Target fast positioning method of FMCW-MIMO radar under non-ideal conditions

Abstract

The application provides a target positioning method of FMCW-MIMO radar under non-ideal conditions, based on deep learning technology, the received signal is solved into a covariance matrix of the signal and a transpose covariance matrix of the signal, the real part, the imaginary part and the phase of the covariance matrix are combined into a three-dimensional tensor, the three-dimensional tensor formed by the covariance matrix is input into a DOA estimation network according to the same method of processing the transpose covariance matrix of the signal, the DOA estimation value is obtained through coarse estimation and fine estimation, the tensor formed by the transpose covariance matrix of the signal is input into a distance estimation network, the distance estimation value is obtained, finally, the DOA value and the distance value are matched in the form of comparing F norm, joint estimation is realized to achieve the positioning effect, in order to eliminate the influence of less array elements, less shots and low signal-to-noise ratio, the application carries out specific training on corresponding data, a low signal-to-noise ratio reinforcement network is added to reduce the error caused by the above conditions, and the fitting capability of deep learning is utilized, so that the positioning of FMCW-MIMO radar under non-ideal conditions is realized.

Description

Technical Field

[0001] This invention relates to the field of radar technology, and in particular to a method for rapid target localization of FMCW-MIMO radar under conditions of fewer array elements, fewer snapshots, and lower signal-to-noise ratio. Background Technology

[0002] Since the birth of radar, target parameter estimation has been a crucial topic in radar system applications, possessing significant research value and importance. In 2004, American scholars Fishler et al. proposed the concept of MIMO radar based on the idea of ​​spatial diversity. Multiple-input multiple-output (MIMO) radar has attracted considerable attention from researchers due to its flexible signal processing methods and higher degrees of freedom in transmission and reception. Continuous wave radar is a type of radar that continuously transmits electromagnetic waves; the transmitted signal is unmodulated, multi-frequency, or modulated. Unmodulated (single-frequency) continuous wave radar calculates the velocity of a moving object by measuring its Doppler shift, enabling velocity measurement of targets with any velocity within a considerable range. Unmodulated continuous wave radar can only measure velocity, not range. Appropriate modulation of the carrier wave is used to simultaneously obtain the target's range and velocity information; that is, the frequency and phase of the appropriately modulated signal are adjusted, and the range and velocity information are obtained through digital signal processing of the echo. Frequency-modulated continuous wave radar can measure the range and velocity information of moving targets. Continuous wave (CW) radar is primarily used for altimetry, proximity fuses, Doppler navigation, target search, tracking and identification, traffic control, battlefield surveillance, and missile guidance. The transmit power of CW radar is relatively evenly distributed over a period of time, with minimal power variation. CW radar has a narrow transmission spectrum, simple circuitry, and low peak power, making it more compatible with simple solid-state transmitters. The signal modulation of CW radar mainly includes several different methods such as triangular wave modulation, sawtooth wave modulation, and sine wave modulation. Since sine wave modulation requires adjusting the signal frequency offset when detecting an object, it is mostly used when there is only one target, such as an altimeter. CW radar needs to detect range and velocity information, and typically aims to detect multiple targets, so it mostly uses triangular wave modulation and sawtooth wave modulation. Sawtooth wave modulation is mainly used to measure the distance of an object, while triangular wave modulation can simultaneously obtain the object's range and velocity information. For CW-MIMO radar, the impact of low signal-to-noise ratio, low snapshot speed, and few array elements on the system is rarely considered. Current algorithms for FMCW-MIMO radar are all based on a large number of array elements and multiple snapshots, and their estimated speed is questionable, which cannot meet the current real-time requirements. Summary of the Invention

[0003] The purpose of this invention is to overcome the limitations of existing FMCW-MIMO radar localization methods, which suffer from low signal-to-noise ratio, low snapshot speed, and slow operation in situations with few array elements. This invention proposes a fast localization method for FMCW-MIMO radar based on deep learning. This method utilizes the nonlinear fitting capabilities of deep learning to achieve high-precision DOA and distance estimation, thereby improving algorithm performance. Simultaneously, leveraging the characteristics of deep learning, this method reduces the time required for algorithm execution.

[0004] To achieve the above-mentioned objective, this invention provides a positioning method for FMCW-MIMO radar under non-ideal conditions, comprising the following steps:

[0005] S1. The receiving array element receives the signal transmitted by the transmitting array element, wherein both the receiving array element and the transmitting array element are half-wavelength uniform linear arrays.

[0006] S2. The received signal is matched and filtered by the matched filter of the receiving array element receiver;

[0007] S3. Combine the real and imaginary parts of the covariance matrix with the phase to form a three-dimensional tensor, and process the transpose covariance matrix of the signal in the same way.

[0008] S4. Based on deep learning technology, we build a coarse DOA estimation network, a low signal-to-noise ratio enhancement network, a fine DOA estimation network, and a distance estimation network, and train the networks.

[0009] S5. Input the 3D tensor into the DOA coarse estimation network, low signal-to-noise ratio enhancement network, DOA fine estimation network and distance estimation network to calculate the target's DOA and distance;

[0010] S6. By comparing the F-norm of the signal matrix formed by different combinations of DOA and distance, the DOA and distance of the smallest F-norm combination are used as the final result to calculate the localization estimate.

[0011] Furthermore, in step S2, the uniform linear array constructed in step S1 is used to receive and model the incident signal. Specifically, assuming there are K narrowband far-field incoherent signal sources, the N snapshot signals received by the linear array with M elements are as follows:

[0012]

[0013] A = [a1, a2, ..., a K ]

[0014]

[0015] S = [s1, s2, ..., s K ]

[0016]

[0017] Where X is the received signal, A is the direction matrix, S is the signal matrix, N is the noise matrix, and a k It is the guide vector, s k It is a signal vector, θ k R is the DOA of the k-th incident signal. k It is the distance of the k-th incident signal, where c represents the speed of light, and f is the distance of the k-th incident signal. c It is the carrier frequency, k s It is the chirping slope, T s This refers to the sampling time. Further, in step S3, the two calculated sampling covariance matrices are R1 and R2, and the corresponding two constructed three-dimensional tensors are T1 and T2, as shown in the following equation:

[0018]

[0019] T1[:,:,1]=Re{R1},T1[:,:,2]=Im{R1},T1[:,:,3]=Angle{R1}

[0020] T2[:,:,1]=Re{R2},T2[:,:,2]=Im{R2},T2[:,:,3]=Angle{R2}

[0021] Further, in step S4, a coarse DOA estimation network, a fine DOA estimation network, a low signal-to-noise ratio enhancement network, and a distance estimation network based on a convolutional neural network are constructed. After training the above networks, the two tensors T1 and T2 obtained in step S3 are input to obtain the DOA and distance. Specifically,

[0022] S41. Construct a coarse DOA estimation network RDN and a low signal-to-noise ratio enhancement network LSN based on a convolutional neural network for use in calling the fine DOA estimation network. RDN and LSN include a classification network built on a convolutional neural network and training data. The training data contains parameter sets for training RDN and LSN for different target numbers K and the partition where the target landing point is located. The entire angle is divided into several sectors. RDN and LSN estimate whether there is a signal in the preset sectors by using the input tensor of the covariance matrix.

[0023] S42. Train the DOA coarse estimation network RDN and the low signal-to-noise ratio reinforcement network LSN to obtain the parameters of the DOA coarse estimation network RDN and the low signal-to-noise ratio reinforcement network LSN.

[0024] S43. Construct a fine-grained DOA estimation network (FDN) based on a convolutional neural network to output the final DOA estimate. The FDN includes a classification network built on a convolutional neural network and training data. Each sector has independent training data. The training data contains the number of different targets K in the sector and the parameter set of the actual DOA value of the target.

[0025] S44. Train the DOA fine estimation network FDN to obtain the parameters of the DOA fine estimation network FDN;

[0026] S45. Construct a distance estimation network REN based on a convolutional neural network to output distance estimates. REN includes a classification network built on a convolutional neural network and training data. The training data includes a parameter set containing the number of different targets K within the complete angular sector and the actual distance values ​​of the targets.

[0027] S46. Train the distance estimation network REN to obtain the parameters of the distance estimation network REN;

[0028] Further, in step S41, a coarse DOA estimation network RDN, a low signal-to-noise ratio enhancement network LSN, and a distance estimation network REN based on a convolutional neural network are constructed, specifically as follows:

[0029] The coarse estimation of DOA is modeled as a multi-label classification task, with the goal of defining the maximum likelihood ML hypothesis as a nonlinear function f from the input space to the output space. RDN Nonlinear functions are composed of multiple layers of networks, i.e.

[0030]

[0031]

[0032] The DOA coarse estimation network RDN architecture includes three 2D convolutional layers, three batch normalization layers, three fully connected layers, and corresponding activation function layers. The aforementioned f... RDN {f in (·) i (·)} i＝1,2,3 Represents a convolutional layer, {f i (·)} i＝4,5,6 The first layer represents a fully connected layer, r(·) and lr(·) represent ReLU and leaky ReLU layers respectively, and bn(·) represents a Batch Normalization layer, dividing the area from -60° to 60° into 10° sectors, for a total of 12 sectors. The final fully connected layer outputs... It is a prediction vector consisting of 12 numbers, each... Both correspond to the probability of whether a signal exists in the corresponding sector. The Low Signal-to-Noise Ratio Enhancement Network (LSN) and the DOA coarse estimation network are structurally identical, and the only difference between them lies in the training data.

[0033] Further, in step S42, the DOA coarse estimation network RDN and the low signal-to-noise ratio enhancement network LSN are trained to obtain the parameters of the DOA coarse estimation network RDN and the low signal-to-noise ratio enhancement network LSN, specifically as follows:

[0034] S421. Traverse the single and dual target angles within the set range. Then, randomly generate target angles from 3 to 8 targets. From the generated target angles, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct tensor T1. Divide the angles into their respective sectors and match them with tensor T1 to generate data pairs. In the form of.

[0035] S422. Use the cross-entropy loss between the actual output and the predicted output as the loss function, specifically,

[0036]

[0037] In the above formula, P represents the probability of the actual output. It is the probability of the predicted output.

[0038] The dataset obtained in step S421 is used as input and expected output to train the DOA coarse estimation network RDN, to obtain the smallest possible L. cross_entropy Network parameters.

[0039] S423. Following the method in step S421, obtain the covariance matrix of the signal-to-noise ratio from -20dB to 0dB and construct a tensor dataset. Use this as input and expected output to train the low signal-to-noise ratio reinforcement network LSN, obtaining the smallest possible L... cross_entropy Network parameters.

[0040] Further, in step S43, a fine estimation network for DOA based on a convolutional neural network, FDN, is constructed, specifically as follows:

[0041] Modeling DOA fine estimation as a multi-label classification task aims to define the maximum likelihood ML hypothesis as a nonlinear function f from the input space to the output space. RDN Nonlinear functions are composed of multiple layers of networks, i.e.

[0042]

[0043]

[0044] The architecture of the DOA fine estimation network FDN includes four 2D convolutional layers, three batch normalization layers, two fully connected layers, and corresponding activation function layers. The aforementioned f... FEN {f in (·) i (·)} i＝1,2,3,4 Represents a convolutional layer, {f i (·)} i＝4,5 Representing a fully connected layer, r(·) and lr(·) represent ReLU and leakyReLU layers respectively, and bn(·) represents a Batch Normalization layer, which constructs an interval of 0.1° for each of the 12 sectors. The output of the final fully connected layer... It is a prediction vector consisting of 100 numbers, each... These are all probabilities of whether a signal exists within the corresponding angular interval.

[0045] Further, in step S44, the DOA fine estimation network FDN is trained to obtain the parameters of the DOA fine estimation network FDN, specifically as follows:

[0046] S441. Traverse the single and dual target angles within the corresponding sector. Then, randomly generate target angles from 3 to 8 targets. Based on the generated target angles, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct tensor T1. Encode the angles into one-hot encoded form and match them with tensor T1 to generate data pairs.

[0047] S442. Using the dataset obtained in step S441 as input and expected output, train the DOA fine estimation network RDN to obtain the smallest possible L. cross_entropy Network parameters. A fine-grained estimation network (FDN) is trained for each sector, for a total of 12 FDNs.

[0048] Further, in step S45, a distance estimation network REN based on a convolutional neural network is constructed, specifically as follows:

[0049] The distance estimation task is modeled as a multi-label classification task, with the goal of defining the maximum likelihood ML hypothesis as a nonlinear function f from the input space to the output space. REN Nonlinear functions are composed of multiple layers of networks, i.e.

[0050]

[0051]

[0052] The architecture of the distance estimation network REN consists of three 2D convolutional layers, three batch normalization layers, two fully connected layers, and corresponding activation function layers. The aforementioned f... REN {f in (·) i (·)} i＝1,2,3 Represents a convolutional layer, {f i (·)} i＝4,5 Representing a fully connected layer, r(·) and lr(·) represent ReLU and leaky ReLU layers respectively, and bn(·) represents a Batch Normalization layer, dividing the range from 0 to 10m into intervals of 0.1m. In the final fully connected layer, the output... It is a prediction vector consisting of 101 numbers, each... These are all probabilities of whether a signal exists within the corresponding distance interval.

[0053] Further, in step S46, the distance estimation network REN is trained to obtain the parameters of the distance estimation network REN, specifically,

[0054] S461. Traverse the single and dual target distances throughout the entire interval. Then, randomly generate target distances within a range of 3 to 8 targets. Based on the generated target distances, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct a tensor T2. Encode the distances into a one-hot encoded form and match it with tensor T2 to generate data pairs.

[0055] S462. Using the dataset obtained in step S461 as input and the expected output, train the distance estimation network REN to obtain the smallest possible L. cross_entropy Network parameters.

[0056] Further, in step S5, the DOA and distance are estimated using two three-dimensional tensors, specifically as follows:

[0057] S51. The three-dimensional tensor T1, composed of the sampled covariance of the received signal, is input into the DOA coarse estimation network RDN to obtain the predicted value. when When the maximum probability is greater than the threshold of 0.5, the Low Signal-to-Noise Ratio Enhancement Network (LSN) is invoked. If the maximum probability is less than the threshold of 0.5, the LSN is invoked, and T1 is input into the LSN to obtain the predicted value.

[0058] S52. Predictions obtained by strengthening the Low Signal-to-Noise Ratio Network (LSN) and the DOA coarse estimation network (RDN) and The form is the same, prediction Sectors with a probability greater than 0.5 call the corresponding DOA fine estimation network FDN.

[0059] S53. Input T1 into the DOA fine estimation network FDN corresponding to the sector called in step S52, and obtain the prediction from each DOA fine estimation network FDN. This corresponds to the DOA (Depth of Ability) location of the sector, and the specific DOA prediction value can be obtained through the spectral peaks.

[0060] S54. The three-dimensional tensor T2, composed of the sampling covariance of the transposed received signal, is input into the distance estimation network REN to obtain the predicted value. Directly obtain the distance estimate of the target.

[0061] Further, in step S6, the DOA and distance estimates obtained in step S5 are combined in different ways to form a received signal using the signal model in step S2. The F-norm of the difference between the formed signal and the received signal is compared, and the DOA and distance estimates of the smallest F-norm combination are paired to achieve positioning estimation.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] The fast localization method for FMCW-MIMO radar under non-ideal conditions provided by this invention solves the problem of fast localization of FMCW-MIMO radar under conditions of low signal-to-noise ratio, few snapshots, and few array elements. This invention utilizes the powerful nonlinear fitting capability of deep learning to effectively improve the localization accuracy. This invention can effectively reduce the computation speed, which is beneficial for real-time target localization and the realization of practical applications.

[0064] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0065] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 This is a flowchart of a rapid localization method for FMCW-MIMO radar under conditions of low signal-to-noise ratio, few snapshots, and few array elements in an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of the FMCW-MIMO radar in an embodiment of the present invention;

[0069] Figure 3 This is a positioning estimation performance diagram in an embodiment of the present invention;

[0070] Figure 4 This is a graph showing the relationship between the root mean square error of the angle estimation of the target by the present invention, the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB) and the change in signal-to-noise ratio.

[0071] Figure 5 This is a graph showing the relationship between the root mean square error of the distance estimation of the target and the signal-to-noise ratio for the present invention, the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB).

[0072] Figure 6 This is a graph showing the relationship between the root mean square error of target localization and the signal-to-noise ratio variation using the present invention, the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB).

[0073] Figure 7 This is a graph showing the relationship between the probability of successful target localization estimation and detection and the change in signal-to-noise ratio using the present invention, the MUSIC algorithm, and the OPM algorithm. Detailed Implementation

[0074] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0075] like Figure 1 The diagram shows a flowchart of a fast localization method for FMCW-MIMO radar under non-ideal conditions according to an embodiment of the present invention. In this embodiment, non-ideal conditions refer to low signal-to-noise ratio, few snapshots, and a limited number of array elements. The fast localization method for FMCW-MIMO radar under low signal-to-noise ratio, few snapshots, and a limited number of array elements provided by the present invention includes the following steps:

[0076] S1. M receiving array elements receive signals transmitted by transmitting array elements, where both receiving and transmitting array elements are half-wavelength uniform linear arrays. Assume there are K narrowband far-field incoherent signal sources with incident angles θ = {θ1, θ2, ..., θ...}. K The noise is incident on M uniform linear arrays (ULA), and it is assumed that the noise is independent Gaussian white noise and is unrelated to the signal.

[0077] S2. The received signal is subjected to matched filtering by the matched filter of the receiving array element receiver. The resulting signal can be expressed as follows:

[0078]

[0079] A = [a1, a2, ..., a K ]

[0080]

[0081] S = [s1, s2, ..., s K ]

[0082]

[0083] Where A is the direction matrix, S is the signal matrix, and θ k R is the DOA of the k-th incident signal. k It is the distance of the k-th incident signal, where c represents the speed of light, and f is the distance of the k-th incident signal. c It is the carrier frequency, k s It is the chirping slope, T s That is the sampling time.

[0084] In step S3, the real and imaginary parts of the covariance matrix and the phase are combined to form a three-dimensional tensor. The transpose covariance matrix of the signal is then processed using the same method. The two calculated sampling covariance matrices are R1 and R2, and the two corresponding three-dimensional tensors are T1 and T2. Specifically,

[0085]

[0086] T1[:,:,1]=Re{R1},T1[:,:,2]=Im{R1},T1[:,:,3]=Angle{R1}

[0087] T2[:,:,1]=Re{R2},T2[:,:,2]=Im{R2},T2[:,:,3]=Angle{R2}

[0088] In step S4, a coarse DOA estimation network, a fine DOA estimation network, a low signal-to-noise ratio enhancement network, and a distance estimation network based on a convolutional neural network are constructed. After training the above networks, the two tensors T1 and T2 obtained in step S3 are input to obtain the DOA and distance. Specifically,

[0089] S41. Construct a coarse DOA estimation network RDN and a low signal-to-noise ratio enhancement network LSN based on convolutional neural networks for use in calling the fine DOA estimation network. RDN and LSN include a classification network built based on convolutional neural networks and training data. The training data contains parameter sets for training RDN and LSN for different target numbers K and the partition where the target landing point is located. The entire angle is divided into several sectors. RDN and LSN estimate whether there is a signal in the preset sectors by using the input tensor of the covariance matrix.

[0090] The coarse estimation of DOA is modeled as a multi-label classification task, with the goal of defining the maximum likelihood ML hypothesis as a nonlinear function f from the input space to the output space. RDN Nonlinear functions are composed of multiple layers of networks, i.e.

[0091]

[0092]

[0093] The architecture of the DOA coarse estimation network RDN includes three 2D convolutional layers, three batch normalization layers, three fully connected layers, and corresponding activation function layers. The aforementioned f... RDN The structure of (·) is: {f i (·)} i＝1,2,3 These represent two-dimensional convolutional layers with 256, 128, and 64 filters and 3 kernels, respectively. Following each two-dimensional convolutional layer is a BatchNormalization layer and a leaky ReLU layer. i (·)} i＝4,5,6 These represent fully connected layers with 2048, 512, and 128 neurons respectively, followed by ReLU layers. The layers are divided into 12 sectors, each 10° apart, ranging from -60° to 60°. These sectors are represented by 12 neurons in the final output layer. It is a prediction vector consisting of 12 numbers, each... Both correspond to the probability of whether a signal exists in the corresponding sector. The Low Signal-to-Noise Ratio Enhancement Network (LSN) and the DOA coarse estimation network are structurally identical; the only difference between them lies in the training data.

[0094] S42. Train the DOA coarse estimation network RDN and the low signal-to-noise ratio enhancement network LSN to obtain the parameters of the DOA coarse estimation network RDN and the low signal-to-noise ratio enhancement network LSN, specifically:

[0095] S421. Traverse the single and dual target angles within the set range. Then, randomly generate target angles from 3 to 8 targets. From the generated target angles, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct tensor T1. Divide the angles into their respective sectors and match them with tensor T1 to generate data pairs. In the form of.

[0096] S422. Use the cross-entropy loss between the actual output and the predicted output as the loss function, specifically,

[0097]

[0098] In the above formula, P represents the probability of the actual output. It is the probability of the predicted output.

[0099] The dataset obtained in step S421 is used as input and expected output to train the DOA coarse estimation network RDN, to obtain the smallest possible L. cross_entropy Network parameters.

[0100] S423. Following the method in step S421, obtain the covariance matrix of the signal-to-noise ratio from -20dB to 0dB and construct a tensor dataset. Use this as input and expected output to train the low signal-to-noise ratio reinforcement network LSN, obtaining the smallest possible L... cross_entropy Network parameters.

[0101] S43. Construct a fine-grained DOA estimation network (FDN) based on a convolutional neural network to output the final DOA estimate. The FDN includes a classification network built on a convolutional neural network and training data. Each sector has independent training data, which contains the number of different targets K in the sector and the parameter set of the actual DOA value of the target. Specifically:

[0102] Modeling DOA fine estimation as a multi-label classification task aims to define the maximum likelihood ML hypothesis as a nonlinear function f from the input space to the output space. RDN Nonlinear functions are composed of multiple layers of networks, i.e.

[0103]

[0104]

[0105] The architecture of the DOA fine estimation network FDN includes four 2D convolutional layers, three batch normalization layers, two fully connected layers, and corresponding activation function layers. The aforementioned f... FEN {f in (·) i (·)} i＝1,2,3,4These represent convolutional layers with 256, 128, 64, and 32 filters respectively, each with a 3-kernel configuration. Following each 2D convolutional layer is a BatchNormalization layer and a leaky ReLU layer. i (·)} i＝4,5 These represent fully connected layers with 512 and 128 neurons respectively, followed by ReLU layers. The angles corresponding to the 12 sectors are divided into intervals of 0.1°. The final output layer has 101 neurons, and the output... It is a prediction vector consisting of 101 numbers, each... These are all probabilities of whether a signal exists within the corresponding angular interval.

[0106] S44. Train the DOA fine estimation network FDN to obtain the parameters of the DOA fine estimation network FDN, specifically:

[0107] S441. Traverse the single and dual target angles within the corresponding sector. Then, randomly generate target angles from 3 to 8 targets. Based on the generated target angles, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct tensor T1. Encode the angles into one-hot encoded form and match them with tensor T1 to generate data pairs.

[0108] S442. Using the dataset obtained in step S441 as input and expected output, train the DOA fine estimation network RDN to obtain the smallest possible L. cross_entropy Network parameters. A fine-grained estimation network (FDN) is trained for each sector, for a total of 12 FDNs.

[0109] S45. Construct a distance estimation network (REN) based on a convolutional neural network to output distance estimates. The REN includes a classification network built on a convolutional neural network and training data. The training data contains a parameter set of the number K different targets within a complete angular sector and the actual distance values ​​of the targets. Specifically...

[0110] The distance estimation task is modeled as a multi-label classification task, with the goal of defining the maximum likelihood ML hypothesis as a nonlinear function f from the input space to the output space. REN Nonlinear functions are composed of multiple layers of networks, i.e.

[0111]

[0112]

[0113] The architecture of the distance estimation network REN consists of three 2D convolutional layers, three batch normalization layers, two fully connected layers, and corresponding activation function layers. The aforementioned f... REN The structure in (·) is: {f i (·)} i＝1,2,3 These represent two-dimensional convolutional layers with 256, 128, and 32 filters and 5, 4, and 3 kernels, respectively. Following the two-dimensional convolutional layers is a BatchNormalization layer and a leaky ReLU layer.

[0114] {f i (·)} i＝4,5 These represent fully connected layers with 2048 and 128 neurons respectively, followed by ReLU layers. Dividing the range from 0 to 10m into 0.1m intervals, the final output layer contains 101 neurons, and the output... It is a prediction vector consisting of 101 numbers, each... These are all probabilities of whether a signal exists within the corresponding distance interval.

[0115] S46. Train the distance estimation network REN to obtain the parameters of the distance estimation network REN. This includes the following steps:

[0116] S461. Traverse the single and dual target distances throughout the entire interval. Then, randomly generate target distances within a range of 3 to 8 targets. Based on the generated target distances, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct a tensor T2. Encode the distances into a one-hot encoded form and match it with tensor T2 to generate data pairs.

[0117] S462. Using the dataset obtained in step S461 as input and the expected output, train the distance estimation network REN to obtain the smallest possible L. cross_entropy Network parameters.

[0118] In step S5, the DOA and distance are estimated using the two three-dimensional tensors constructed in the aforementioned embodiments, specifically including the following steps:

[0119] S51. The three-dimensional tensor T1, composed of the sampled covariance of the received signal, is input into the DOA coarse estimation network RDN to obtain the predicted value. when When the maximum probability is greater than the threshold of 0.5, the Low Signal-to-Noise Ratio Enhancement Network (LSN) is invoked. If the maximum probability is less than the threshold of 0.5, the LSN is invoked, and T1 is input into the LSN to obtain the predicted value.

[0120] S52. Predictions obtained by using Low Signal-to-Noise Ratio Enhancement Network (LSN) and DOA Coarse Estimation Network (RDN) and The form is the same, prediction Sectors with a probability greater than 0.5 call the corresponding DOA fine estimation network FDN.

[0121] S53. Input T1 into the DOA fine estimation network FDN corresponding to the sector called in step S52, and obtain the prediction from each DOA fine estimation network FDN. This corresponds to the DOA (Depth of Ability) location of the sector, and the specific DOA prediction value can be obtained through the spectral peaks.

[0122] S54. The three-dimensional tensor T2, composed of the sampling covariance of the transposed received signal, is input into the distance estimation network REN to obtain the predicted value. Directly obtain the distance estimate of the target.

[0123] In step S6, the DOA and distance estimates obtained in step S5 are combined in different ways to form a received signal according to the signal model constructed in step S2. The F-norm of the difference between the formed signal and the received signal is compared, and the DOA and distance estimates of the combination with the smallest F-norm are paired to achieve positioning estimation.

[0124] The effects of this invention can be illustrated by the following simulations, as follows:

[0125] (I) Simulation conditions and content:

[0126] 1. Target localization performance of FMCW-MIMO radar:

[0127] In the simulation, the FMCW-MIMO radar has M=10 receiving antennas, and the number of targets is set to K=2. The three targets are located at (θ1,r1)=(-14.3°,2.5m) and (θ2,r2)=(17.8°,7.5m) respectively. The signal-to-noise ratio is set to 10dB, and the number of snapshots is set to L=10.

[0128] 2. The root mean square error of the azimuth angle of two targets by the FDA-MIMO radar as a function of the signal-to-noise ratio:

[0129] In the simulation, the FMCW-MIMO radar has M=10 receiving antennas, and the number of targets is set to K=2. Three targets are located at (θ1,r1)=(-14.3°, 2.5m) and (θ2,r2)=(17.8°, 7.5m), respectively. The simulation results are evaluated using the root mean square error (RMSE) for angle estimation performance, defined as:

[0130]

[0131] in It is DOAθ k The estimated result of the i-th Monte Carlo trial. Q is the total number of Monte Carlo trials, which is Q = 2000 in this simulation, and the number of snapshots is set to L = 10. In the following simulation results, this invention is compared with the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB).

[0132] 3. The root mean square error of the distance between two targets by the FMCW-MIMO radar as a function of the signal-to-noise ratio:

[0133] In the simulation, the FMCW-MIMO radar has M=10 receiving antennas, and the number of targets is set to K=2. The two targets are located at (θ1,r1)=(-14.3°,2.5m) and (θ2,r2)=(17.8°,7.5m), respectively. The simulation results are evaluated using the root mean square error (RMSE) for angle estimation performance, defined as:

[0134]

[0135] in It is the distance r k The estimated result of the i-th Monte Carlo trial. Q is the total number of Monte Carlo trials, which is Q = 2000 in this simulation, and the number of snapshots is set to L = 10. In the following simulation results, this invention is compared with the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB).

[0136] 4. The relationship between the root mean square error of positioning by FMCW-MIMO radar and the signal-to-noise ratio:

[0137] In the simulation, the FMCW-MIMO radar has M=10 receiving antennas, and the number of targets is set to K=3. The three targets are located at (θ1,r1)=(-14.3°,2.5m) and (θ2,r2)=(17.8°,7.5m), respectively. The simulation results are evaluated using the root mean square error (RMSE) for angle estimation performance, defined as:

[0138]

[0139] Where, (x k ,y k ) is the position coordinate of the Kth target. This is the localization result of the i-th Monte Carlo trial. Q is the total number of Monte Carlo trials; in this simulation, Q = 2000, and the number of snapshots is set to L = 10. In the following simulation results, this invention is compared with the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB).

[0140] Another parameter used to evaluate the performance of angle estimation is the probability of successful detection (PSD), which is defined as:

[0141] PSD = (D / Q) × 100% (84)

[0142] In this simulation, the parameter settings are the same as in Experiment 2, and 500 Monte Carlo experiments are performed independently. The following simulation results compare the results of this invention, the MUSIC algorithm, and the OPM algorithm.

[0143] (II) Simulation Results

[0144] 1. Target localization performance of FMCW-MIMO radar:

[0145] like Figure 2 and Figure 3 The figures shown are the performance graphs of the angle-distance estimation of the target according to the present invention. We can clearly see that the angle-distance is accurately identified and matched, which proves the effectiveness of the present invention. Therefore, it is applicable to estimating multiple targets simultaneously.

[0146] 2. The root mean square error of the azimuth angle of two targets by the FMCW-MIMO radar as a function of the signal-to-noise ratio:

[0147] Figure 4 This is a graph showing the relationship between the root mean square error of the estimation of the azimuth angle of two targets and the signal-to-noise ratio, using the present invention, the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB). From... Figure 6 It can be seen that the performance of this algorithm is better than the other two algorithms.

[0148] 3. The root mean square error of the distance between two targets by the FMCW-MIMO radar as a function of the signal-to-noise ratio:

[0149] Figure 5 This is a graph showing the relationship between the root mean square error of the distance estimation for three targets and the signal-to-noise ratio (SNR) using the present invention, the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB). From... Figure 5 It can be seen that the performance of this algorithm is better than the other two algorithms.

[0150] 4. The relationship between the root mean square error of FMCW-MIMO radar in locating two targets and the signal-to-noise ratio:

[0151] Figure 6 This is a graph showing the relationship between the root mean square error of the localization estimation of three targets and the signal-to-noise ratio (SNR) using the present invention, the MUSIC algorithm, the OPM algorithm, and the Cramer-Rao bound (CRB). From... Figure 5 It can be seen that the performance of this algorithm is better than the other two algorithms.

[0152] 5. The relationship between the probability of successful target localization and detection by FMCW-MIMO radar and the signal-to-noise ratio:

[0153] Figure 7 This is a graph showing the relationship between the root mean square error of the target azimuth angle estimation and the probability of successful detection, based on the present invention, the MUSIC algorithm, and the OPM algorithm. Figure 6 This represents the PSD of angle estimates from different algorithms. Figure 7 As can be seen from the curves, when the signal-to-noise ratio (SNR) reaches 30dB, the PSD of both algorithms can reach 100%. At the same SNR, the PSD of this algorithm is significantly higher than the other two algorithms under low SNR conditions. In other words, under low SNR conditions, the angle estimation accuracy of this algorithm is higher than the other two algorithms.

[0154] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0155] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0156] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of the invention. Therefore, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A method for rapid target localization using FMCW-MIMO radar under non-ideal conditions, including the following steps: S1, the receiving element receives the signal transmitted by the transmitting element, wherein... Both the receiving and transmitting array elements are half-wavelength uniform linear arrays; S2. The received signal is matched and filtered by the matched filter of the receiving array element receiver; S3. Combine the real and imaginary parts of the covariance matrix with the phase to form a three-dimensional tensor, and process the transpose covariance matrix of the signal in the same way. S4. Based on deep learning technology, we build a coarse DOA estimation network, a low signal-to-noise ratio enhancement network, a fine DOA estimation network, and a distance estimation network, and train the networks. S5. Input the 3D tensor into the DOA coarse estimation network, low signal-to-noise ratio enhancement network, DOA fine estimation network and distance estimation network to calculate the target's DOA and distance; S6. By comparing the F-norm of the signal matrix formed by different combinations of DOA and distance, the DOA and distance of the smallest F-norm combination are used as the final result to calculate the localization estimate.

2. The target rapid localization method for FMCW-MIMO radar under non-ideal conditions according to claim 1, characterized in that: In step S2, the uniform linear array constructed in step S1 is used to receive and model the incident signal. Specifically, assuming there are K narrowband far-field incoherent signal sources, the N snapshot signals received by the linear array with M elements are as follows: Where X is the received signal, A is the direction matrix, S is the signal matrix, and N is the noise matrix. It is a guide vector. It is a signal vector. It is the DOA of the k-th incident signal. It is the distance to the k-th incident signal, and c represents the speed of light. It is the carrier frequency. It is the chirping slope. The sampling time is used. In step S3, the two calculated sampling covariance matrices are respectively... , The two corresponding three-dimensional tensors are respectively , The specific formula is as follows: 。 3. The target rapid localization method for FMCW-MIMO radar under non-ideal conditions according to claim 2, characterized in that: In step S4, a coarse DOA estimation network, a fine DOA estimation network, a low signal-to-noise ratio enhancement network, and a distance estimation network based on a convolutional neural network are constructed. After training the above networks, the two tensors obtained in step S3 are input. , To obtain the DOA and distance, specifically: S41. Construct a coarse DOA estimation network RDN and a low signal-to-noise ratio enhancement network LSN based on a convolutional neural network for use in calling the fine DOA estimation network. RDN and LSN include a classification network built on a convolutional neural network and training data. The training data contains parameter sets for training RDN and LSN for different target numbers K and the partition where the target landing point is located. The entire angle is divided into several sectors. RDN and LSN estimate whether there is a signal in the preset sectors by using the input tensor of the covariance matrix. S42. Train the DOA coarse estimation network RDN and the low signal-to-noise ratio reinforcement network LSN to obtain the parameters of the DOA coarse estimation network RDN and the low signal-to-noise ratio reinforcement network LSN. S43. Construct a fine-grained DOA estimation network (FDN) based on a convolutional neural network to output the final DOA estimate. The FDN includes a classification network built on a convolutional neural network and training data. Each sector has independent training data. The training data contains the number of different targets K in the sector and the parameter set of the actual DOA value of the target. S44. Train the DOA fine estimation network FDN to obtain the parameters of the DOA fine estimation network FDN; S45. Construct a distance estimation network REN based on a convolutional neural network to output distance estimates. REN includes a classification network built on a convolutional neural network and training data. The training data includes a parameter set containing the number of different targets K within the complete angular sector and the actual distance values ​​of the targets. S46. Train the distance estimation network REN to obtain the parameters of the distance estimation network REN.

4. The target rapid localization method for FMCW-MIMO radar under non-ideal conditions according to claim 3, characterized in that: In step S41, a coarse DOA estimation network RDN, a low signal-to-noise ratio enhancement network LSN, and a distance estimation network REN based on a convolutional neural network are constructed, specifically as follows: The coarse estimation of DOA is modeled as a multi-label classification task, with the goal of defining the maximum likelihood ML hypothesis as a nonlinear function from the input space to the output space. Nonlinear functions are composed of multiple layers of networks, i.e. The architecture of the DOA coarse estimation network RDN includes three 2D convolutional layers, three batch normalization layers, three fully connected layers, and corresponding activation function layers. In Represents a convolutional layer. Represents a fully connected layer. , These represent the ReLU layer and the leaky ReLU layer, respectively. This represents the BatchNormalization layer, which divides the area from -60° to 60° into 12 sectors, each segment measuring 10°. The output is then processed in the final fully connected layer. It is a prediction vector consisting of 12 numbers, each... Both correspond to the probability of whether a signal exists in the corresponding sector. The Low Signal-to-Noise Ratio Enhancement Network (LSN) and the DOA coarse estimation network are structurally identical, and the only difference between them lies in the training data.

5. The target rapid localization method for FMCW-MIMO radar under non-ideal conditions according to claim 3, characterized in that: In step S42, the DOA coarse estimation network RDN and the low signal-to-noise ratio enhancement network LSN are trained to obtain their parameters, specifically: S421. Traverse the single and dual target angles within the set range, and then randomly generate target angles from 3 to 8 targets. From the generated target angles, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct a tensor. The angle is divided into its corresponding sector and compared with the tensor. Matching generates data pairs The form; S422. Use the cross-entropy loss between the actual output and the predicted output as the loss function, specifically, In the above formula This represents the probability of the actual output. It is the probability of predicting the output. The dataset obtained in step S421 is used as input and the expected output to train the DOA coarse estimation network RDN, to obtain the smallest possible probability. Network parameters; S423. Following the method in step S421, obtain the covariance matrix of the signal-to-noise ratio from -20dB to 0dB and construct a tensor dataset. Use this as input and expected output to train the low signal-to-noise ratio reinforcement network (LSN) to obtain the smallest possible value. Network parameters.

6. The target rapid localization method for FMCW-MIMO radar under non-ideal conditions according to claim 4, characterized in that: In step S43, a fine estimation network for DOA based on a convolutional neural network, FDN, is constructed, specifically as follows: Modeling DOA fine estimation as a multi-label classification task aims to define the maximum likelihood ML hypothesis as a nonlinear function from the input space to the output space. Nonlinear functions are composed of multiple layers of networks, i.e. The architecture of the DOA fine estimation network FDN includes four 2D convolutional layers, three batch normalization layers, two fully connected layers, and corresponding activation function layers. In Represents a convolutional layer. Represents a fully connected layer. , These represent the ReLU layer and the leaky ReLU layer, respectively. This represents the BatchNormalization layer, which constructs an interval of 0.1° for each of the 12 sectors' corresponding angles. The output in the final fully connected layer... It is a prediction vector consisting of 100 numbers, each... These are all probabilities of whether a signal exists within the corresponding angular interval.

7. The target rapid localization method for FMCW-MIMO radar under non-ideal conditions according to claim 3, characterized in that: In step S44, the DOA fine estimation network FDN is trained to obtain the parameters of the DOA fine estimation network FDN, specifically, S441. Traverse the single and dual target angles within the corresponding sector. Then, randomly generate target angles from 3 to 8 targets. From the target angles generated above, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct a tensor. Encode the angle into a one-hot encoding form and a tensor. Matching generates data pairs ; S442. Using the dataset obtained in step S441 as input and expected output, train the DOA fine estimation network RDN to obtain the smallest possible DOA. The network parameters correspond to training a DOA fine estimation network FDN for each sector, for a total of 12 FDNs.

8. The target rapid localization method for FMCW-MIMO radar under non-ideal conditions according to claim 3, characterized in that: In step S45, the distance estimation network REN based on a convolutional neural network is constructed, specifically as follows: The distance estimation task is modeled as a multi-label classification task, with the goal of defining the maximum likelihood ML hypothesis as a nonlinear function from the input space to the output space. Nonlinear functions are composed of multiple layers of networks, i.e. ； The architecture of the distance estimation network REN includes three 2D convolutional layers, three batch normalization layers, two fully connected layers, and corresponding activation function layers. In Represents a convolutional layer. Represents a fully connected layer. , These represent the ReLU layer and the leaky ReLU layer, respectively. This represents the Batch Normalization layer, which divides the range from 0 to 10m into intervals of 0.1m. The output is then processed in the final fully connected layer. It is a prediction vector consisting of 100 numbers, each... These are all probabilities of whether a signal exists within the corresponding distance interval.

9. The target rapid localization method for FMCW-MIMO radar under non-ideal conditions according to claim 3, characterized in that: In step S46, the distance estimation network REN is trained to obtain its parameters, specifically: S461. Traverse the single and dual target distances throughout the entire interval. Then, randomly generate target distances within a range of 3 to 8 targets. Based on the generated target distances, generate the corresponding covariance matrix with a signal-to-noise ratio of -20dB to 20dB and construct a tensor. The distance is encoded into a one-hot encoding form and a tensor. Matching generates data pairs ; S462. Using the dataset obtained in step S461 as input and the expected output, train the distance estimation network REN to obtain the smallest possible distance. Network parameters.

10. The method for rapid target localization of FMCW-MIMO radar under non-ideal conditions according to any one of claims 2-9, characterized in that: In step S5, the DOA and distance are estimated using two three-dimensional tensors, specifically as follows: S51, The three-dimensional tensor formed by the sampled covariance of the received signal The input is fed into the DOA coarse estimation network RDN to obtain the predicted value. ,when If the maximum probability is greater than or equal to the threshold of 0.5, the Low Signal-to-Noise Ratio Enhancement Network (LSN) is not invoked; if the maximum probability is less than the threshold of 0.5, the LSN is invoked. The input is fed into a low signal-to-noise ratio (SNR) enhancement network (LSN) to obtain the predicted value. ; S52. Predictions obtained by using Low Signal-to-Noise Ratio Enhancement Network (LSN) and DOA Coarse Estimation Network (RDN) and The form is the same, prediction Sectors with a probability greater than 0.5 are assigned the corresponding DOA fine estimation network FDN; S53, will The data is input into the fine-grained DOA estimation network (FDN) corresponding to the sector called in step S52, and the predictions obtained by each FDN are... This corresponds to the DOA (Depth of Ability) of the sector, and the specific DOA prediction value can be obtained through the spectral peaks. S54, The three-dimensional tensor formed by the sampling covariance of the transposed received signal. The input is fed into the distance estimation network REN to obtain the predicted value. Directly obtain the distance estimate of the target; In step S6, the DOA and distance estimates obtained in step S5 are combined in different ways to form a received signal using the signal model in step S2. The F-norm of the difference between the formed signal and the received signal is compared, and the DOA and distance estimates of the smallest F-norm combination are paired to achieve positioning estimation.

Images