Design of a Neural Network-Based Radio Frequency Network in a Vehicle Radar System

Abstract

A method for designing a radar system includes implementing a supervised learning process for a neural network to determine weights corresponding to each of a plurality of patch antennas in the radar system. Each of the plurality of patch antennas has its size determined based on these weights.

Description

Technical Field

[0001] This subject matter discloses the design of a neural network-based radio frequency (RF) network in a vehicle radar system. Background Technology

[0002] Vehicles (such as cars, trucks, construction equipment, agricultural equipment, and automated factory equipment) increasingly include sensors to acquire information about the vehicle and its surrounding environment. This information, for example, aids in the semi-autonomous and autonomous operation of the vehicle. Exemplary sensors include cameras, light detection and ranging (LiDAR) systems, and radio detection and ranging (radar) systems. Radar systems may include a radio frequency (RF) network with multiple patch antennas to receive reflected energy generated by one or more objects within the radar system's field of view reflecting emitted energy. The RF network also includes components for processing the reflected energy prior to additional processing. Therefore, it is desirable to provide a neural network-based RF network design for vehicle radar systems. Summary of the Invention

[0003] In one exemplary embodiment, a method of designing a radar system includes implementing a supervised learning process of a neural network to determine a weight corresponding to each of a plurality of patch antennas of the radar system. Each of the plurality of patch antennas has its size determined based on the weights.

[0004] In addition to one or more features described herein, the supervised learning process also determines feed weights corresponding to each of the plurality of RF feeds used in the output of the plurality of patch antennas combined within the RF network of the radar system. The relative thickness of each of the plurality of RF feeds is determined based on these feed weights.

[0005] In addition to one or more features described herein, the supervised learning process involves obtaining a weighted output from each of the plurality of representative patch antennas representing the radar system, based on selected weights associated with each of the plurality of representative patch antennas. The selected weights for the final iteration of the supervised learning process are the weights corresponding to each of the plurality of patch antennas of the radar system.

[0006] In addition to one or more features described herein, implementing the supervised learning process includes summing weighted outputs from two or more of the plurality of representative patch antennas to obtain sums of two or more of each group of the plurality of representative patch antennas, performing a nonlinear operation on each sum to obtain a nonlinear output, and at the output of each nonlinear operation, weighting each nonlinear output into a weighted nonlinear output based on selected feed weights associated with each representative RF feed, wherein the selected feed weights for the final iteration of the supervised learning process are feed weights corresponding to each of the plurality of RF feeds of the radar system's RF network.

[0007] In addition to one or more features described herein, implementing the supervised learning process includes performing one or more additional combination levels, including performing the summation of weighted nonlinear outputs generated by two or more nonlinear operations, performing the next nonlinear operation on the result of the summation, and providing the weighted result of the next nonlinear operation by selecting additional feed weights, to ultimately provide one or more outputs of the RF network.

[0008] In addition to one or more features described herein, the supervised learning process involves obtaining a baseband signal from one or more downconverters. Each of the one or more downconverters obtains one of one or more outputs of the radio frequency network.

[0009] In addition to one or more features described herein, the supervised learning process includes further processing of baseband signals from each of the one or more downconverters to obtain the location of one or more objects whose reflected energy is a portion of the weighted output from each of the plurality of representative patch antennas.

[0010] In addition to one or more features described herein, as part of the supervised learning process, weighted outputs from each of the plurality of representative patch antennas are simulated, including reflected energy.

[0011] In addition to one or more features described in this article, further processing includes filtering, amplification, and conversion using an analog-to-digital converter.

[0012] In addition to one or more features described in this paper, further processing includes the use of detection neural networks.

[0013] In addition to one or more features described herein, the supervised learning process includes obtaining ground-based data indicating the actual location of the one or more objects.

[0014] In addition to one or more features described in this paper, implementing a supervised learning process involves obtaining a loss by comparing the actual location with a location obtained based on further processing.

[0015] In addition to one or more features described in this paper, the supervised learning process involves updating each selected weight and each selected feed weight in each of multiple iterations, the number of iterations being based on the loss.

[0016] In another exemplary embodiment, the radar system includes a plurality of patch antennas as part of the radar system's radio frequency network. The relative dimensions of each of the plurality of patch antennas are non-uniform, and the learning is based on a supervised learning process implemented via a neural network. The radar system also includes processing circuitry for detecting and locating one or more objects based on reflected signals received by the plurality of patch antennas.

[0017] In addition to one or more features described herein, the radar system also includes multiple radio frequency (RF) feeds as part of an RF network and feeding module to combine outputs from two or more of the multiple patch antennas, thereby ultimately providing one or more RF network outputs. The thickness of each of the multiple RF feeds is non-uniform and based on a supervised learning process implemented via a neural network.

[0018] In addition to one or more features described herein, the radar system also includes one or more downconverters, each of which provides a baseband signal based on one of one or more radio frequency network outputs.

[0019] In addition to one or more features described herein, the radar system also includes one or more digital channels. Each of the one or more digital channels includes a power amplifier, a filter, and an analog-to-digital converter, and corresponds to one of the one or more downconverters to provide a digital signal based on a baseband signal output from one of the one or more downconverters.

[0020] In addition to one or more features described herein, the processing circuitry obtains the position of each of the one or more objects based on digital signals from each of the one or more digital channels.

[0021] In addition to one or more features described herein, the processing circuitry implements a second neural network to detect and locate the one or more objects, and the parameters of the second neural network are updated as part of a supervised learning process implemented by the neural network.

[0022] In addition to one or more features described herein, the radar system is located in the vehicle, and the location of each of the one or more objects provided by the radar system is used to control the operation of the vehicle.

[0023] The above-described features and advantages, as well as other features and advantages, of this disclosure will become apparent when taken in conjunction with the accompanying drawings and the following detailed description. Attached Figure Description

[0024] Other features, advantages, and details appear by way of example only in the following detailed description, which refers to the accompanying drawings, wherein:

[0025] Figure 1 This is a block diagram of a vehicle with a radar system, which includes a neural network-based radio frequency (RF) network design according to one or more embodiments;

[0026] Figure 2 Aspects of an exemplary radar system designed from a neural network based on one or more embodiments are described in detail.

[0027] Figure 3 This is a process flow of performing a method for designing a neural network based on one or more embodiments; and

[0028] Figure 4 It shows Figure 3 The process of the method shown. Detailed Implementation

[0029] The following description is merely exemplary in nature and is not intended to limit this disclosure, its application, or use. It should be understood that in all the drawings, corresponding reference numerals denote similar or corresponding parts and features.

[0030] As previously mentioned, a radar system is one of the exemplary sensors that can be used to acquire vehicle information. A set of patch antennas can be used to receive reflected energy from one or more objects that reflect transmitted energy emitted from the radar system. Increasing the number of patch antennas improves the signal-to-noise ratio and extends the maximum detectable range. Increasing the number of patch antennas also allows for a larger antenna array size, thus increasing the angular resolution of the radar system. However, receiver front-end components that perform processes such as down-converting the received reflected frequency to a baseband frequency and analog-to-digital conversion are costly and therefore limited in number. Each group of these components can be referred to as a receive channel. The output of one or more receive channels is provided to a digital channel for further processing. Each digital channel provides its output to the same processor, which acquires information about the object that produced the reflection.

[0031] Because the number of patch antennas is typically greater than the number of receive channels, energy received by more than one patch antenna can be combined and provided to a single receive channel. In existing methods, a linear combination of signals from a set of patch antennas is provided to each receive channel. However, this linear combination can result in poor angular resolution of the detected object. Embodiments of the systems and methods detailed herein relate to the design of a neural network-based radio frequency (RF) network. An RF network refers to patch antennas and components used to combine energy received from two or more patch antennas before further processing via corresponding digital channels. Neural networks are used in radar system design. Specifically, a neural network is used to determine weights for each patch antenna. This weighting is used to physically determine the size of the patch antennas (i.e., to determine the dimensions of the patch antennas) such that a linear combination of the energy received by each patch antenna will result in a weighted combination. Additionally, weights are determined for the RF feed lines (i.e., RF lines) used to combine the outputs of the patch antennas. The weighting is used to determine the physical thickness of the feed lines in the RF network.

[0032] According to an exemplary embodiment, Figure 1 This is a block diagram of a vehicle 100 with a radar system 110, which includes a neural network-based radio frequency network design. Figure 1 The exemplary vehicle 100 shown is a car 101. Vehicle 100 includes a radar system 110 and may include other sensors 130 (e.g., cameras, lidar systems). Typically, the radar system 110 emits a transmitted signal 105. When the transmitted signal 105 encounters one or more objects 140 (e.g., cars, pedestrians, buildings), the one or more objects 140 reflect some of the energy from the transmitted signal 105 back as reflected energy 115. The number and location of the radar system 110 and other sensors 130 are not limited. Figure 1 The exemplary illustrations in the diagram are limited. Controller 120 obtains information from radar system 110 and one or more other sensors 130 to perform semi-autonomous or autonomous operation of vehicle 100.

[0033] The neural network-based radio frequency network design can be performed by the controller of radar system 110, an external controller, controller 120, or a combination thereof. Each includes processing circuitry, which may include application-specific integrated circuits (ASICs), electronic circuitry, processors (shared, dedicated, or grouped), and memory executing one or more software or firmware programs, combinational logic circuitry, and / or other suitable components providing the aforementioned functionality.

[0034] Figure 2 A number of aspects of an exemplary radar system 110 derived from a neural network-based radio frequency network design according to one or more embodiments are described in detail. Figure 2Multiple patch antennas 210 are shown. These patch antennas 210 are part of the radio frequency network 205 of the radar system 110. The radio frequency network 205 also includes additional components C1, C2, C3 for combining the reflected energy 115 received by two or more patch antennas 210. (Reference) Figure 4 The exemplary mathematical representation 400 shown is discussed further. For clarity, Figure 4 The exemplary mathematical representation shown is not 400. Figure 2 The exemplary radar system 110 shown is illustrated. Figure 2 The exemplary radar system 110 shown and Figure 4 The difference between the exemplary mathematical representations shown (400) is used to illustrate that the number of combination levels can be different. That is, for Figure 2 The exemplary radar system 110 in the example illustrates a three-level combination (i.e., a three-level summing module 217), while for Figure 4 The exemplary mathematical representation 400 in the figure shows a two-level combination (i.e., a two-level summation module 217).

[0035] like Figure 2 As shown, in the exemplary radar system 110, each patch antenna 210 has a different size. Furthermore, the RF feed line 215 (i.e., the RF feed line 215 of the feed-in summing module 217) which facilitates different levels of combination has a non-uniform thickness. (See reference...) Figure 3 The neural network-based process for determining the relative dimensions of each patch antenna 210 and the relative thickness of each RF feed 215 is further discussed. This process uses a mathematical representation 400, which is similar in various respects to... Figure 4 The mathematical representation shown is 400, but it corresponds to... Figure 2 The radar system 110 shown is an example. Figure 2 As shown, the patch antennas 210 of different sizes and the RF feed lines 215 of different thicknesses are the result of weights determined iteratively by a neural network-based RF network design according to one or more embodiments.

[0036] As illustrated in the exemplary case, sixteen patch antennas 210 undergo three levels of combination to produce two outputs from the RF network 205. According to alternative embodiments, the number of patch antennas 210, the number of combination levels, and the final number of outputs can vary. As previously mentioned, the number of digital channels 230 is often a driving factor in designing the number of outputs from the RF network 205 (i.e., the input Si of the downconverter 220). It should be understood that when one or more additional combination levels are part of the RF network 205, one or more additional groups of RF feed lines 215 and summing modules 217 are required. The additional groups of RF feed lines 215 will have a relative thickness reflecting weights determined according to a neural network-based method.

[0037] As previously described, based on the transmitted signal 105 emitted by the radar system 110, each patch antenna 210 receives some of the reflected energy 115. Larger patch antennas 210 will receive more reflected energy 115 than smaller patch antennas. The output Pi of each patch antenna 210 (P0 to P15 in the exemplary case of sixteen patch antennas 210) is combined in three different stages or levels and provided as input Si to the corresponding downconverter 220. Inputs S1 and S2 are as follows... Figure 2 As shown. Based on the different sizes of the patch antenna 210 and the different thicknesses of the RF feed line 215, the combination of reflected energy 115 in each group is actually a weighted combination. See reference... Figure 3 The neural network-based method discussed is used to determine the weights, and the relative weights associated with each patch antenna 210 and each RF feed line 215 are used to determine the size of each patch antenna 210 and the thickness of each RF feed line 215.

[0038] Each downconverter 220 converts the frequency of the corresponding input Si to a lower baseband frequency. This input Si frequency is on the same order of magnitude as the frequency of the transmitted signal 105, which generates reflected energy 115 (e.g., in the gigahertz (GHz) range) received at the patch antenna 210. According to an exemplary embodiment, downconversion requires multiplying the input Si by the transmitted signal 105 or another carrier signal having a frequency on the order of magnitude of the transmitted signal 105. This multiplication results in a difference between the frequency of the input signal Si and the frequency of the transmitted signal 105 or the carrier signal. Because these frequencies are all in the GHz range, the difference (i.e., the result of the multiplication) will be small (i.e., the baseband frequency).

[0039] The baseband signal Bi output by each downconverter 220 is further processed by the digital channel 230. Figure 2 In the exemplary embodiment shown, each of the two digital channels 230 receives a corresponding baseband signal B1, B2 from two downconverters 220. Components as part of each digital channel 230 may include a power amplifier, a filter, and an analog-to-digital converter. The filter is a low-pass filter for eliminating noise in the baseband signal B1. Detection of one or more objects 140 that cause reflected energy 115 to be received by the patch antenna 210 is performed by a processor 240, which receives the output (i.e., digital signal Di) of each digital channel 230. For example, the processor 240 may be part of a radar system 110 or a controller 120. The processor 240 may implement another neural network to perform detection based on the digital signal Di.

[0040] Figure 3This is a process flow of method 300 for performing a neural network-based radio frequency network design according to one or more embodiments. Method 300 includes supervised learning to determine weights applied to the output Pi of each patch antenna 210 and weights applied to each radio frequency feed 215 in the radio frequency network 205. These weights then determine the relative dimensions of the patch antennas 210 and radio frequency feeds 215 in the radio frequency network 205 of the physical radar system 110. Supervised learning can be performed offline before the radar system 110 is installed in the vehicle 100. During implementation... Figure 3 Before the method 300 shown, the number of patch antennas 210 and the number of combination levels (i.e., the number of levels including the summing module 217) are determined.

[0041] At block 310, according to an exemplary embodiment, a weighted output signal Pi*wi can be simulated from each patch antenna 210. The weights wi applied to the output Pi of each patch antenna 210 are refined in iterations, as detailed herein. At block 320, combining the weighted output signals Pi*wi using the weighted RF feed line 215 results in the downconverter 220 providing input Si, as referenced. Figure 4 As detailed below, similar to the weight applied to the output Pi of each patch antenna 210, the weight applied to the output Ri of each RF feed 215 is also improved with iteration according to method 300, as described below.

[0042] Figure 4 The processes at blocks 310 and 320 are shown; therefore, they will be discussed in conjunction with their descriptions. That is, Figure 4 This is a mathematical representation 400 of the process by which a downconverter 220, performed by an exemplary RF network 205, provides input Si. (Continue to refer to...) Figure 1 and Figure 2 However, it must be reiterated that the mathematical representation of 400 is not... Figure 2 The mathematical representation of the exemplary radio frequency network 205 shown includes more than one combination level. According to the exemplary mathematical representation 400, four exemplary subarrays 410-1, 410-2, 410-3, and 410-4 (collectively referred to as 410) are shown, each subarray having four patch antennas 210. The patch antennas 210 provide output signals P0 to P15, respectively weighted by weights w0 to w15. As described above, the neural network-based method may include simulating the weighted output Pi*wi at block 310. At block 320, the process includes combining the weighted output signals Pi*wi (from block 310) using weighted radio frequency feeds 215.

[0043] exist Figure 4 In the exemplary case shown, the weighted output Pi*wi of the four patch antennas 210 of each subarray 410 is first summed at four summing modules 217 to obtain sumi (i.e., at...). Figure 4 The exemplary cases shown are sum1, sum2, sum3, and sum4. For example, sum1 is the weighted sum of the patch antennas 210 of the first subarray 410-1 (P0*w0+P1*w1+P2*w2+P3*w3). Each sumi is provided to the rectified linear unit (ReLU) 420. If the input is positive, the ReLU 420 output Ri is the same as its input, or if the input is negative, the ReLU 420 output Ri is zero. Therefore, the ReLU 420 performs nonlinear processing on sumi to output Ri. Figure 4 The four exemplary ReLU420 outputs shown are R0, R1, R2, and R3. These are implemented as additional components C1 implemented at the first combination level, such as... Figure 2 As shown.

[0044] As shown in the figure, these outputs Ri are weighted by weights w16, w17, w18, and w19 respectively to reflect the relative thickness of the RF feed line 215 in the corresponding physical radar system 110. At the two summing modules 217, the weighted outputs Ri*wi corresponding to each pair of ReLUs 410 are summed to produce sum5 and sum6. These sums, sum5 and sum6, represent the second combination level. Each of the sums sum5 and sum6 is provided to another ReLU 420, as shown in the figure. These two ReLUs 420 represent the additional component C2 implemented at the second combination level. In the exemplary case, there are no further combination levels, and the outputs of the two ReLUs 420 provide the input Si to the downconverter 220. For example, as Figure 4 As shown, (R0*w16+R1*w17) is provided to ReLU 420 to obtain S1. As previously stated, the number of levels of output combination and the number of patch antennas 210 involved in each combination are not limited to the examples here. For example, Figure 2 The exemplary radar system 110 in the diagram will include two patch antennas 210 in each subarray 410 and a corresponding... Figure 2 The additional components C1, C2, and C3 shown are three-stage ReLU 410. (As shown...) Figure 4 As shown, the weighted input is provided to each summing module 217 at each combination level. That is, the weighted output Pi*wi of the patch antenna 210 or the weighted RF feed 215 is input to each summing module 217 at each stage.

[0045] Now back Figure 3 The processing flow shown, at block 330, downconverting the frequency of input Si from RF network 205 (obtained in block 320) may include an analog downconverter 220. In the exemplary case where two inputs S1 and S2 are received via analog mathematical representation 400 (…), Figure 4(As shown), this involves two downconverters 220. As previously described, the frequency of the downconversion input Si may include multiplying each input Si by the transmit signal 105. In an exemplary embodiment of the weighted output Pi*wi of the analog patch antenna 210, the transmit signal 105 is the corresponding analog transmit signal 105 that results in the analog output Pi. As previously described, the downconversion result at block 330 is the baseband signal Bi. In the exemplary case of the two inputs S1 and S2 to the two downconverters 220, two baseband signals B1 and B2 are output.

[0046] At block 340, digital channels 230 are implemented (e.g., simulated). As previously described, each digital channel 230 may include a power amplifier, a filter, and an analog-to-digital converter to process the corresponding baseband signal Bi. Then, at block 350, processor 240 processes the output digital signal Di. That is, at block 350, the process includes detecting one or more objects 140 and estimating the position (e.g., angle of arrival, range) of each detected object 140. This processing may include machine learning (i.e., implementation of another neural network). To reiterate, this design process may include simulating the transmitted signal 105 and the reflected energy 115 to simulate the output Pi of the patch antenna 210. Therefore, objects 140 may also be simulated.

[0047] At block 310, the process of determining the size of each patch antenna 210 of the physical radar system 110 includes determining the correct weights applied to the output Pi of each patch antenna 210 to obtain a weighted output Pi*wi. The relative size of each patch antenna 210 is then designed to match its relative weight. That is, for example, the size of a patch antenna 210 with a lower weight will be smaller than that of a patch antenna 210 with a higher weight. The process of determining the thickness of the RF feed lines 215 of the physical radar system 110 (at each combination level) includes determining the correct weights applied to the output Ri of each ReLU 420 at block 320 to obtain a weighted output Ri*wi. As previously described, at block 320, additional combination levels in the RF network 205 will result in additional weighted output Ri*wi. The relative thickness of each RF feed line 215 is then designed to match its relative weight.

[0048] To determine the correct weights, an iterative process (i.e., supervised learning) is performed at box 360 by implementing a loss function based on ground reality. Ground reality refers to the known locations of (multiple) objects 140. As previously mentioned, (multiple) objects 140 can be simulated. Supervised learning requires that the actual location of each object 140 must be known. Then, at box 360, the ground reality is compared with the estimated location of each detected object 140 obtained at box 350 (with the weights wi assigned at boxes 310 and 320) to obtain the loss. This loss is compared with the loss from the previous iteration to determine if the loss is stable. This will indicate that improvements or refinements based on the weights from additional iterations are negligible. An exemplary embodiment of using a neural network based on the detection and position estimation of each of the objects 140 (at box 350) may include updating the parameters of the neural network (at box 320) in addition to updating the weights applied to the output Pi of the patch antenna 210 (at box 310) and the output Ri of the ReLU 420 (i.e., to the RF feed line 215).

[0049] Specifically, at box 370, it is determined whether the loss is stable. At box 370, any known neural network training method can be used. According to an exemplary embodiment, the iterative process can use a stochastic gradient descent algorithm. According to an alternative embodiment, the loss (from box 360) can be subtracted from the loss obtained in the previous iteration (at box 360) to obtain a difference. This difference can be compared with a predefined (near-zero) value to determine if additional iterations are unnecessary. As shown by the check at box 370, the iteration of updating the weights (Pi*wi) of the output Pi of the patch antenna 210, which began at box 310, and continuing to update the weights (Ri*wi) of the output Ri of the ReLU 420 (also at box 350 according to an exemplary embodiment) is repeated until it is determined that the loss determined at box 360 is sufficiently stable. Then, at box 380, as part of the fabrication of the physical radar system 110, during the last iteration, the dimensions of the patch antenna 210 are determined based on the weights set at boxes 310 and 320, and the RF feed line 215 is formed. The arrangement of the radar system 110 in vehicle 100 can also be part of the process at frame 380.

[0050] While the foregoing disclosure has been described with reference to exemplary embodiments, those skilled in the art will understand that various changes can be made without departing from the scope of the invention, and equivalents can replace its elements. Furthermore, many modifications can be made to adapt particular situations or materials to the teachings of this disclosure without departing from the basic scope of this disclosure. Therefore, it is intended that this disclosure be limited to the specific embodiments disclosed, but will include all embodiments falling within its scope.

Claims

1. A method for designing a radar system, the method comprising: A supervised learning process for a neural network is implemented to determine the weights corresponding to each of the multiple patch antennas in the radar system; and The size of each of the plurality of patch antennas is determined according to weights. in, The supervised learning process also involves determining a feed weight corresponding to each of the plurality of radio frequency (RF) feeds used when combining the outputs of the plurality of patch antennas within the radio frequency network of the radar system, and determining the relative thickness of each of the plurality of RF feeds based on the feed weights. The supervised learning process includes: obtaining a weighted output from each of the plurality of representative patch antennas representing the radar system based on selected weights associated with each of the plurality of representative patch antennas, wherein the selected weights of the last iteration of the supervised learning are the weights corresponding to each of the plurality of patch antennas of the radar system.

2. The method according to claim 1, wherein, Implementing the supervised learning process involves summing the weighted outputs from two or more of the plurality of representative patch antennas to obtain the sum of each group of two or more of the plurality of representative patch antennas; Perform a nonlinear operation on each sum to obtain a nonlinear output; Each nonlinear output is weighted into a weighted nonlinear output based on the selected feed weight associated with each representative RF feed at the output of each nonlinear operation. The feed weight selected in the last iteration of supervised learning is the feed weight corresponding to each of the plurality of RF feeds in the RF network of the radar system.

3. The method according to claim 2, wherein implementing the supervised learning process comprises: Perform one or more additional combination levels, the one or more combination levels including performing the summation of weighted nonlinear outputs generated by two or more nonlinear operations; Perform the next nonlinear operation on the result of the addition; And by selecting additional feed weights to provide a weighted result for the next nonlinear operation, one or more outputs of the RF network are ultimately provided.

4. The method according to claim 3, wherein the supervised learning process comprises: The baseband signal is obtained from one or more downconverters, each of which obtains one of one or more outputs of the radio frequency network; The baseband signal from each of the one or more downconverters is further processed to obtain the location of one or more objects, the reflected energy of which is a portion of the weighted output from each of the plurality of representative patch antennas, wherein the weighted output from each of the plurality of representative patch antennas, including the reflected energy, is simulated as part of a supervised learning process, and the further processing includes filtering, amplification, and conversion using an analog-to-digital converter, and using a detection neural network, and the supervised learning process includes obtaining ground-based data indicating the actual location of the one or more objects.

5. The method of claim 4, wherein implementing the supervised learning process includes obtaining a loss by comparing the actual position with a position obtained based on the further processing, and further includes updating each selected weight and each selected feed weight in each of multiple iterations, the number of iterations being based on the loss.

6. A radar system, comprising: Multiple patch antennas, which are part of the radio frequency (RF) network of the radar system, wherein the relative size of each of the multiple patch antennas is non-uniform and based on a supervised learning process implemented by a neural network; The processing circuitry is configured to detect and locate one or more objects based on reflected signals received by the plurality of patch antennas; and Multiple RF feed lines, which are part of the RF network and configured as feed modules, combine outputs from two or more of the multiple patch antennas to ultimately provide one or more RF network outputs, wherein the thickness of each of the multiple RF feed lines is non-uniform and based on a supervised learning process implemented by the neural network.

7. The radar system of claim 6, further comprising one or more downconverters, each of the one or more downconverters being configured to provide a baseband signal based on one or more digital channels of one or more radio frequency network outputs, wherein each of the one or more digital channels includes a power amplifier, a filter, and an analog-to-digital converter, and corresponds to one of the one or more downconverters to provide a digital signal based on the baseband signal output by one of the one or more downconverters, and the processing circuitry being configured to obtain the position of each of the one or more objects based on the digital signal from each of the one or more digital channels, the processing circuitry being configured to implement a second neural network to detect and locate the one or more objects, and the parameters of the second neural network being updated as part of a supervised learning process implemented by the neural network.

8. The radar system according to claim 6, wherein, The radar system is installed in the vehicle, and the location of each of the one or more objects provided by the radar system is used to control the operation of the vehicle.

Images