elements (12) by a plurality of receivers (14) and digitizing the received signal, at least one sample (S) of the received signal being obtained, each of which has at least one measured value (x, x) the receiver (14) contains, b) detecting at least one signal pulse in the received signal, whereby noise is rejected, c) preprocessing the at least one sample (S) of the received signal by determining a complex covariance matrix (M) at least one sample (S), d) converting the complex covariance matrix (M) into a real vector (V), unde) supplying the vector (V) to an artificial neural network (26), the artificial neural network (26) den The angle of incidence of the signal, an indication of the quality of the signal and / or an indication of the confidence of the bearing are determined. A bearing system (10) is also shown.

Description

The invention relates to a DF method and a DF system with at least two antenna elements.

Known DF methods are very complex and are therefore resource-intensive with acceptable performance. Among other things, this leads to large systems. However, a compact design is desirable in particular for mobile direction finding systems.

It is therefore an object of the invention to provide a DF method and a DF system which have a lower resource consumption with constant performance.

1. a) Empfangen von einem Signal mittels der wenigstens zwei Antennenelemente durch die Empfänger und Digitalisieren des empfangenen Signals, wobei wenigstens ein Sample des empfangenen Signals erhalten wird, das jeweils wenigstens einen Messwert der Empfänger enthält,
2. b) Detektieren von wenigstens einem Signalpuls im empfangenen Signal, wobei Rauschen verworfen wird,
3. c) Vorverarbeiten des wenigstens einen Samples des empfangenen Signals durch Bestimmen einer komplexen Kovarianzmatrix des wenigstens einen Samples,
4. d) Konvertieren der komplexen Kovarianzmatrix in einen reellwertigen Vektor, und
5. e) Zuführen des Vektors zu dem künstlichen neuronalen Netzwerk des Maschinenlernmoduls, wobei das künstliche neuronale Netzwerk den Einfallswinkel des Signals, eine Angabe zur Qualität des Signals und/oder eine Angabe zur Konfidenz der Peilung bestimmt.

The object is achieved by a DF method using a DF system which has at least one antenna unit with at least two antenna elements and at least two receivers, each associated with one of the antenna elements, and a machine learning module with an artificial neural network. The process comprises the following steps:

1. a) Receiving a signal by means of the at least two antenna elements by the receivers and digitizing the received signal, at least one sample of the received signal being obtained, each of which contains at least one measured value of the receivers,
2. b) detecting at least one signal pulse in the received signal, noise being rejected,
3. c) preprocessing the at least one sample of the received signal by determining a complex covariance matrix of the at least one sample,
4. d) converting the complex covariance matrix into a real vector, and
5. e) supplying the vector to the artificial neural network of the machine learning module, the artificial neural network determining the angle of incidence of the signal, an indication of the quality of the signal and / or an indication of the bearing confidence.

By using an artificial neural network, which in particular only receives real-value input values, the complexity of the direction finding process is reduced, as a result of which the resource consumption can also be reduced without any loss of performance. On the one hand, this enables an improved performance with unchanged use of resources, but on the other hand also a higher integration density and therefore a more compact design with constant performance. As a result, the size of DF systems, in particular mobile DF systems, can be reduced.

The signal or the measured values of the signal can be expressed using I data and Q data (in-phase & quadrature method).

For example, the real-valued vector is twice as long as the number of elements contained in the complex covariance matrix.

The angle of incidence or the bearing can include the azimuth and the elevation.

For example, all noise is discarded.

To further improve the performance, a predetermined frequency band can be selected from the signal or the measured value for further processing. This can already be done by the recipient. For example, the measured values are the antenna voltages.

The predetermined frequency band is preselected by the user, for example.

The at least one sample preferably contains measured values from each of the at least two receivers, as a result of which the signal information of each antenna element is used to determine the bearing.

In one embodiment of the invention, a pulse-describing data word is determined for the signal pulse, in particular for each signal pulse, as a result of which the signal is precisely described.

A pulse describing data word (PDW) is also called "pulse descriptor word".

In one embodiment of the invention, the information about the quality of the signal contains an estimated error in signal reception and / or the polarization of the signal and / or the information about the confidence contains an estimated error of the angle of incidence. In this way, the signal can be clearly classified.

In order to further reduce the complexity of the direction finding process and to increase the accuracy, further information can be added to the vector, in particular the frequency of the signal and / or the signal strength of the signal.

In one embodiment of the invention, the signal received by each of the at least two receivers is integrated over the corresponding reception duration, as a result of which an accumulated sample, in particular an accumulated measured value, is generated Antenna element is obtained. This can further improve the accuracy of the bearing.

The accumulated sample can be processed in the same way as the previously described samples.

For example, the last layer of the artificial neural network provides a soft max function and / or combines the I data and Q data of the signal. This can simplify the output of the artificial neural network.

• - Vorwärtsspeisen des künstlichen neuronalen Netzwerks mit den Soll-Trainingsdaten,
• - Bestimmen der Signaleigenschaften mittels des künstlichen neuronalen Netzwerks,
• - Ermitteln eines Fehlers zwischen den bestimmten Signaleigenschaften und den tatsächlichen Signaleigenschaften, und
• - Ändern von Gewichtungsfaktoren des künstlichen neuronalen Netzwerks durch Rückwärtsspeisen des künstlichen neuronalen Netzwerks mit dem Fehler.

In one aspect of the invention, the artificial neural network is trained with target training data, the target training data measured signals and / or synthetic signals as well as information on the actual signal properties, in particular the actual angle of incidence of the signal, the quality of the signal and / or Bearing Confidence Include. The following training steps are carried out:

• - feeding the artificial neural network with the target training data,
• Determining the signal properties by means of the artificial neural network,
• Determining an error between the specific signal properties and the actual signal properties, and
• - Changing weighting factors of the artificial neural network by feeding back the artificial neural network with the error.

In this way, the accuracy of the bearing can be greatly improved.

In a further embodiment of the invention, the determined angle of incidence of the signal, the information on the quality of the signal and / or the information on the confidence of the bearing are made available to an evaluation module of the direction-finding system, as a result of which further processing steps are possible.

For example, the information and / or the angle of incidence are used as a pre-classifier for a known DF method, for example a DF method based on the maximum likelihood method.

Furthermore, the object is achieved by a direction finding system with at least one antenna unit, the at least two antenna elements, at least two receivers, each associated with one of the antenna elements, and at least two analog-digital converters for digitizing the signal, each associated with one of the receivers , has a machine learning module with an artificial neural network and a control unit. The receivers are set up to receive a signal at the respective antenna element and to select a predetermined frequency band therefrom. The control unit is set up to control the direction finder system for carrying out a previously described method according to the invention.

The features and advantages described for the DF method apply equally to the DF system and vice versa.

For example, the artificial neural network has at least one completely connected position in order to further improve the classification and / or determination of the bearing.

The artificial neural network can be a convolutional neural network (CNN).

In one embodiment of the invention, the machine learning module is a field programmable gate array (FPGA), a microcontroller or a graphics processor (GPU), as a result of which compact designs are achieved.

• - 1 ein erfindungsgemäßes Peilsystem,
• - 2 ein Ablaufdiagramm eines erfindungsgemäßen Verfahrens, und
• - 3 eine mathematische Veranschaulichung von Teilen des Verfahrens nach 2.

Further features and advantages of the invention will become apparent from the following description and from the accompanying drawings, to which reference is made. The drawings show:

• - 1 a bearing system according to the invention,
• - 2nd a flowchart of a method according to the invention, and
• - 3rd a mathematical illustration of parts of the method according to 2nd .

In 1 is a schematic DF system 10th with which the bearing to a signal source can be determined, that is to say to determine the angle of incidence of the signal, including azimuth and elevation.

The DF system 10th has an antenna unit 11 with several antenna elements 12th , seven antenna elements in the exemplary embodiment shown 12th , multiple recipients 14 and several analog-digital converters 16 . Each is a recipient 14 and an analog-to-digital converter 16 one of the antenna elements 12th assigned and electrically connected to it. That is, in the exemplary embodiment shown there are also seven receivers 14 and seven analog-to-digital converters 16 available.

The DF system 10th also has a control unit 18th which is a preprocessing module 20 , a machine learning module 22 and a post-processing module 24th or evaluation module, and an output unit 32 .

The preprocessing module 20 , the machine learning module 22 and the post-processing module 24th can also be separated from the control unit 18th be executed. The control unit 18th controls the function of the receiver in any case 14 , the analog-digital converter 16 , the preprocessing module 20 of the machine learning module 22 and the post-processing module 24th and the output unit 32 .

The machine learning module 22 is, for example, as a field programmable gate array (FPGA), as a microcontroller or as a graphics processor (GPU) as part of the control unit 18th or carried out separately. It is also conceivable that the machine learning module 22 is a software module created by the control unit 18th is carried out, for example by an FPGA, a microcontroller or a GPU of the control unit 18th .

The machine learning module 22 has an artificial neural network 26 . The artificial neural network 26 is, for example, a convolutional neural network (CNN) that has several layers. At least one of the layers is a fully connected layer 28 (Fully Connected Layer).

The last location 30th of the artificial neural network 26 is also trained to provide a Softmax function - a normalized exponential function.

The recipients 14 are designed to measure the antenna voltage of the antenna element assigned to them 12th to measure and this measured value of the antenna voltage to the associated analog-digital converter 16 pass on.

The analog-digital converter 16 are trained to digitize this measured value.

In addition, the recipients 14 suitable for selecting a frequency band from the measured value determined so that only those from the antenna elements 12th received signals are processed that lie within this frequency band.

The control unit is for bearing 18th trained the DF system 10th to carry out the in 2nd to control illustrated DF procedure.

In 3rd The method is illustrated mathematically, with the simplification of a DF system 10th with only two antenna elements 12th was assumed. Of course, the method can also be applied to any number of antenna elements.

A signal from a signal source to be targeted strikes the antenna elements 12th the DF system 10th . The first step is to determine the bearing of the signal source S1 by the recipient 14 each on the antenna elements 12th incoming signal received. The signal becomes, for example, by the receiver 14 , selected the part that lies within a predetermined frequency band. This frequency band is, for example, from the user of the DF system 10th previously selected.

In addition, in this first step S1 that of the recipients 14 received signal digitized, taking samples S of the signal can be obtained.

As in 3rd to see contains every sample S the different measurements x ₁ , x ₂ the recipient 14 at the time of the sample S .

The measured values are, for example, the antenna voltage of the respective antenna element 12th .

The measured values x ₁ , x ₂ the recipient 14 form a sample. The measured values x ₁ , x ₂ can also be represented by I data and Q data in the so-called I & Q method (in-phase & quadrature method), as by I ₁ , Q ₁ which is the reading x ₁ can represent, and I ₂ , Q ₂ which is the reading x ₂ can represent in 3rd is indicated.

The samples S will be in the next step S2 to the control unit 18th , more precisely the preprocessing module 20 , transfer. The preprocessing module 20 preferably completely discards the noise of the signal and detects a signal pulse in the received signal.

In the next step S3 is at least for each sample S within the signal pulse, especially for all samples S individually, the complex covariance matrix M formed as in the second representation of the 3rd can be seen. Determining the complex covariance matrix M takes place according to usual procedures.

It is easy to see that the covariance matrix M the measured values x ₁ , x ₂ all antenna elements 12th or recipient 14 contains.

In the third representation in 3rd is shown by way of illustration that each matrix element is complex and can therefore be represented as I data and Q data.

In the next step S4 converts the preprocessing module 20 the complex covariance matrix M into a real vector V as in the fourth representation of the 3rd is shown.

Again, this is done with the help of known methods and algorithms.

The real valued vector V contains at least twice as many elements or dimensions as the covariance matrix M Had elements. This is because both the real part and the imaginary part of each matrix element or I data and Q data of the covariance matrix M as own elements in the vector V be taken into account. In the exemplary embodiment shown, the I data and Q data are used to represent the complex matrix elements and form individual elements of the vector V .

You can also use the vector V additional elements containing information about the signal may be added to facilitate the bearing.

In the exemplary embodiment shown, this is, for example, the frequency f of the signal and the signal strength s of the signal.

However, other parameters are also conceivable that are helpful for determining the bearing.

The real vector obtained in this way V will be in the next step S5 the machine learning module 22 fed, more precisely to the artificial neural network 26 .

The artificial neural network 26 determined on the basis of the vector V the angle of incidence of the signal, i.e. the bearing, in particular by specifying the azimuth and the elevation.

The last location 30th of the artificial neural network 26 combines the I data and Q data of the signal and / or provides a Softmax function to determine the bearing.

In addition, the artificial neural network 26 Determine information on the quality of the signal, for example an estimated error in signal reception and / or the polarization of the signal, or an indication of the confidence of the bearing, for example by specifying an estimated error of the angle of incidence.

The bearing and the information on quality and / or confidence can then be made using the output unit 32 be issued.

The bearing, the indication of the quality of the signal and / or the indication of the confidence of the bearing can be the post-processing module 24th to be provided.

The post-processing module 24th can determine a pulse-describing data word (PDW) - also called "Pulse Descriptor Word" - for each signal pulse, which describes the type of signal pulse (step S6 ).

It is also conceivable that the particular bearing, the information on quality and / or confidence in the post-processing module 24th used as a pre-classifier to determine the bearing using previously known methods, for example the maximum likelihood method (step S7 ). In this case, the post-processing module performs 24th the bearing and therefore represents an evaluation module.

As in 2nd recognizing is the step S1 a training step T upstream, in which the artificial neural network 26 was trained.

In training step T, the artificial neural network is first 26 Target training data supplied (feed forward), the target training data comprising measured and / or synthetic signals that the DF system 10th , more precisely the control unit 18th or the artificial neural network 26 , are fed.

In addition, the target training data contain information on the actual signal properties, that is to say the actual angle of incidence of the signal, the quality of the signal and / or the confidence in the bearing. The target training data thus contain a signal such as that from the antenna elements 12th and the bearing to the signal source of this training signal, ie the correct result of the bearing procedure to be carried out.

The artificial neural network 26 thus determines the signal properties of the measured and / or synthetic signal, ie the bearing, the quality of the signal and / or the confidence of the bearing.

Then the error between those of the artificial neural network 26 determined signal properties and the actual signal properties - which are known from the target training data - and the weighting factors of the artificial neural network 26 are adjusted by feeding the error backwards.

These steps are repeated with different target training data until the determined error is sufficiently small. Of course, the error for the application of the pre-classification does not have to be as small as if the artificial neural network 26 determines the bearing completely independently.

It is also conceivable that not individual samples S processed, but the recipient 14 integrate the received signal over the corresponding reception duration, which was previously determined, for example. In this way, an accumulated measurement value per antenna element 12th get that together an accumulated sample S form.

The accumulated sample S can - like the individual samples described above S - by means of the preprocessing module 20 , the machine learning module 22 or the artificial neural network 26 and the post-processing module 24th to be processed further.

In this way, the bearing can be determined using an artificial neural network 26 perform, which enables better performance and scalability, a higher integration density and thus a more compact design of the DF system 10th allow.

Claims (13)

DF method using a DF system (10), which has at least one antenna unit (11) with at least two antenna elements (12) and at least two receivers (14), each associated with one of the antenna elements (12), and a machine learning module (22) with an artificial one Neural network (26), the method comprising the following steps: a) receiving a signal by means of the at least two antenna elements (12) by the receiver (14) and digitizing the received signal, at least one sample (S) of the received Signal is obtained, each containing at least one measured value (x ₁ , x ₂ ) of the receiver (14), b) detecting at least one signal pulse in the received signal, rejecting noise, c) preprocessing the at least one sample (S) of the received signal by determining a complex covariance matrix (M) of the at least one sample (S), d) converting the complex covariance matrix (M) into a real-valued Ve ktor (V), and e) feeding the vector (V) to the artificial neural network (26) of the machine learning module (22), the artificial neural network (26) the angle of incidence of the signal, an indication of the quality of the signal and / or determines an indication of the bearing's confidence. DF procedure according to Claim 1 , characterized in that a predetermined frequency band is selected from the signal for further processing. DF procedure according to Claim 1 or 2nd , characterized in that the at least one sample (S) contains measured values (x ₁ , x ₂ ) of each of the at least two receivers (14). Direction finding method according to one of the preceding claims, characterized in that a pulse-describing data word is determined for the signal pulse, in particular for each signal pulse. Direction finding method according to one of the preceding claims, characterized in that the information about the quality of the signal contains an estimated error in signal reception and / or the polarization of the signal and / or that the information about the confidence contains an estimated error of the angle of incidence. Direction finding method according to one of the preceding claims, characterized in that the vector (V) further information is added, in particular the frequency (f) of the signal and / or the signal strength (s) of the signal. Direction finding method according to one of the preceding claims, characterized in that the signal received by each of the at least two receivers (14) is integrated over the corresponding reception duration, as a result of which an accumulated sample (S), in particular an accumulated measured value (x ₁ , x ₂ ) each Antenna element (12) is obtained. Direction finding method according to one of the preceding claims, characterized in that the last layer (30) of the artificial neural network (26) provides a Softmax function and / or combines the I data and Q data of the signal. Direction finding method according to one of the preceding claims, characterized in that the artificial neural network (26) is trained with target training data, the target training data being measured signals and / or synthetic signals as well as information on the actual signal properties, in particular the actual angle of incidence of the signal , the quality of the signal and / or the confidence of the bearing, with the following training steps: - feeding the artificial neural network (26) with the target training data, - determining the signal properties using the artificial neural network, - determining an error between the determined signal properties and the actual signal properties, and - changing weighting factors of the artificial neural network (26) by feeding back the artificial neural network (26) with the error. DF method according to one of the preceding claims, characterized in that the certain angles of incidence of the signal, the information on the quality of the signal and / or the information on the confidence of the bearing are made available to an evaluation module of the bearing system. DF system with at least one antenna unit (11), the at least two antenna elements (12), at least two receivers (14), each associated with one of the antenna elements (12), and at least two analog-digital converters (16) for digitizing the signal each of which is assigned to one of the receivers (14), has a machine learning module (22) with an artificial neural network (26) and a control unit (18), wherein the receivers (14) are set up to receive a signal at the respective antenna element (12) and to select a predetermined frequency band therefrom, and wherein the control unit (18) is set up to control the direction finding system (10) for carrying out a method according to one of the preceding claims. DF system according to Claim 11 , characterized in that the artificial neural network (26) has at least one completely connected layer (28). DF system according to Claim 11 or 12th , characterized in that the machine learning module (22) is a field programmable gate array (FPGA), a microcontroller or a graphics processor (GPU).

Images