Method for providing calibration data for antenna pattern
By training a neural network to identify the correlation of the radar sensor antenna pattern and generate a complete target pattern, it solves the problem of high cost resulting from the large number of calibration measurements in the prior art, and realizes a cost-effective calibration method.
Patent Information
- Application Number
- CN202380072449.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-10-13
- Filing Date
- 2023-07-25
- Publication Date
- 2025-05-30
AI Technical Summary
The prior art requires a lot of measurements when calibrating the antenna pattern of radar sensors, resulting in high costs.
By selecting devices in a set of structural series to create training files, measuring target and source maps, the neural network is trained to identify correlations, thereby generating a complete target map for calibration.
The number of calibration measurements required is significantly reduced and the cost of sensor calibration is reduced.
Smart Images
Figure CN120077587A_ABST
Abstract
Description
Field of the Invention
[0001] The present invention relates to a method for providing calibration data for calibrating the antenna pattern of devices configured for transmitting and / or receiving electromagnetic radiation and belonging to a common construction series.
[0002] The present invention particularly relates to the calibration of the antenna pattern of radar sensors, for example, using these radar sensors in motor vehicles to detect the traffic environment. Background Art
[0003] Angularly resolved radar sensors for motor vehicles typically have a transmitting antenna and a receiving antenna with a plurality of antenna elements arranged offset from one another. The angular information of the located radar target is encoded in the amplitude relationship and phase relationship between the signals received by different antenna elements and can be extracted from the received data by aligning the amplitude and phase of the received signals with an antenna pattern that describes the angular dependence of the amplitude and phase. Similarly, the radar cross-section of the located target can be determined by aligning with the antenna pattern, and thus the extent measure of the located object can be determined.
[0004] Radar sensors belonging to the same construction series and thus having the same structure should ideally all have the same antenna pattern. However, due to inevitable manufacturing tolerances and other interfering influences, the actual antenna patterns of the radar sensors deviate from each other to some extent. In order to obtain accurate positioning data, it is therefore necessary to compensate for these deviations by calibrating the antenna pattern individually for each radar sensor. For this purpose, calibration data is required, which is obtained by measuring the amplitude and phase of the received signal under standardized conditions. For this purpose, the radar echoes of standardized reflectors are analyzed, which are arranged at known angles relative to the radar sensor. Each reflector position represents a measurement point for which a single measurement must be performed. In order to obtain calibration data that gives the angular dependence of the amplitude and phase with high angular resolution for the relevant sensor, the angular spacing between the individual measurement points should be as small as possible. But this means that a relatively large number of measurements must be performed for each individual sensor. The more measurements are made, the larger the angular range that should be covered by the calibration data, and the number of measurements increases many times for radar sensors that are angularly resolved in two dimensions (in azimuth and in elevation). Since the required calibration measurements are time-consuming and laborious, high costs are incurred for the precise calibration of the sensors. Summary of the Invention
[0005] The object of the present invention is to minimize the number of required calibration measurements given the quality of the calibration data.
[0006] This task is solved according to the invention by a method which is characterized by the following steps:
[0007] a) Select a set of devices from the structural series for creating a training file,
[0008] b) For each selected device: measure the target pattern with a first number of measurement points;
[0009] and create the respective source pattern with a second number of measurement points, which second number is less than the first number; and store the target pattern and the source pattern in the training file,
[0010] c) Use the training file to train a neural network to determine, from the source pattern, the respective target pattern,
[0011] d) Measure the source patterns of the remaining devices from the structural series,
[0012] e) Create the respective target patterns with the aid of the neural network, and
[0013] f) Use these target patterns as calibration data for the remaining devices.
[0014] The above-mentioned "device" can be a radar sensor or other system that can be used to transmit and / or receive electromagnetic radiation. The "target pattern" should be understood as an antenna pattern that reflects the angular dependence of the amplitude and phase of the received signal with the required angular resolution. The "source pattern" should be understood as an antenna pattern based on a reduced number of measurement points compared to the target pattern.
[0015] The invention is based on the recognition that, in the case where sensors (devices) belong to the same structural series, there are characteristic correlations between different parts of the antenna pattern, so that the appearance of other parts of the pattern can be inferred from the measurement results obtained for the measurement points in one part of the pattern without actually measuring these other parts of the pattern. According to the invention, the neural network is trained to recognize these correlations based on suitable training data. If the neural network has been trained, it is sufficient to measure only the source pattern with a relatively small number of measurement points for a single sensor and input this data into the neural network, and then the neural network generates the complete target pattern, which can then be used for the calibration of the sensor.
[0016] In this way, the cost for calibrating the sensors can be significantly reduced.
[0017] Advantageous configurations and further improvements of the invention are given in the dependent claims.
[0018] The source direction pattern can be a direction pattern that covers the same angular range as the respective target direction pattern, but has a reduced angular resolution due to fewer measurement points.
[0019] On the other hand, the source direction pattern can also cover an angular range smaller than that of the target direction pattern. In this case, neural networks are used to extrapolate the calibration data for the larger angular range.
[0020] For a sensor that is angularly resolved in two dimensions, the source direction pattern can represent a one-dimensional cross-section (e.g., in azimuth) through the two-dimensional measurement field, while the target direction pattern determined by the neural network additionally includes calibration data for at least one cross-section (in elevation) in the second dimension.
[0021] For the sensors (devices) selected to create the training file, the complete target direction pattern is measured. Then, without additional measurement, the respective source direction pattern can be simply generated by reducing the target direction pattern to a smaller group of measurement points. Then, during training of the network, the source direction pattern forms the input data, and the respective target direction pattern provides the feedback based on which, for example, the weights of the neural connections in the network are adjusted by means of backpropagation.
[0022] In one embodiment, calibration data is stored for each sensor, which includes both the source direction pattern and the target direction pattern, where, if the source direction pattern is obtained by reducing the measurement points, the source direction pattern is already included in the target direction pattern.
[0023] In another embodiment, the measured source direction pattern is stored in each sensor to be calibrated, but instead of the target direction pattern, a trained neural network is stored in the sensor, which then provides the required calibration data during sensor operation. This embodiment is advantageous when the data volume of the parameters of the trained neural network is smaller than the data volume of the complete target direction pattern.
[0024] The subject matter of the present invention also includes a neural network trained according to the above method for generating calibration data, and also includes a radar sensor in which such a neural network for generating calibration data is stored. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Embodiments are further explained below with reference to the drawings. The drawings show:
[0026] Figure 1 An example of the measurement field of a radar sensor;
[0027] Figure 2 Examples of four amplitude antenna direction patterns for alongFigure 1 Cross-section of the measurement field through the median line II-II;
[0028] Figure 3 and Figure 2 Example of the amplitude antenna pattern of the same four radar sensors as in Figure 1 Cross-section of the measurement field through the median line III-III;
[0029] Figure 4 Measurement field with cross-section lines having different trends;
[0030] Figure 5 Examples of amplitude antenna patterns and phase antenna patterns, which are equivalent to the cross-section of the measurement field along Figure 4 Median line V-V;
[0031] Figure 6 Amplitude antenna pattern and phase antenna pattern, equivalent to the cross-section of the measurement field along Figure 4 Median line VI-VI;
[0032] Figure 7 Diagram for explaining the method according to an embodiment of the present invention;
[0033] Figure 8 Flowchart of a method for training a neural network;
[0034] Figure 9 Flowchart of the first embodiment of the method for providing calibration data according to the present invention;
[0035] Figure 10 Flowchart of the method according to the second embodiment of the present invention; and
[0036] Figure 11 and 12 Diagram similar to Figure 7 for the method according to a modified embodiment of the present invention. Detailed description
[0037] Figure 1 The measurement field 10 of a radar sensor is shown, which is angularly resolved in two dimensions, i.e., in azimuth (horizontal axis in the pattern) and in elevation (vertical axis). The amplitude of the signals received from different directions (modulus of the complex amplitude) is shown by the contour lines 12.
[0038] In the conventional method for providing calibration data for a radar sensor, horizontal cross-sections are placed through the measurement field 10, for example corresponding to Figure 1 the line II in Figure 1the line III therein, and measures the amplitude of the received signal and the phase difference between different antenna elements of the radar sensor at the highest possible angular resolution on each cross-sectional line.
[0039] Figure 2 shows the Figure 1 results of amplitude measurements for four structurally identical radar sensors on line II in the Figure 2 Four curves 14a, 16a, 18a, and 20a visible in the figure give the amplitude as a function of the azimuth angle for the four sensors, respectively. Each curve is part of the antenna pattern of the associated radar sensor. It can be seen that all four curves have a similar trend, but due to manufacturing tolerances, material properties, and similar reasons, there are certain deviations between the curves and from the ideal antenna pattern, which is theoretically expected for sensors belonging to this structural series.
[0040] Figure 3 shows the corresponding curves 14b, 16b, 18b, and 20b measured along Figure 1 line III.
[0041] In Figure 2 and 3 The examples shown illustrate a phenomenon exploited by the invention described herein. This phenomenon lies in the fact that the deviations from sensor to sensor in the antenna pattern Figure 1 portion are highly correlated with the deviations in other portions of the antenna pattern. For example Figure 2 shows that curves 14a and 16a have a very similar trend and deviate from each other only slightly. This slight deviation between the curves is correlated with the Figure 3 corresponding slight deviation between curves 14b and 16b in the Figure 3 portion, i.e., in the portion of the antenna pattern corresponding to a cross-section in the elevation angle direction. In this example, curve 14b has a distinct anomaly particularly in the elevation angle range between 12° and 15°, which is marked with an arrow in the
[0042] This anomaly consists in the fact that the curve has a flatter trend in this range than in the adjacent angle ranges. The same anomaly is also shown in curve 16b.
[0043] Thus, according to the present invention, a neural network is used to identify such regularities or patterns in the antenna patterns of different sensors, and then these regularities are used to predict the shape of the antenna pattern in regions where no measurements have been carried out.
[0044] First, according to Figures 4 to 6 A conventional method for providing calibration data by measuring an antenna pattern will be described. Figure 4 The measurement field 22 of a radar sensor is shown, which has cross-section lines V and VI running through the center of the measurement field, and the antenna pattern should be measured on these cross-section lines. Figure 5 The amplitude antenna pattern and the phase antenna pattern collected along cross-section line V are shown. For this purpose, a large number of measurements were carried out at the measurement points 24 distributed on cross-section line V. Each measurement point 24 represents a measurement in which the reflector whose radar echo is measured is positioned on cross-section line V such that its azimuth angle corresponds to the position of the measurement point on the horizontal axis in Figure 5 In Figure 5 The upper part of Figure 5 gives the amplitude antenna pattern for a single radar sensor, that is, the measured amplitude (in dB) as a function of the azimuth angle (in °). The lower part of Figure 5 gives the phase antenna pattern, that is, the phase difference between adjacent antenna elements as a function of the azimuth angle. This phase difference varies in the range from -π to +π, where the values -π and +π should be considered the same because they both correspond to a phase angle of 180°. Of course, these phase differences are measured at the same measurement points 24 as the amplitude.
[0045] Figure 6 The corresponding amplitude antenna pattern and phase antenna pattern collected on cross-section line VI are shown. Here, amplitude measurements and phase measurements were carried out at a plurality of measurement points 26 evenly distributed on cross-section line VI.
[0046] Even if the determination of the antenna pattern is limited to cross-section lines V and VI, a relatively large number of single measurements (equivalent to the sum of measurement points 24 and 26) have to be carried out for each sensor in this method.
[0047] Figure 7 A method that can significantly reduce the number of single measurements required will be described. In this method, actual measurements are only carried out at a reduced number of measurement points for each sensor to be calibrated, so as to obtain a source pattern 28, which only represents a part of the complete antenna pattern.
[0048] In Figure 7 In the example shown in Figure 4 the source pattern 28 is the amplitude antenna pattern and the phase antenna pattern for cross-section line V in . Therefore, the source pattern 28 only represents the azimuth cross-section through the measurement field, so no measurements need to be carried out on cross-section line VI that runs in the elevation angle. Therefore, the number of single measurements required is only equivalent to the number of measurement points 24 on cross-section line V.
[0049] Then, a synthetic radiation pattern 32 is generated with the aid of a neural network 30 trained specifically for this application scenario, the synthetic radiation pattern giving the amplitude and phase differences on the Figure 4 mid-section line VI, thereby refining the antenna radiation pattern. The source radiation pattern 28 and the synthetic radiation pattern 32 together form a so-called target radiation pattern 34, which provides calibration data for calibrating the radar sensor.
[0050] The neural network 30 has an input stage IN, which feeds the real and imaginary parts of the complex amplitude of each measurement point 24 of the source radiation pattern 28 into the neurons 38 of the first layer 40 of the neural network. Then, in the hidden layer 42, the information of the neurons 38 of the first layer 40 is gradually further processed until finally, in the output layer 44, a predicted value of the real or imaginary part of the complex amplitude is obtained in each neuron, which corresponds to the measurement point 26 on the Figure 4 mid-section line VI. Then, the amplitude values and phase values are output via the output stage OUT of the neural network, and these values together form the synthetic antenna radiation pattern 32.
[0051] The neural network 30 can have any architecture known for neural networks. In the example shown, at least the following layers 40, 42 form a fully connected network, in which each neuron of the first layer 40 affects the state of each neuron in the subsequent layer 42. Alternatively, the neural network 30 can also be a convolutional neural network (Convolutional Neural Network), as is also often used in algorithms for pattern recognition.
[0052] However, before the neural network can fulfill the functions shown in Figure 7 , it must be trained with suitable training data. The main steps of the method for training the network are shown in Figure 8 in the form of a flowchart.
[0053] In step S1, a certain number N of sensors are selected from the series of structures of the radar sensors for which calibration data is to be provided, and the source radiation pattern 28 and the target radiation pattern 34 are measured in a conventional manner for each of these sensors according to the method shown in Figures 4 to 6 . However, the data for the source radiation pattern 28 does not need to be measured separately, since these measurement results are already obtained when measuring the complete target radiation pattern 34. The totality of the source radiation patterns 28 and target radiation patterns 34 for all N sensors obtained in this way in step S1 forms a training file 36, which provides training data for the neural network 30.
[0054] Then, the actual training of the neural network 30 occurs in step S2. Before training, the network 30 is in an initial configuration in which each neural connection between the neurons 38 of one layer and the neurons of the subsequent layer has a defined weight that determines with what strength and in which direction the state of the upstream neuron changes the state of the downstream neuron. Then, in a first training step, the first source orientation map 28 is input into the network via the input stage IN. The result obtained from the output stage OUT is compared with the respective target orientation map 34 from the training file 36. The weights of the neural connections are changed according to the deviation of this result from the target orientation map 34 such that when the same input values are re - input, the result is closer to the desired target orientation map 34. Then, in the next training step, the source orientation map 28 for another sensor is input into the network and the weights are changed again according to the result. If the number N of pairs of source files and target files in the training file 36 is large enough, the weights gradually converge to a configuration in which the network can relatively accurately predict the respective target orientation map 34 for each source orientation map 28. Then, the overall defined transfer function 38 is assigned the weights, which assigns the respective target orientation map to each source orientation map.
[0055] In practice, mostly a validation phase then follows, in which several additional sensors are selected from this series of structures in order to test the performance of the neural network 30. If the test results are positive, then the network can be put into use, and in this way calibration data (i.e., the target file 34) can be generated for each sensor belonging to this series of structures by means of Figure 7 the method shown in
[0056] Figure 9 is a flow chart of a possible method for calibrating a sensor. In this method, in a first step S11, the source orientation map 28 is measured for the sensor to be calibrated. In step S12, this source file is input into the neural network 30 and the target orientation map 34 is generated by means of the neural network or by means of the transfer function 38 defined thereby. Then, in step S13, the target orientation map 34 is stored, which also contains the original source orientation map 28 as a sub - data set. Then, this target orientation map 34 provides the calibration data with which the sensor is calibrated before being put into operation.
[0057] In Figure 10 a further alternative method is shown. In this method, the first step S11 and Figure 9is the same as step S11 therein, i.e., the source direction pattern of the sensor to be calibrated is measured. In the subsequent step 12', the source direction pattern 28 and the neural network 30 are stored in the digital memory of the sensor. Here, "storing the neural network" should be understood as storing the connection pattern of the neurons (connectome (Konnektom)) and the weights of all neural connections.
[0058] After step S12', the sensor can already be put into use. When the sensor is put into operation, the neural network 30 in the electronic device of the sensor can be run in another step S13' to generate the target direction pattern 34 by means of the transfer function 38, and thus generate calibration data. Here, the calibration can be respectively restricted to those parts of the antenna direction pattern for which current positioning data exists.
[0059] Figure 11 In a figure similar to Figure 7 an alternative method for providing calibration data is illustrated. For simplicity, it should be assumed here that the sensor to be calibrated is only angularly resolved in the azimuth angle.
[0060] In this case, the source direction pattern 28' only includes measurement points 24' located within a limited azimuth angle range. The neural network 30 extrapolates the amplitude and phase differences to the entire angle range based on the source direction pattern 28', thereby providing a synthetic direction pattern 32' that extends through the complete detection angle range of the radar sensor and contains the source direction pattern 28' as a sub-dataset. The amplitude trend outside the angle range of the source direction pattern 28 can be represented, for example, by numerical pairs for discrete "measurement points" (or better: support points (Stützpunkte)). Here, the number of support points can be greater than the number of measurement points 24' in the original source direction pattern 28'. In this case, the neural network 30' has a larger number of neurons in the output layer than in the first layer.
[0061] In Figure 12 another method variant is illustrated.
[0062] The source direction pattern 28” has measurement points 24” evenly distributed over the entire azimuth angle range. However, the number of these measurement points is relatively small, so the direction pattern only has a low angular resolution. In this case, the neural network 30” is trained to interpolate intermediate values between the measurement points 24” to obtain a synthetic direction pattern 32” with a higher angular resolution.
[0063] Of course, the methods presented above in conjunction with Figure 7 , 11 and Figure 12 can be combined with each other as needed.
Claims
1. A method for providing calibration data for calibrating the antenna pattern of a device, the device being configured for transmitting and / or receiving electromagnetic radiation and belonging to a common structural series, characterized in that the following steps are performed: a) Selecting a set of devices from the structural series for creating a training file (36), b) For each selected device: Measuring a target pattern (34) having a first number of measurement points (24, 26; 24'; 24”) and creating a respective source pattern (28; 28'; 28”) having a second number of measurement points (24; 24'; 24”), the second number being less than the first number; and storing the target pattern and the source pattern (34, 28; 28'； 28”) in the training file (36), c) Using the training file (36) to train a neural network (30; 30'; 30”) to, Determine the respective target pattern (34) according to the source pattern (28; 28'; 28”), d) Measuring the source pattern (28; 28'; 28”)， e) Creating the respective target pattern (34) with the aid of the neural network (30; 30'; 30”), and f) Using these target patterns (34) as calibration data for the remaining devices.
2. The method according to claim 1, wherein the neural network (30; 30'; 30”) is trained to generate a synthetic pattern (32; 32'; 32”), the synthetic pattern jointly forming the target pattern (34) with the source pattern (28; 28'; 28”).
3. The method according to claim 1 or 2, wherein the source pattern (28') has a smaller value range than the target pattern (34), and the neural network (30; 30'; 30”) is trained to, by extrapolating values from the source pattern (28') to determine the target pattern (34).
4. The method according to any one of the above claims, wherein the neural network (30”) is trained to, when creating the target pattern (34), interpolate between the measurement points (24”) of the source pattern (28”).
5. The method according to any one of the above claims, for a device configured for two-dimensional angle measurement, wherein the source pattern (28) only contains data for one dimension, and the neural network (30) is trained to, from the data of the source pattern (28), generate data for a second dimension.
6. A method for calibrating the antenna pattern of a device, the device being configured for transmitting and / or receiving electromagnetic radiation and belonging to a common structural series, the method using the method for providing calibration data according to any one of the above claims, characterized in that the target pattern (34) is used to calibrate the device before it is put into operation.
7. A method for calibrating the antenna pattern of a device, the device being configured for transmitting and / or receiving electromagnetic radiation and belonging to a common structural series, the method using the method for providing calibration data according to any one of claims 1 to 6, characterized in that, for each device to be calibrated, the measured source pattern (28; 28'; 28”) is stored jointly with the parameters of the neural network (30; 30'; 30”) in the device and is used during operation of the device to provide calibration data and to calibrate the antenna pattern.
8. A neural network, which is trained for performing step e) of the method according to claim 1.
9. A radar sensor for a motor vehicle, having a memory for storing the measured source pattern (28) in step d) of the method according to claim 1, and having an electronic device in which the neural network (30; 30'; 30”) according to claim 9 is implemented.