Method for providing calibration data for antenna diagrams
Patent Information
- Application Number
- EP2023748473
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-10-13
- Filing Date
- 2023-07-25
- Publication Date
- 2025-08-20
AI Technical Summary
The calibration of antenna patterns for radar sensors used in motor vehicles is time-consuming and labor-intensive due to the need for numerous measurements to achieve high angular resolution, resulting in high costs.
A method utilizing a neural network to generate calibration data by training on a reduced set of measurement points, allowing for the prediction of complete antenna diagrams from a smaller source diagram, thereby reducing the number of required calibration measurements.
Significantly reduces the costs and time associated with calibrating radar sensors by leveraging characteristic correlations in antenna diagrams, enabling accurate calibration with fewer measurements.
Smart Images
Figure 1.1
Abstract
Description
[0001] Description
[0002] title
[0003] Method for providing calibration data for antenna patterns
[0004] The invention relates to a method for providing calibration data for calibrating antenna patterns of devices configured to transmit and / or receive electromagnetic radiation and belonging to a common series.
[0005] In particular, the invention relates to the calibration of antenna patterns of radar sensors which are used, for example, in motor vehicles to detect the traffic environment.
[0006] State of the art
[0007] An angle-resolving radar sensor for motor vehicles typically has a transmitting and receiving antenna with several offset antenna elements. The angular information of located radar targets is encoded in the amplitude and phase relationships between the signals received by different antenna elements and can be extracted from the received data by comparing the amplitudes and phases of the received signals with an antenna pattern that describes the angular dependence of the amplitudes and phases. Likewise, by comparing the amplitudes and phases with the antenna pattern, the radar cross-section of a located target can be determined, thus providing a measure of the extent of the located object.
[0008] Radar sensors belonging to the same series, and therefore identical in design, should ideally all have the same antenna pattern. However, due to unavoidable manufacturing tolerances and other interference, the actual antenna patterns of the radar sensors differ slightly. To obtain precise positioning data, it is therefore necessary to compensate for these deviations by calibrating the antenna patterns individually for each radar sensor. This requires calibration data obtained by measuring the amplitudes and phases of the received signals under standardized conditions. For this purpose, radar echoes from standardized reflectors arranged at known angles relative to the radar sensor are evaluated. Each reflector position represents a measurement point for which an individual measurement must be performed.To obtain calibration data that specify the angular dependence of the amplitudes and phases for the respective sensor with high angular resolution, the angular distances between the individual measurement points should be as small as possible. However, this means that a relatively large number of measurements must be performed for each individual sensor. The number of measurements increases the larger the angular range to be covered by the calibration data, and this increases exponentially for radar sensors that have angular resolution in two dimensions (azimuth and elevation). Since the necessary calibration measurements are time-consuming and labor-intensive, accurate sensor calibration is very expensive.
[0009] Disclosure of the invention The object of the invention is to minimize the number of required calibration measurements for a given quality of the calibration data.
[0010] This object is achieved according to the invention by a method which is characterized by the following steps: a) selecting a set of devices from the series for creating a training file, b) for each selected device: measuring a target diagram with a first number of measuring points and creating an associated source diagram with a second number of measuring points which is smaller than the first number, and storing the target and source diagrams in the training file, c) using the training file to train a neural network to determine the associated target diagrams based on the source diagrams, d) measuring the source diagrams of the remaining devices from the series, e) creating the associated target diagrams using the neural network, and f) using these target diagrams as calibration data for the remaining devices.
[0011] The above-mentioned "devices" can be radar sensors or other systems with which electromagnetic radiation can be transmitted and / or received. "Target diagrams" are understood to be antenna diagrams that reproduce the angular dependence of the amplitudes and phases of the received signals with the required angular resolution. "Source diagrams" are understood to be antenna diagrams based on a reduced number of measurement points compared to the target diagrams. The invention is based on the finding that, for sensors (devices) of the same series, there are characteristic correlations between different parts of the antenna diagram, so that, based on measurement results obtained for measurement points in one part of the diagram, feedback can be obtained.
[0012] 5 conclusions can be drawn about the appearance of other parts of the diagram without actually having to measure these other parts of the diagram. According to the invention, a neural network is trained to recognize these correlations using suitable training data. Once the neural network is trained, it is sufficient to measure only the source diagram for a single sensor with a relatively small number of measurement points and input this data into the neural network, which then generates the complete target diagram, which can then be used to calibrate the sensor.
[0013] In this way, the costs for calibrating the sensors can be reduced considerably.
[0014] Advantageous embodiments and further developments of the invention are specified in the subclaims. 0
[0015] The source diagrams can be diagrams that cover the same angular range as the corresponding target diagrams, but with reduced angular resolution due to the smaller number of measurement points. ^
[0016] On the other hand, the source diagrams can also cover an angular range that is smaller than the angular range of the target diagrams. In this case, the calibration data for the larger angular range are extrapolated using the neural network. For sensors with angle resolution in two dimensions, the source diagram can represent a one-dimensional section through the two-dimensional measurement field (e.g., in azimuth), while the target diagram determined by the neural network additionally includes calibration data for at least one section in the second dimension (in elevation).
[0017] The complete target graphs are measured for the sensors (devices) selected for creating the training file. The corresponding source graphs can then be generated without additional measurements simply by reducing the target graph to a smaller set of measurement points. When training the network, the source graphs then form the input data, and the corresponding target graphs provide the feedback, based on which the weights of the neural connections in the network are adjusted, for example, using backpropagation.
[0018] In one embodiment, calibration data is stored for each sensor, which contains both the source diagram and the target diagram, wherein, if the source diagram is obtained by reducing the measurement points, the source diagram is already contained in the target diagram.
[0019] In another embodiment, the measured source diagram is stored in each sensor to be calibrated, but instead of the target diagram, the trained neural network is stored in the sensor, which then provides the necessary calibration data during sensor operation. This embodiment is advantageous in cases where the data volume of the parameters of the trained neural network is smaller than the data volume of the complete target diagram.
[0020] The invention also relates to a neural network which has been trained for generating the calibration data according to the method described above, as well as to a radar sensor in which such a neural network is stored for generating the calibration data.
[0021] In the following, exemplary embodiments are explained in more detail using the drawings.
[0022] They show:
[0023] Fig. 1 shows an example of a measuring field of a radar sensor;
[0024] Fig. 2 Examples of four amplitude antenna patterns for a
[0025] Section through the measuring field along the line II - II in Fig. 1 ;
[0026] Fig. 3 examples of amplitude antenna diagrams of the same four radar sensors as in Fig. 2 corresponding to a section through the measuring field along the line III - III in Fig. 1;
[0027] Fig. 4 shows a measurement field with differently extending section lines; Fig. 5 shows an example of an amplitude and phase antenna diagram corresponding to a section through the measurement field along the line V-V in Fig. 4;
[0028] Fig. 6 an amplitude and phase antenna diagram corresponding to a section through the measuring field along the line VI -
[0029] VI in Fig. 4 ;
[0030] Fig. 7 is a diagram for explaining a method according to a
[0031] Embodiment of the invention; Fig. 8 shows a flowchart for a method for training a neural network;
[0032] Fig. 9 is a flowchart for a first embodiment of the inventive method for providing calibration data;
[0033] Fig. 10 is a flowchart for a method according to a second embodiment of the invention; and
[0034] Fig. 11 and 12 diagrams analogous to Fig. 7 for methods according to modified embodiments of the invention.
[0035] Fig. 1 shows a measuring field 10 of a radar sensor that has angle resolution in two dimensions, i.e., both in azimuth (horizontal axis in the diagram) and in elevation (vertical axis). The amplitudes (magnitudes of the complex amplitudes) of the signals received from different directions are represented by contour lines 12. In a conventional method for providing calibration data for a radar sensor, a horizontal section, for example, corresponding to line II in Fig. 1, and a vertical section, for example, corresponding to line III in Fig. 1, are drawn through the measuring field 10, and on each section line, the amplitudes of the received signals and the phase differences between the various antenna elements of the radar sensor are measured with the highest possible angular resolution.
[0036] Fig. 2 shows the results of amplitude measurements on line II in Fig. 1 for four identical radar sensors. The four curves 14a, 16a, 18a, and 20a visible in Fig. 2 indicate the amplitude as a function of the azimuth angle for each of the four sensors. Each curve is part of an antenna pattern for the respective radar sensor. It can be seen that all four curves have a similar shape; however, due to manufacturing tolerances, material properties, and the like, there are certain deviations between the curves and from an ideal antenna pattern that is typical for sensors of this type.
[0037] series would theoretically be expected.
[0038] Fig. 3 shows corresponding curves 14b, 16b, 18b and 20b measured along line III in Fig. 1.
[0039] The example shown in Figures 2 and 3 illustrates a phenomenon that the invention described here makes use of. This phenomenon consists in the fact that deviations from sensor to sensor in one part of the antenna pattern correlate strongly with deviations in other parts of the antenna pattern. For example, Figure 2 shows that curves 14a and 16a have a very similar course and deviate only slightly from each other. This small deviation between the curves correlates with a likewise only small deviation between the corresponding curves 14b and 16b in Figure 3, i.e. in the part of the antenna pattern that corresponds to the section in the elevation direction. In particular, in this example, curve 14b exhibits a noticeable anomaly in the elevation angle range between 12 and 15°, which is marked by an arrow in Figure 3. This anomaly consists in the fact that the curve is flatter in this range than in the adjacent angular ranges.The same anomaly is also evident in curve 16b. ^.
[0040] Observations of this kind indicate that sensors that exhibit very similar behavior in one part of the antenna pattern (here in the section along line II) also exhibit similar behavior in other parts of the pattern (here in the section along line III). According to the invention, a neural network is therefore used to recognize such regularities or patterns in the antenna patterns of different sensors and then, based on these regularities, to predict the shape of the antenna pattern in areas where no measurements were performed.
[0041] Figures 4 to 6 first illustrate a conventional method for providing calibration data by measuring antenna patterns. Fig. 4 shows a measurement field 22 of a radar sensor with section lines V, VI running through the center of the measurement field, on which the antenna pattern is to be measured. Fig. 5 shows the amplitude and phase antenna pattern recorded along section line V. For this purpose, numerous measurements were carried out at measuring points 24 distributed along section line V. Each measuring point 24 represents a measurement in which a reflector, whose radar echo was measured, was positioned on section line V such that its azimuth angle corresponds to the position of the measuring point on the horizontal axis in Fig. 5. The upper part of Fig. 5 shows the amplitude antenna pattern for an individual radar sensor, i.e. the measured amplitude (in dB) as a function of the azimuth angle (in °). In the lower part in Fig.Figure 5 shows the phase antenna diagram, i.e., the phase difference between adjacent antenna elements as a function of the azimuth angle. This phase difference varies in the range from -% to +%, with the values -% and +% being equivalent, as they both correspond to a phase angle of 180°. It should be understood that the phase differences were measured at the same measuring points 24 as the amplitudes.
[0042] Fig. 6 shows the corresponding amplitude and phase antenna pattern recorded on section line IV. Here, too, the amplitude and phase measurements were performed at a large number of measurement points 26 evenly distributed along section line IV. Even if one restricts oneself to measuring the antenna pattern on section lines V and VI, this method requires a large number of individual measurements (corresponding to the sum of measurement points 24 and 26) for each sensor.
[0043] Fig. 7 illustrates a method that can significantly reduce the number of individual measurements required. In this method, for each sensor to be calibrated, actual measurements are performed only at a reduced number of measurement points, resulting in a source pattern 28 that represents only a portion of the complete antenna pattern.
[0044] In the example shown in Fig. 7, the source diagram 28 is the amplitude and phase antenna diagram for section line V in Fig. 4. The source diagram 28 thus represents only the azimuthal section through the measurement field, so that no measurements need to be performed on the elevation section line VI. Accordingly, the number of individual measurements required corresponds only to the number of measurement points 24 on section line V.
[0045] Using a neural network 30 specifically trained for this application, a synthetic diagram 32 is then generated, which indicates the amplitudes and phase differences on section line VI in Fig. 4 and thus completes the antenna diagram. The source diagram 28 and the synthetic diagram 32 together form a so-called target diagram 34, which provides the calibration data for calibrating the radar sensor.
[0046] The neural network 30 has an input stage IN, which feeds the real and imaginary parts of the complex amplitudes for each measurement point 24 of the source diagram 28 into neurons 38 of a first layer 40 of the neural network. In hidden layers 42, the information from the neurons 38 of the first layer 40 is then processed step by step until, finally, in an output layer 44, a predicted value of the real or imaginary parts of the complex amplitude is obtained in each neuron. This value corresponds to a measurement point 26 on the section line VI in Fig. 4. The amplitude and phase values, which together form the synthetic antenna diagram 32, are then output via an output stage OUT of the neural network.
[0047] The neural network 30 can have any architecture known for neural networks. In the example shown, at least the lower layers 40, 42 form a fully interconnected network, in which each neuron in the first layer 40 influences the state of every neuron in the subsequent layer 42. Alternatively, the neural network 30 can also be a convolutional neural network, as is also frequently used in pattern recognition algorithms.
[0048] However, before the neural network can perform the function depicted in Fig. 7, it must be trained using suitable training data. The essential steps of a network training procedure are depicted as a flowchart in Fig. 8.
[0049] In step S1, a certain number N of sensors is selected from the series of radar sensors for which the calibration data is to be provided, and for each of these sensors, a source diagram 28 and a target diagram 34 are measured in a conventional manner according to the method shown in Figures 4 to 6. However, the data for the source diagrams 28 do not need to be measured separately, since the measurement results are obtained anyway when measuring the complete target diagram 34. The totality of the source diagrams 28 and target diagrams 34 obtained in this way in step S1 for all N sensors forms a training file 36, which provides the training data for the neural network 30.
[0050] In step S2 the actual training of the neural network 30 takes place
[0051] 5. Before training, the network 30 is in an initial configuration in which each neural connection between a neuron 38 of one layer and a neuron of the downstream layer has a specific weight that determines how strongly and in which direction the state of the upstream neuron changes compared to the state of the downstream neuron. In a first training step, a first source diagram 28 is then input to the network via the input stage IN. The result obtained from the output stage OUT is compared with the corresponding target diagram 34 from the training file 36. Based on the deviation of the result from the target diagram 34, the weights of the neural connections are changed such that when the same input is input again, the result is closer to the desired target diagram 34.In the next training step, a source diagram 28 for a different sensor is input into the network, and the weights are changed again based on the result. If the number N of pairs of source and target files in the training file 36 is large enough, the weights gradually converge to a configuration in which the network can predict the corresponding target diagram 34 relatively precisely for each source diagram 28. The totality of the weights then defines a transfer function 38 that assigns the corresponding target diagram to each source diagram. ^.
[0052] In practice, a validation phase will usually follow, in which a further number of sensors from the series are selected to test the performance of the neural network 30. If the test is positive, the network is ready for use, so that calibration data (ie, target files 34) can be generated for each sensor in the series using the method shown in Fig. 7.
[0053] Fig. 9 is a flowchart for a possible method for calibrating a sensor. In this method, a source diagram 28 for the sensor to be calibrated is measured in a first step S11. In step S12, this source file is input into the neural network 30, and the target diagram 34 is generated using the neural network or the transfer function 38 defined thereby. In step S13, the target diagram 34, which also contains the original source diagram 28 as a partial data set, is then saved. This target diagram 34 then provides the calibration data with which the sensor is calibrated before commissioning.
[0054] An alternative method is shown in Fig. 10. In this method, the first step S11 is identical to step S11 in Fig. 9, i.e., a source diagram is measured for the sensor to be calibrated. In a subsequent step 12', the source diagram 28 and the neural network 30 are stored in a digital memory of the sensor. Storing the neural network means storing the network pattern of the neurons (the connectome) as well as the weights of all neural connections.
[0055] After step S12', the sensor is already ready for use. Only when the sensor is put into operation is the neural network 30 run in the sensor's electronics in a further step S13' to generate the target diagram 34 and thus the calibration data with the aid of the transfer function 38. Calibration can be limited to those parts of the antenna diagram for which current positioning data are available. Fig. 11 illustrates an alternative method for providing calibration data in a diagram analogous to Fig. 7. For the sake of simplicity, it will be assumed here that the sensor to be calibrated has angle resolution only in azimuth.
[0056] In this case, a source diagram 28' only includes measurement points 24' that lie within a limited azimuth angle range. Based on the source diagram 28, the neural network 30 extrapolates the amplitudes and phase differences to the entire angular range, thus delivering a synthetic diagram 32' that extends over the entire detection angular range of the radar sensor and contains the source diagram 28' as a partial data set. The amplitude curve outside the angular range of the source diagram 28 can be represented, for example, by value pairs for discrete "measurement points" (or better: support points). The number of support points can be greater than the number of measurement points 24' in the original source diagram 28'. In this case, the neural network 30' has a larger number of neurons in the output layer than in the first layer.
[0057] Another process variant is illustrated in Fig. 12.
[0058] A 28" source diagram has 24" measurement points evenly distributed across the entire azimuth angle range. However, the number of these measurement points is relatively small, resulting in a low angular resolution. In this case, a 30" neural network is trained to interpolate intermediate values between the 24" measurement points, resulting in a 32" synthetic diagram with higher angular resolution.
[0059] It is understood that the methods described above in connection with Figures 7, 11 and 12 can be combined as required.
Claims
Claims 1 . A method for providing calibration data for calibrating antenna patterns of devices configured to transmit and / or receive electromagnetic radiation and belonging to a common series, characterized by the following steps: a) selecting a set of devices from the series for creating a training file (36), b) for each selected device: measuring a target pattern (34) with a first number of measurement points (24, 26; 24'; 24") and creating an associated source pattern (28; 28'; 28") with a second number of measurement points (24; 24'; 24") that is smaller than the first number, and storing the target and source patterns (34, 28; 28'; 28") in the training file (36), c) using the training file (36) to create a neural network (30; 30'; 30") to determine the respective associated target diagrams (34) on the basis of the source diagrams (28; 28'; 28"), d) measuring the source diagrams (28; 28'; 28") of the remaining devices from the series, e) creating the associated target diagrams (34) with the aid of the neural network (30; 30'; 30"), and f) using these target diagrams (34) as calibration data for the remaining devices.
2. Method according to claim 1, wherein the neural network (30; 30'; 30") is trained to generate a synthetic diagram (32; 32'; 32") which, together with the source diagram (28; 28'; 28"), forms the target diagram (34).
3. Method according to claim 1 or 2, wherein the source diagrams (28') have a smaller range of values than the target diagrams (34) and the neural network (30') is trained to determine the target diagram (34) by extrapolating the values from the source diagram (28').
4. Method according to one of the preceding claims, in which the neural network (30") is trained to interpolate between measurement points (24") of the source diagram (28") when creating the target diagram (34).
5. Method according to one of the preceding claims, for devices configured for two-dimensional angle measurements, wherein the source diagrams (28) contain data for only one dimension and the neural network (30) is trained to generate data for the second dimension from the data of the source diagram (28).
6. A method for calibrating antenna patterns of devices configured to transmit and / or receive electromagnetic radiation and belonging to a common series, comprising a method for providing calibration data according to any one of the preceding claims, characterized in that the target patterns (34) are used to calibrate the devices before commissioning.
7. Method for calibrating antenna patterns of devices configured to transmit and / or receive electromagnetic radiation and belonging to a common series, with a method for providing calibration data according to one of claims 1 to 6, characterized in that for each device to be calibrated, the measured source pattern (28; 28'; 28") is stored in the device together with parameters of the neural network (30; 30'; 30") and during the Operation of the facility to provide calibration data and calibrate the antenna pattern.
9. A neural network trained to perform step e) of the method according to claim 1.
10. Radar sensor for motor vehicles, comprising a memory for a source diagram (28) measured in step d) of a method according to claim 1, and comprising electronics in which a neural network (30; 30'; 30") according to claim 9 is implemented.