Efficient DOA estimation method and system for intelligent driving vehicle based on neural network driving

Abstract

The invention discloses an efficient DOA estimation method and system for an intelligent driving vehicle based on neural network driving, and the method comprises the steps: carrying out the sampling of a single-time sampling signal through a preset sparse co-prime array position set, and obtaining a single snapshot signal; inputting the single snapshot signal into a preset neural network, and outputting a covariance matrix; based on a preset sparse co-prime array position set, utilizing differential operation to generate a differential co-array position, utilizing the differential co-array position to carry out vectorization processing on the covariance matrix, and carrying out deduplication and reordering to obtain a signal vector; designing an adaptive filter by using a minimum mean square principle, inputting the signal vector into the filter and carrying out iteration to obtain a filtering signal updating coefficient of a minimum error; and constructing a polynomial by using a filtering signal updating coefficient, solving and searching a root close to a unit circle to obtain a target arrival direction estimation result, and the method is small in calculation amount, high in real-time performance, wide in application range, low in dependency on a high snapshot number and high in accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication and signal processing, and more specifically, to an efficient DOA estimation method and system for an intelligent driving vehicle driven by a neural network. Background Art

[0002] In recent years, with the rapid development and widespread application of AI technology, the automotive industry has ushered in an unprecedented intelligent transformation. Especially in the field of autonomous driving assistance technology, the importance of vehicle-mounted radar systems has become increasingly apparent. When autonomous vehicles are driving at high speeds, these systems must accurately perceive the surrounding environment, including real-time detection and positioning of surrounding vehicles, pedestrians, obstacles and other targets. The vehicle-mounted radar system positioning scene diagram is shown below: Figure 2 As shown. This places extremely high demands on the accurate estimation of the target's direction of arrival (DOA), because the accuracy of DOA is directly related to the navigation, obstacle avoidance and driving safety of autonomous vehicles. However, existing vehicle-mounted radar systems face many challenges in practical applications. First, autonomous vehicles travel in complex natural electromagnetic environments. For example, structures such as high-rise buildings, tunnels and bridges in cities will reflect and scatter radar signals, resulting in multipath propagation problems of signals. These problems will seriously affect the signal quality and target positioning accuracy of the radar system. In addition, sudden natural disasters may also affect the rapid real-time response of the radar system to target signals. It is reported that vehicle-mounted radar systems are often subject to various natural or man-made interferences in practical applications. These interferences not only reduce the navigation accuracy of autonomous driving, but may even cause the operation of the autonomous driving system to be interrupted. Every year, there are endless incidents of vehicle-mounted radar system performance degradation caused by electromagnetic interference, and these incidents show an increasing trend year by year. Although most interferences are short-lived and can be mitigated by conventional signal processing techniques, some interferences that last longer may cause the entire system to fail to operate normally. Therefore, how to quickly and effectively reduce or even shield the impact of various complex interference signals and quickly and accurately locate targets has become an important research topic for improving the performance and safety protection of vehicle-mounted radar systems.

[0003] In the existing vehicle-mounted radar system-assisted autonomous driving navigation technology, the MUSIC algorithm is used for DOA estimation. The MUSIC algorithm constructs the autocovariance matrix by receiving the signal, and uses the eigenvalue decomposition technology to distinguish the signal subspace and the noise subspace, and then realizes DOA estimation by spectral peak search. This method performs well in high signal-to-noise ratio environments, and can effectively improve the resolution and break through the Rayleigh limit. However, the MUSIC algorithm involves a large number of spectral peak search calculations, resulting in a huge amount of calculation. To solve this problem, signal subspace algorithms such as the ESPRIT algorithm have emerged. This type of algorithm can complete DOA estimation without spectral peak search, which significantly improves the estimation speed, but they have specific requirements for array distribution and have a relatively narrow application range. In order to further overcome the limitations of the MUSIC algorithm, researchers proposed a single-shot MUSIC algorithm based on beam space. This method combines conventional beamforming technology and MUSIC super-resolution methods, reduces the amount of calculation and improves real-time performance. However, these improved algorithms still have problems such as dependence on the number of snapshots, low resolution and large estimation bias, which limits their application in complex signal environments.

[0004] The invention patent with the prior art publication number CN116699507A proposes a sparse array complete model error self-correction DOA estimation method based on the atomic norm, which includes the following steps: 1. Receive the signal emitted by the source and obtain the denoised covariance matrix; 2. Establish the inverse matrix constraint of the error matrix; 3. Establish the matrix complete constraint; 4. Construct an atomic norm error self-correction model that meets the sparse array amplitude and phase error correction and data completeness, and convert the self-correction model into an equivalent semi-positive definite programming problem; 5. Perform Vandermonde decomposition on the semi-positive definite Toeplitz matrix in the semi-positive definite programming problem to recover the DOA parameters of the incident source. This solution has large computational complexity and poor real-time performance. Summary of the invention

[0005] In order to overcome the problems of large amount of calculation, poor real-time performance, limited scope of application, over-reliance on the number of snapshots and low accuracy in the prior art DOA estimation methods, the present invention provides an efficient DOA estimation method and system for intelligent driving vehicles driven by a neural network.

[0006] The primary purpose of the present invention is to solve the above technical problems. The technical solution of the present invention is as follows:

[0007] The first aspect of the present invention provides an efficient DOA estimation method for an intelligent driving vehicle based on a neural network drive, comprising the following steps:

[0008] The single-shot signal is sampled using a preset sparse coprime array position set to obtain a single-snapshot signal;

[0009] Inputting the single snapshot signal into a preset neural network and outputting a covariance matrix;

[0010] Based on a preset sparse coprime array position set, a differential common array position is generated by using a differential operation, and the covariance matrix is ​​vectorized and de-duplicated and re-ordered by using the differential common array position to obtain a signal vector;

[0011] The adaptive filter is designed by using the least mean square principle, the signal vector is input into the filter and iterated to obtain the filter signal update coefficient that minimizes the error;

[0012] The polynomial is constructed and solved by using the filter signal update coefficients to find the root close to the unit circle and obtain the target arrival direction estimation result.

[0013] Furthermore, a method for sampling a single-shot sampled signal using a preset sparse coprime array position set to obtain a single-snapshot output signal comprises the following steps:

[0014] Inherit the preset coprime arrays to get the first subarray and the second subarray. The expressions are as follows:

[0015] P1={Nmd|0≤m≤M-1}

[0016] P2={Mnd|0≤n≤N-1}

[0017] Wherein, M and N are the number of sensors in the first subarray and the second subarray respectively, d is the array element spacing, Nm is the position index of a single array element in the first subarray, Mn is the position index of a single array element in the second subarray, Nmd is the position of the mth array element in the first subarray, which is determined by N, m and d, and Mnd is the position of the nth array element in the second subarray, which is determined by M, n and d;

[0018] Construct the third subarray with the following expression:

[0019] P3={-q(M+N)d|1≤q≤Q}

[0020] Where, Q is the number of sensors in the third subarray, and d is the sensor spacing;

[0021] Constructing the fourth subarray, containing an additional sensor, the expression is as follows:

[0022] P4={MN}

[0023] The first subarray, the second subarray, the third subarray and the fourth subarray are combined to form a sparse mutually prime array position set P. ACA , the expression is as follows:

[0024] P ACA =P1∪P2∪P3∪P4

[0025] Using the position set P of sparse coprime arrays ACA The incident signal is sampled once, and the received signal matrix X is obtained as the output single-snap signal, which is expressed as follows:

[0026] X=AS+N

[0027] Among them, X is the T×1-dimensional received signal vector, S is the K×1-dimensional incident signal vector, N is the T×1-dimensional noise vector, and A is the T×K-dimensional direction matrix. The expression is as follows:

[0028] A=[a(θ1),a(θ2),...,a(θ K )]

[0029] Among them, the column vector a(θ i ) is the steering vector of each signal source, θ i is the directional arrival angle of the i-th signal source, and its expression is as follows:

[0030]

[0031] Among them, e represents the base of natural logarithm, j represents the imaginary unit, T represents transpose,

[0032] β i =2πdsinθ i / λ

[0033] Wherein, d is the physical distance between the array sensors, d=λ / 2, and λ is the wavelength of the signal source.

[0034] Furthermore, the preset neural network includes:

[0035] Input layer: used to determine the number of neurons in the input layer according to the dimension of the single snapshot signal vector. Each neuron corresponds to the one-dimensional data of the signal vector.

[0036] Hidden layer: includes one or more fully connected layers. The number of neurons in each layer is adapted to the feature dimension of the input single snapshot signal. It is used to combine the ReLU activation function to enhance the network's nonlinear feature mapping ability for the input signal.

[0037] Output layer: used to generate the predicted values ​​of the covariance matrix.

[0038] Furthermore, the training method of the preset neural network comprises the following steps:

[0039] Obtain a single snapshot signal and its true covariance matrix dataset for training;

[0040] Inputting the single snapshot signal vector data set into a feature extraction module for feature extraction, and outputting a feature data set;

[0041] Input the feature data set into the preset neural network and output the covariance matrix;

[0042] Compare the covariance matrix with the true covariance matrix and calculate the loss using a preset method;

[0043] Adjust network weights and parameters based on loss values ​​and gradients combined with preset algorithms;

[0044] Repeat the above steps until the loss value converges to the set threshold or reaches the maximum number of training iterations, and output the trained neural network model.

[0045] Furthermore, the default method for calculating the loss is to calculate the mean square error. When calculating the mean square error, only the upper or lower triangular part of the matrix is ​​considered to avoid repeated calculations. The expression is as follows:

[0046]

[0047] Among them, R i and R y are the corresponding elements in the true covariance matrix and the predicted covariance matrix, respectively, and T is the dimension of the covariance matrix;

[0048] The default algorithm for adjusting network weights and parameters is the momentum-modified stochastic gradient descent optimization algorithm, and its update rule is shown in the following expression:

[0049]

[0050] θ t =θ t-1 -ηv t

[0051] Where t is the number of iterations, v t-1 、v t are the momentum terms of the t-1th and tth iterations respectively, β is the momentum coefficient, For the parameter θ t-1 The gradient of the loss function J, η is the learning rate, which controls the size of the update step, θ t-1 is the parameter value for the t-1th iteration.

[0052] Furthermore, regularization technology is introduced in the hidden layer to suppress overfitting of the model. The expression is as follows:

[0053]

[0054] Among them, R1(θ) is the regularization term of the neural network weight, θ iis the absolute value of the i-th weight in the hidden layer, θ is the set of weights of the hidden layer, λ is the regularization strength, and n is the number of weights.

[0055] Furthermore, the method of vectorizing the covariance matrix using the differential common array position and performing deduplication and reordering to obtain a virtual signal vector comprises the following steps:

[0056] Flatten the covariance matrix R into a one-dimensional vector Z, as shown below:

[0057]

[0058] Where K is the number of signal sources, p k is the power of the kth signal source, a(θ k ) is the steering vector of the kth signal source, a(θ k ) * is the conjugate transpose of the steering vector, is the Kronecker product, σ 2 is the noise power term, where i is the identity matrix;

[0059] De-duplicate and re-order Z to obtain the signal vector Z1, which is expressed as follows:

[0060]

[0061] Wherein, J is a matrix operator, which is used to deduplicate and reorder the one-dimensional vector Z, and after sorting, generate an equivalent signal vector Z1 corresponding to the number of differential common array positions MN+(2Q+1)(M+N), wherein MN and (2Q+1)(M+N) are calculated according to the positions of the sparse coprime array, M and N are the numbers of sensors in the first subarray and the second subarray, respectively, and Q is the number of sensors in the third subarray.

[0062] Furthermore, the signal vector is input into a filter and iterated to obtain a method for updating coefficients of a filtering signal that minimizes an error, comprising the following steps:

[0063] The positive part of the differential common array position is defined as dca + , the expression is as follows:

[0064] dca + ={0,...,(MN+(2Q+1)(M+N)) / 2}

[0065] Wherein, M and N are the number of sensors in the first subarray and the second subarray respectively, and Q is the number of sensors in the third subarray;

[0066] According to the positive part position dca of the differential common array position +The adaptive filter is used to capture the signal vector and obtain the reference signal vector Z corresponding to the 0 position. r and the auxiliary signal vector Z corresponding to the positive part position a , Z r The expression is as follows:

[0067]

[0068] Where K is the number of signal sources, p k is the power of the kth signal source, σ 2 is the noise power term; Z a The expression is as follows:

[0069]

[0070] Where K is the number of signal sources, a(θ k ) + is the steering vector of the positive part of the differential matrix, p k is the power of the kth signal source;

[0071] The reference array signal Z r With the auxiliary array signal Z a The error signal e(i) is generated by comparison according to the following formula, which is expressed as follows:

[0072] e(i)=Z r (i)-g H (i) Z a

[0073] Among them, i is the number of iterations, g H (i) is the conjugate transpose of the current filter weight;

[0074] According to the error signal e(i) and the auxiliary array signal Z a , combined with the step size parameter u, the adaptive filter weight g is updated according to the following formula, the expression is as follows:

[0075] g(i+1)=g(i)+ue(i) * Z a

[0076] Repeat the above steps, use the updated filter weight g(i+1) to regenerate the error signal e(i+1), and update the weight again until the loss function converges to the set threshold or reaches the maximum number of iterations. The loss function expression is as follows:

[0077]

[0078] When the iteration stops, the final weights are output as the filter signal update coefficients that minimize the error.

[0079] Furthermore, the method for obtaining the target arrival direction estimation result includes the following steps:

[0080] Update the coefficient g using the filtered signal that minimizes the error m Construct the polynomial Q(z), the expression is as follows:

[0081]

[0082] in,

[0083] z=e -j2πdsinθ / λ

[0084] Where, e represents the base of the natural logarithm, j represents the imaginary unit, d is the physical spacing between the array sensors, θ is the signal arrival angle, λ is the wavelength of the signal source, n is the number of positive partial positions of the virtual array, [g m ] (n+1) Indicates the filter signal update coefficient g m The n+1th element of , M and N are the number of sensors in the first subarray and the second subarray respectively, and Q is the number of sensors in the third subarray;

[0085] By solving the polynomial Q(z), we get the root z k , the expression is as follows:

[0086] z k =e -j2πdsinθk / λ

[0087] Where k represents the number of the signal source, using the root z k Extract the target arrival direction estimation result θ k , the expression is as follows:

[0088] θ k =cos -1 (real(z k )) / (2π)

[0089] Among them, cos -1 represents the arccosine function, and real(·) represents the operation of taking the real part of a complex number.

[0090] A second aspect of the present invention provides an efficient DOA estimation system for an intelligent driving vehicle driven by a neural network, comprising a memory and a processor, wherein the memory includes an efficient DOA estimation method program for an intelligent driving vehicle driven by a neural network, and when the efficient DOA estimation method program for an intelligent driving vehicle driven by a neural network is executed by the processor, the steps of an efficient DOA estimation method for an intelligent driving vehicle driven by a neural network are implemented.

[0091] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0092] The present invention adopts a new sparse coprime array (ACA) to replace the traditional uniform array design. While reducing the number of receiving sensors, it significantly expands the number of estimable signal sources, reduces the mutual coupling effect between receiving array elements, and further expands the effective aperture of the virtual array. This design not only improves the positioning accuracy and resolution of DOA estimation, but also effectively reduces the complexity of covariance matrix calculation, greatly improving system performance. At the same time, ACA can maintain efficient signal source estimation performance even when the number of snapshots is limited by increasing the degrees of freedom (DoF) of the array. In addition, the ACA design has stronger adaptability to multipath interference and complex noise environments, broadens the scope of application of the system, and significantly reduces the estimation deviation caused by non-ideal environments.

[0093] The present invention uses a neural network to automatically learn the complex mapping between input data and covariance matrix, which has greater flexibility and adaptability than traditional model-based methods. Through training, the neural network can handle a variety of noise conditions and signal distortions, thereby improving the robustness of the algorithm in non-ideal environments. At the same time, the neural network can automatically extract useful features from the data and capture the nonlinear relationship between the input data and the covariance matrix. These capabilities further improve the DOA estimation performance, especially under complex signal conditions.

[0094] The present invention also introduces an adaptive method based on the least mean square (LMS) principle to accurately locate target signals during driving. This method not only effectively improves the positioning accuracy of the vehicle-mounted radar system for multi-target sources, but also significantly reduces the hardware cost and data processing burden. The LMS algorithm can quickly respond to changes in the signal environment through dynamic weight adjustment, reduce dependence on high snapshot numbers, optimize the positioning accuracy of the signal, and reduce estimation bias. At the same time, the LMS method performs well in adapting to dynamic environments, making the system more flexible and adaptable, and further enhancing the stability and efficiency of the algorithm in actual application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to make the purpose and technical solution of the present invention clearer, the present invention provides the following drawings and descriptions:

[0096] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0097] Figure 2 Localize scene graphs for vehicle-mounted radar systems;

[0098] Figure 3ACA array geometric physical position diagram provided for an embodiment of the present invention;

[0099] Figure 4 A schematic diagram of the neural network model structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0100] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0101] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0102] Embodiment 1:

[0103] The present invention provides an efficient DOA estimation method for intelligent driving vehicles based on neural network drive, such as Figure 1 The figure shows a flow chart of an efficient DOA estimation method for an intelligent vehicle driven by a neural network. The specific steps are as follows:

[0104] Assume that there is a set of single-shot sampling data from a vehicle-mounted radar system, which simulates the incidence of three far-field narrowband signals. The wavelength of the signal is λ, and the incident angles are θ1 = 30°, θ2 = 60°, and θ3 = 90°. These signals are received using a new sparse coprime array (ACA).

[0105] S1: Sample the single-shot signal using a preset sparse coprime array position set to obtain a single-snapshot signal;

[0106] The specific process is:

[0107] Inheriting the preset coprime array, we get the first subarray and the second subarray. By optimizing the arrangement, we can improve the sparsity and effectiveness of the signal. The expression is as follows:

[0108] P1={Nmd|0≤m≤M-1}

[0109] P2={Mnd|0≤n≤N-1}

[0110] Wherein, M and N are the number of sensors in the first subarray and the second subarray respectively, d is the array element spacing, Nm is the position index of a single array element in the first subarray, Mn is the position index of a single array element in the second subarray, Nmd is the position of the mth array element in the first subarray, which is determined by N, m and d, and Mnd is the position of the nth array element in the second subarray, which is determined by M, n and d;

[0111] Construct the third subarray with the following expression:

[0112] P3={-q(M+N)d|1≤q≤Q}

[0113] Wherein, Q is the number of sensors in the third subarray, d is the sensor spacing, and in a specific embodiment, M is 2, N is 3, and Q is 5.

[0114] Constructing the fourth subarray, containing an additional sensor, the expression is as follows:

[0115] P4={MN}

[0116] ACA array geometric physical location diagram Figure 3 As shown, the first subarray, the second subarray, the third subarray and the fourth subarray are combined to form a sparse mutually prime array position set P ACA , the expression is as follows:

[0117] P ACA =P1∪P2∪P3∪P4

[0118] In a specific embodiment, the positions of the four subarrays P1, P2, P3, and P4 on the one-dimensional horizontal axis are [0,3]d, [0,2,4]d, [-5,-10,-15,-20,-25]d, and [6]d, respectively. The position sets of these four subarrays are then merged into a large ACA array position set P ACA It is [-25,-20,-15,-10,-5,0,3,2,4,6]d, where d = λ / 2, λ is the wavelength of the signal source, λ = 0.1 meter.

[0119] The present invention adopts a new sparse coprime array (ACA) to replace the traditional uniform array design. While reducing the number of receiving sensors, it significantly expands the number of estimable signal sources, reduces the mutual coupling effect between receiving array elements, and further expands the effective aperture of the virtual array. This design not only improves the positioning accuracy and resolution of DOA estimation, but also effectively reduces the complexity of covariance matrix calculation, greatly improving system performance. At the same time, ACA can maintain efficient signal source estimation performance even when the number of snapshots is limited by increasing the degrees of freedom (DoF) of the array. In addition, the ACA design has stronger adaptability to multipath interference and complex noise environments, broadens the scope of application of the system, and significantly reduces the estimation deviation caused by non-ideal environments.

[0120] Using the position set P of sparse coprime arrays ACA The incident signal is sampled once to obtain the received signal matrix X as the output single snapshot signal. The matrix X comprehensively reflects the information such as the direction of the signal source, noise interference and signal strength. The expression is as follows:

[0121] X=AS+N

[0122] Among them, X is the T×1-dimensional received signal vector, S is the K×1-dimensional incident signal vector, N is the T×1-dimensional noise vector, and A is the T×K-dimensional direction matrix. The expression is as follows:

[0123] A=[a(θ1),a(θ2),...,a(θ K )]

[0124] Among them, the column vector a(θ i ) is the steering vector of each signal source, θ i is the directional arrival angle of the i-th signal source, and its expression is as follows:

[0125]

[0126] Among them, e represents the base of natural logarithm, j represents the imaginary unit, T represents transpose,

[0127] β i =2πdsinθ i / λ

[0128] Wherein, d is the physical distance between the array sensors, d=λ / 2, λ is the wavelength of the signal source, and in a specific embodiment, λ is 0.1.

[0129] S2: inputting the single snapshot signal into a preset neural network, and outputting a covariance matrix estimated during the learning process of the single snapshot data;

[0130] More specifically, the preset neural network includes:

[0131] Input layer: used to determine the number of neurons in the input layer according to the dimension of the single snapshot signal vector. Assume that the input layer has T neurons. In a specific embodiment, T is 10. Each neuron corresponds to one-dimensional data of the signal vector, corresponding to each element of the received signal vector X.

[0132] Hidden layer: includes one or more fully connected layers (Dense Layers). The number of neurons in each layer is adapted to the feature dimension of the input single snapshot signal. It is used to combine the ReLU activation function to enhance the network's nonlinear feature mapping ability for the input signal.

[0133] More specifically, regularization techniques are introduced in the hidden layer to suppress model overfitting. The expression is as follows:

[0134]

[0135] Among them, R1(θ) is the regularization term of the neural network weight, θ i is the absolute value of the i-th weight in the hidden layer, θ is the set of weights of the hidden layer, λ is the regularization strength, and n is the number of weights.

[0136] Output layer: used to generate the predicted value of the covariance matrix. The number of neurons in the output layer is equal to the number of elements in the covariance matrix, that is, T*T.

[0137] In a specific embodiment, the neural network model structure diagram is as follows: Figure 4 shown.

[0138] More specifically, the training method of the preset neural network includes the following steps:

[0139] Obtain a single snapshot signal and its true covariance matrix dataset for training;

[0140] Input the single snapshot signal vector data set into a feature extraction module to extract features, including the amplitude, phase and frequency components of the signal, and output a feature data set;

[0141] Input the feature data set into the preset neural network and output the covariance matrix;

[0142] Compare the covariance matrix with the true covariance matrix and calculate the loss using a preset method;

[0143] Adjust network weights and parameters based on loss values ​​and gradients combined with preset algorithms;

[0144] Repeat the above steps until the loss value converges to the set threshold or reaches the maximum number of training iterations, and output the trained neural network model.

[0145] More specifically, the default method for calculating the loss is to calculate the mean square error. When calculating the mean square error, only the upper or lower triangular part of the matrix is ​​considered to avoid repeated calculations. The expression is as follows:

[0146]

[0147] Among them, R i and R y are the corresponding elements in the true covariance matrix and the predicted covariance matrix, respectively, and T is the dimension of the covariance matrix;

[0148] The default algorithm for adjusting network weights and parameters is the momentum-modified stochastic gradient descent optimization algorithm, and its update rule is shown in the following expression:

[0149]

[0150] θ t =θ t-1 -ηv t

[0151] Where t is the number of iterations, v t-1 、v t are the momentum terms of the t-1th and tth iterations respectively, β is the momentum coefficient, which usually takes a value between [0.5, 0.9], and determines the proportion of the previous update in the current update. For the parameter θ t-1 The gradient of the loss function J is , η is the learning rate, which controls the size of the update step. In a specific embodiment, η is 0.01, θ t-1 is the parameter value for the t-1th iteration.

[0152] The present invention uses a neural network to automatically learn the complex mapping between input data and covariance matrix, which has greater flexibility and adaptability than traditional model-based methods. Through training, the neural network can handle a variety of noise conditions and signal distortions, thereby improving the robustness of the algorithm in non-ideal environments. At the same time, the neural network can automatically extract useful features from the data and capture the nonlinear relationship between the input data and the covariance matrix. These capabilities further improve the DOA estimation performance, especially under complex signal conditions.

[0153] S3: Based on a preset sparse coprime array position set, a differential common array position is generated by using a differential operation, and the covariance matrix is ​​vectorized and de-duplicated and re-ordered by using the differential common array position to obtain an equivalent signal vector corresponding to the differential common array position;

[0154] The specific process is:

[0155] Flatten the covariance matrix R into a one-dimensional vector Z, as shown below:

[0156]

[0157] Wherein, K is the number of signal sources. In a specific embodiment, K is 3, p k is the power of the kth signal source, a(θ k ) is the steering vector of the kth signal source, a(θ k ) * is the conjugate transpose of the steering vector, is the Kronecker product, σ 2 is the noise power term, where i is the identity matrix;

[0158] De-duplicate and re-order Z to obtain the signal vector Z1, which is expressed as follows:

[0159]

[0160] Wherein, J is a matrix operator, which is used to deduplicate and reorder the one-dimensional vector Z, and after sorting, generate an equivalent signal vector Z1 corresponding to the number of differential common array positions MN+(2Q+1)(M+N), wherein MN and (2Q+1)(M+N) are calculated according to the positions of the sparse coprime array, M and N are the numbers of sensors in the first subarray and the second subarray, respectively, and Q is the number of sensors in the third subarray.

[0161] S4: designing an adaptive filter using the least mean square (LMS) principle, inputting the signal vector into the filter and iterating to obtain a filter signal update coefficient that minimizes the error;

[0162] The specific process is:

[0163] The positive part of the differential common array position is defined as dca + , the expression is as follows:

[0164] dca + ={0,...,(MN+(2Q+1)(M+N)) / 2}

[0165] Wherein, M and N are the number of sensors in the first subarray and the second subarray respectively, and Q is the number of sensors in the third subarray;

[0166] According to the positive part position dca of the differential common array position + The adaptive filter is used to capture the signal vector and obtain the reference signal vector Z corresponding to the 0 position. r and the auxiliary signal vector Z corresponding to the positive part position a , Z r The expression is as follows:

[0167]

[0168] Where K is the number of signal sources, p k is the power of the kth signal source, σ 2 is the noise power term; Z a The expression is as follows:

[0169]

[0170] Where K is the number of signal sources, a(θ k ) + is the steering vector of the positive part of the differential matrix, p k is the power of the kth signal source;

[0171] The reference array signal Z r With the auxiliary array signal Z a The error signal e(i) is generated by comparison according to the following formula, which is expressed as follows:

[0172] e(i)=Z r (i)-g H (i) Z a

[0173] Among them, i is the number of iterations, g H (i) is the conjugate transpose of the current filter weight;

[0174] According to the error signal e(i) and the auxiliary array signal Z a , combined with the step size parameter u, the adaptive filter weight g is updated according to the following formula, the expression is as follows:

[0175] g(i+1)=g(i)+ue(i) * Z a

[0176] Repeat the above steps, use the updated filter weight g(i+1) to regenerate the error signal e(i+1), and update the weight again until the loss function converges to the set threshold or reaches the maximum number of iterations. The loss function expression is as follows:

[0177]

[0178] When the iteration stops, the final weights are output as the filter signal update coefficients that minimize the error.

[0179] The present invention introduces an adaptive method based on the least mean square (LMS) principle to accurately locate target signals during driving. This method not only effectively improves the positioning accuracy of the vehicle-mounted radar system for multi-target sources, but also significantly reduces the hardware cost and data processing burden. The LMS algorithm can quickly respond to changes in the signal environment through dynamic weight adjustment, reduce dependence on high snapshot numbers, optimize the positioning accuracy of the signal, and reduce estimation bias. At the same time, the LMS method performs well in adapting to dynamic environments, making the system more flexible and adaptable, and further enhancing the stability and efficiency of the algorithm in actual application scenarios.

[0180] S5: construct a polynomial using the filter signal update coefficients and solve it, find the root close to the unit circle, and obtain the target arrival direction DOA estimation result.

[0181] The specific process is:

[0182] Update the coefficient g using the filtered signal that minimizes the error m Construct the polynomial Q(z), the expression is as follows:

[0183]

[0184] in,

[0185] z=e -j2πdsinθ / λ

[0186] Where, e represents the base of the natural logarithm, j represents the imaginary unit, d is the physical spacing between the array sensors, θ is the signal arrival angle, λ is the wavelength of the signal source, n is the number of positive partial positions of the virtual array, [g m ] (n+1) Indicates the filter signal update coefficient g m The n+1th element of , M and N are the number of sensors in the first subarray and the second subarray respectively, and Q is the number of sensors in the third subarray;

[0187] By solving the polynomial Q(z), we get the root z k , the expression is as follows:

[0188] z k =e -j2πdsinθk / λ

[0189] Where k represents the number of the signal source, using the root z k Extract the target arrival direction estimation result θ k , the expression is as follows:

[0190] θ k =cos -1 (real(z k )) / (2π)

[0191] Among them, cos -1 represents the arccosine function, and real(·) represents the operation of taking the real part of a complex number.

[0192] Embodiment 2:

[0193] The present embodiment provides an efficient DOA estimation system for an intelligent driving vehicle driven by a neural network, including a memory and a processor. The memory includes an efficient DOA estimation method program for an intelligent driving vehicle driven by a neural network. When the efficient DOA estimation method program for an intelligent driving vehicle driven by a neural network is executed by the processor, the steps of the efficient DOA estimation method for an intelligent driving vehicle driven by a neural network as described in Example 1 are implemented.

[0194] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. An efficient DOA estimation method for intelligent vehicles driven by neural networks, characterized in that: The steps include: The single-shot signal is sampled using a preset sparse coprime array position set to obtain a single-snapshot signal; Inputting the single snapshot signal into a preset neural network and outputting a covariance matrix; Based on a preset sparse coprime array position set, a differential common array position is generated by using a differential operation, and the covariance matrix is ​​vectorized and de-duplicated and re-ordered by using the differential common array position to obtain a signal vector; The adaptive filter is designed by using the least mean square principle, the signal vector is input into the filter and iterated to obtain the filter signal update coefficient that minimizes the error; The polynomial is constructed and solved by using the filter signal update coefficients to find the root close to the unit circle and obtain the target arrival direction estimation result.

2. The efficient DOA estimation method for an intelligent driving vehicle based on a neural network drive according to claim 1 is characterized in that: A method for sampling a single-shot sampled signal using a preset sparse coprime array position set to obtain a single-snapshot output signal comprises the following steps: Inherit the preset coprime arrays to get the first subarray and the second subarray. The expressions are as follows: P1={Nmd|0≤m≤M-1} P2={Mnd|0≤n≤N-1} Wherein, M and N are the number of sensors in the first subarray and the second subarray respectively, d is the array element spacing, Nm is the position index of a single array element in the first subarray, Mn is the position index of a single array element in the second subarray, Nmd is the position of the mth array element in the first subarray, which is determined by N, m and d, and Mnd is the position of the nth array element in the second subarray, which is determined by M, n and d; Construct the third subarray with the following expression: P3={-q(M+N)d|1≤q≤Q} Wherein, Q is the number of sensors in the third subarray, and d is the sensor spacing; Constructing the fourth subarray, containing an additional sensor, the expression is as follows: P4={MN} The first subarray, the second subarray, the third subarray and the fourth subarray are combined to form a sparse mutually prime array position set P. ACA , the expression is as follows: P ACA =P1∪P2∪P3∪P4 Using the position set P of sparse coprime arrays ACA The incident signal is sampled once, and the received signal matrix X is obtained as the output single-snap signal, which is expressed as follows: X=AS+N Among them, X is the T×1-dimensional received signal vector, S is the K×1-dimensional incident signal vector, N is the T×1-dimensional noise vector, and A is the T×K-dimensional direction matrix. The expression is as follows: A=[a(θ1),a(θ2),...,a(θ K )] Among them, the column vector a(θ i ) is the steering vector of each signal source, θ i is the directional arrival angle of the i-th signal source, and its expression is as follows: Among them, e represents the base of natural logarithm, j represents the imaginary unit, T represents transpose, b i =2πdsinθ i / l Wherein, d is the physical distance between the array sensors, d=λ / 2, and λ is the wavelength of the signal source.

3. The efficient DOA estimation method for an intelligent driving vehicle based on a neural network drive according to claim 1 is characterized in that: The preset neural network includes: Input layer: used to determine the number of neurons in the input layer according to the dimension of the single snapshot signal vector. Each neuron corresponds to the one-dimensional data of the signal vector. Hidden layer: includes one or more fully connected layers. The number of neurons in each layer is adapted to the feature dimension of the input single snapshot signal. It is used to combine the ReLU activation function to enhance the network's nonlinear feature mapping ability for the input signal. Output layer: used to generate predicted values ​​of the covariance matrix.

4. The efficient DOA estimation method for an intelligent driving vehicle based on a neural network drive according to claim 3 is characterized in that: The training method of the preset neural network comprises the following steps: Obtain a single snapshot signal and its true covariance matrix dataset for training; Inputting the single snapshot signal vector data set into a feature extraction module for feature extraction, and outputting a feature data set; Input the feature data set into the preset neural network and output the covariance matrix; Compare the covariance matrix with the true covariance matrix and calculate the loss using a preset method; Adjust network weights and parameters based on loss values ​​and gradients combined with preset algorithms; Repeat the above steps until the loss value converges to the set threshold or reaches the maximum number of training iterations, and output the trained neural network model.

5. The efficient DOA estimation method for an intelligent driving vehicle based on neural network drive according to claim 4 is characterized in that: The default method for calculating loss is to calculate the mean square error. When calculating the mean square error, only the upper or lower triangular part of the matrix is ​​considered to avoid repeated calculations. The expression is as follows: Among them, R i and R y are the corresponding elements in the true covariance matrix and the predicted covariance matrix, respectively, and T is the dimension of the covariance matrix; The default algorithm for adjusting network weights and parameters is the momentum-modified stochastic gradient descent optimization algorithm, and its update rule is shown in the following expression: i t =θ t-1 -nv t Where t is the number of iterations, v t-1 、v t are the momentum terms of the t-1th and tth iterations respectively, β is the momentum coefficient, For the parameter θ t-1 The gradient of the loss function J, η is the learning rate, which controls the size of the update step, θ t-1 is the parameter value for the t-1th iteration.

6. The efficient DOA estimation method for an intelligent driving vehicle based on neural network drive according to claim 3 is characterized in that: Regularization technology is introduced in the hidden layer to suppress overfitting of the model. The expression is as follows: Among them, R1(θ) is the regularization term of the neural network weight, θ i is the absolute value of the i-th weight in the hidden layer, θ is the set of weights of the hidden layer, λ is the regularization strength, and n is the number of weights.

7. The efficient DOA estimation method for an intelligent driving vehicle based on a neural network drive according to claim 1 is characterized in that: The method of vectorizing the covariance matrix by using the differential common array position and performing deduplication and reordering to obtain a virtual signal vector comprises the following steps: Flatten the covariance matrix R into a one-dimensional vector Z, as shown below: Where K is the number of signal sources, p k is the power of the kth signal source, a(θ k ) is the steering vector of the kth signal source, a(θ k ) * is the conjugate transpose of the steering vector, is the Kronecker product, σ 2 is the noise power term, where i is the identity matrix; De-duplicate and re-order Z to obtain the signal vector Z1, which is expressed as follows: Wherein, J is a matrix operator, which is used to deduplicate and reorder the one-dimensional vector Z, and after sorting, generate an equivalent signal vector Z1 corresponding to the number of differential common array positions MN+(2Q+1)(M+N), wherein MN and (2Q+1)(M+N) are calculated according to the positions of the sparse coprime array, M and N are the numbers of sensors in the first subarray and the second subarray, respectively, and Q is the number of sensors in the third subarray.

8. The efficient DOA estimation method for an intelligent driving vehicle based on a neural network drive according to claim 1 is characterized in that: The method of inputting the signal vector into a filter and iterating to obtain a filtering signal update coefficient that minimizes the error comprises the following steps: The positive part of the differential common array position is defined as dca + , the expression is as follows: dca + ={0,...,(MN+(2Q+1)(M+N)) / 2} Wherein, M and N are the number of sensors in the first subarray and the second subarray respectively, and Q is the number of sensors in the third subarray; According to the positive part position dca of the differential common array position + The adaptive filter is used to capture the signal vector and obtain the reference signal vector Z corresponding to the 0 position. r and the auxiliary signal vector Z corresponding to the positive part position a , Z r The expression is as follows: Where K is the number of signal sources, p k is the power of the kth signal source, σ 2 is the noise power term; Z a The expression is as follows: Where K is the number of signal sources, a(θ k ) + is the steering vector of the positive part of the differential matrix, p k is the power of the kth signal source; The reference array signal Z r With the auxiliary array signal Z a The error signal e(i) is generated by comparison according to the following formula, which is expressed as follows: e(i)=Z r (i)-g H (i)Z a Among them, i is the number of iterations, g H (i) is the conjugate transpose of the current filter weight; According to the error signal e(i) and the auxiliary array signal Z a , combined with the step size parameter u, the adaptive filter weight g is updated according to the following formula, the expression is as follows: g(i+1)=g(i)+ue(i) * Z a Repeat the above steps, use the updated filter weight g(i+1) to regenerate the error signal e(i+1), and update the weight again until the loss function converges to the set threshold or reaches the maximum number of iterations. The loss function expression is as follows: When the iteration stops, the final weights are output as the filter signal update coefficients that minimize the error.

9. The efficient DOA estimation method for an intelligent driving vehicle based on a neural network drive according to claim 1, characterized in that: The method for obtaining the target arrival direction estimation result comprises the following steps: Update the coefficient g using the filtered signal that minimizes the error m Construct the polynomial Q(z), the expression is as follows: in, with=e -j2πdsinθ / λ Where, e represents the base of the natural logarithm, j represents the imaginary unit, d is the physical spacing between the array sensors, θ is the signal arrival angle, λ is the wavelength of the signal source, n is the number of positive partial positions of the virtual array, [g m ] (n+1) Indicates the filter signal update coefficient g m The n+1th element of , M and N are the number of sensors in the first subarray and the second subarray respectively, and Q is the number of sensors in the third subarray; By solving the polynomial Q(z), we get the root z k , the expression is as follows: Where k represents the number of the signal source, using the root z k Extract the target arrival direction estimation result θ k , the expression is as follows: i k =cos -1 (real(z k )) / (2π) Among them, cos -1 represents the arccosine function, and real(·) represents the operation of taking the real part of a complex number.

10. An efficient DOA estimation system for intelligent vehicles driven by neural networks, characterized in that: The system includes: a memory and a processor, wherein the memory includes a program of an efficient DOA estimation method for an intelligent driving vehicle driven by a neural network, and when the program of the efficient DOA estimation method for an intelligent driving vehicle driven by a neural network is executed by the processor, the steps of the efficient DOA estimation method for an intelligent driving vehicle driven by a neural network as described in any one of claims 1 to 9 are implemented.

Images