Low-power-consumption scene unmanned aerial vehicle DOA estimation method, computer storage medium, electronic equipment and system

Abstract

The invention provides a low-power-consumption scene unmanned aerial vehicle DOA estimation method, a computer storage medium, an electronic device and a system, relates to the technical field of array signal processing and deep learning, and improves the non-Gaussian characteristic of a signal during unmanned aerial vehicle DOA estimation by constructing a fourth-order cumulant matrix of a received signal of a sparse array. According to the method, the hardware cost and complexity are reduced, the DOA estimation precision is improved, pruning and quantization processing are combined, the deep learning model is optimized, the calculation complexity and the storage requirement are reduced, and the method, the computer storage medium, the electronic equipment and the system are more suitable for accurate estimation of the information source direction by the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the technical fields of array signal processing and deep learning, and more specifically, to a method for estimating the direction of arrival (DOA) of an unmanned aerial vehicle (UAV) in a low-power scenario, a computer storage medium, an electronic device, and a system. Background Art

[0002] In recent years, due to advantages such as small size, high flexibility, and low cost, UAVs have been rapidly developed and widely used in fields such as agricultural plant protection, police security, power line inspection, disaster relief, logistics transportation, and environmental monitoring. DOA estimation can accurately estimate the incident direction of UAV signals, thereby improving the positioning accuracy of UAVs, which is crucial for navigation, obstacle avoidance, and path planning of UAVs in the above applications. However, on the one hand, these applications not only require UAVs to have high performance but also require UAVs to minimize energy consumption while completing tasks, meaning that there are limitations in the computing resources and energy consumption of UAVs in embedded devices. On the other hand, UAVs have real-time and efficient requirements for signal processing in low-power application scenarios. Due to the limited computing resources of UAVs, traditional array signal processing algorithms for DOA estimation, such as MUSIC, may not be applicable due to excessive computational complexity.

[0003] Higher-order statistics methods (such as methods based on the fourth-order cumulant matrix) are effective in suppressing Gaussian noise. In addition, deep learning techniques have been introduced into the field of array signal processing, providing new solutions for DOA estimation of unmanned aerial vehicles (UAVs). Data-driven methods based on convolutional neural networks (CNNs) or recurrent neural networks (RNNs) can directly learn the complex mapping relationship between the signal covariance matrix or spatial spectrum and DOA, and significantly improve the robustness under non-Gaussian noise conditions and the DOA estimation accuracy in low signal-to-noise ratio scenarios through large-scale data training. For example, a DOA estimation method based on deep learning is disclosed in the prior art. An array data model is constructed using received array data, an output signal is obtained from the array data model, the fourth-order cumulant of the output signal is constructed, the fourth-order cumulant matrix of the output signal is obtained and vectorized, and then the duplicate processing results are removed. Then, the covariance matrix is calculated by combining the spatial smoothing algorithm. Finally, the calculated covariance matrix is input into a pre-trained CNN model to obtain the DOA estimation result. This solution effectively improves the accuracy of the DOA estimation algorithm, reduces the computational complexity, and lowers the hardware cost and operation burden by combining sparse arrays, spatial smoothing algorithms, and deep neural network models to locate interference signals. However, such models usually rely on high-precision floating-point weights and activation functions, with high computational complexity and storage requirements. Due to the computational resources and energy consumption limitations of UAVs, these deep learning models are difficult to directly deploy. In addition, existing deep learning methods generally ignore the hardware adaptability requirements in resource-constrained edge scenarios such as UAVs and lack optimized designs for low-power application scenarios.

[0004] Currently, for problems of resource-constrained devices such as UAVs, some low-power optimization techniques have been attempted in existing research. For example, model quantization significantly reduces storage and computational complexity by quantizing floating-point weights and activation values into low-bit values (such as binary or ternary), but this approach has been less explored in array signal processing. Redundant neurons or channels are removed through pruning and compression techniques to reduce model complexity, but these methods mainly focus on the field of image processing and have not fully designed effective pruning strategies for the task characteristics of array signal processing. Knowledge distillation transfers knowledge from high-precision models to low-precision models to improve the performance of lightweight models, but its application in complex signal processing scenarios such as DOA estimation is still insufficient. The harsh requirements of UAV scenarios for real-time performance, power consumption, and algorithm robustness make it difficult for existing technologies to directly meet the actual application requirements. Summary of the Invention

[0005] To solve the problem that the current methods for UAV DOA estimation using deep learning and array signal processing cannot meet the requirements in resource-constrained scenarios with low altitude of UAVs, the present invention proposes a method and system for UAV DOA estimation in low-power scenarios. Based on the fourth-order cumulant, the non-Gaussian characteristics of the signal during UAV DOA estimation are improved. And by combining with an optimized deep learning method, the computational complexity and storage requirements are reduced. Moreover, by combining with sparse array design, while reducing the hardware cost and complexity, the DOA estimation accuracy is improved. The method proposed by the present invention is more suitable for accurate estimation of the signal source direction by UAVs.

[0006] To achieve the above technical effects, the technical solution of the present invention is as follows:

[0007] In the first aspect, the present application proposes a method for UAV DOA estimation in low-power scenarios, including:

[0008] S1: Use a sparse array to receive data and construct an array signal model to obtain the array received signal;

[0009] S2: Construct the fourth-order cumulant matrix of the array received signal;

[0010] S3: Convert the fourth-order cumulant matrix into a covariance matrix;

[0011] S4: Prune and quantize the deep learning model;

[0012] S5: Use the covariance matrix to perform quantization-aware training on the deep learning model to obtain a trained deep learning model;

[0013] S6: Use the trained deep learning model for UAV DOA estimation.

[0014] Preferably, a sparse array is designed using an extended Cantor array, and the set ε of the positions of the elements of the designed sparse array is:

[0015]

[0016] where Δ represents the element spacing reference, d m represents the position of the m-th element, M represents the total number of elements, k m represents the element index, represents the set of positive integers.

[0017] According to the above technical means, by designing a sparse array, the virtual aperture is significantly extended using non-uniform distribution, and the resolution and accuracy of DOA estimation can be improved under the condition of a smaller number of physical elements.

[0018] Preferably, assume that there are a total of K UAV signals in the spatial scenario at an angle θ k ={θ 1, θ 2 , ..., θ K}, reaching the sparse array, k = 1, 2, ..., K, the array signal model is: X(t) = A(Θ)S(t) + N(t), t = 1, 2, ..., T, where t represents the t-th sampling moment; among them, represents the array received signal, A(Θ) = [a(θ 1 ), ..., a(θ K )] represents the direction vector matrix, which is an M×K matrix; represents the signal source; represents the additive noise, which is non-Gaussian colored noise; each direction vector is defined as: where d m represents the position of the m-th element in the sparse array.

[0019] Preferably, the fourth-order cumulant matrix of the array received signal is expressed as:

[0020]

[0021] where * represents conjugation, represents the Kronecker product, (·) H represents the matrix transpose, and E{·} represents the expectation calculation.

[0022] Preferably, the process of converting the fourth-order cumulant matrix into a covariance matrix includes the following steps:

[0023] S31: Perform vectorization processing on the fourth-order cumulant matrix to obtain a vector F, and the expression is:

[0024] F = vec(C 4 ) = Vu

[0025]

[0026] where, is the equivalent direction vector matrix composed of virtual elements in the fourth-order cumulant, is the equivalent signal vector, represents the fourth-order cumulant of the signal source;

[0027] Since there are duplicate signals in F, remove the duplicate signals and process them according to the position relationship of the distributed array, and the relationship is shown in the following formula:

[0028]

[0029]

[0030] Remove duplicates and sort F to construct a new vector F′;S32: Reconstruct the vector F′ to form a new vector F 1, the expression is:

[0031] F 1 = (J T J) -1 J T F

[0032] Among them, J represents the screening matrix;

[0033] S33: Divide F 1 into q overlapping sub - matrices Each sub - matrix has q virtual elements, and the covariance matrix of the i - th sub - array is defined as:

[0034] S34: Adopt spatial smoothing technology, divide the equivalent covariance matrix into L sub - matrices, and average the covariance matrices of the sub - matrices:

[0035]

[0036] Among them, R i is the covariance matrix of the i - th sub - matrix, and R smooth represents the covariance matrix after averaging the sub - matrices.

[0037] According to the above - mentioned technical means, the fourth - order cumulant matrix can extract the features of non - Gaussian signals and effectively suppress the influence of Gaussian noise, thereby significantly improving the robustness of DOA estimation in a complex noise environment. Using spatial smoothing technology to block sub - matrices and averaging their covariance matrices can effectively solve the correlation problem of array signals, enhance the positive definiteness of the matrix, and eliminate the influence of array position offset.

[0038] Preferably, the deep - learning model adopts a multi - layer fully - connected network. The multi - layer fully - connected network consists of an input layer, a hidden layer, and an output layer. The input layer receives input data and passes it to the hidden layer. The hidden layer consists of multiple neurons. Each neuron receives the outputs of all neurons from the previous layer, and after processing through weighted summation and an activation function, generates its own output and passes it to the next layer. The output layer is responsible for generating the final network output; the deep - learning model uses the ReL u function as the activation function, and the activation - function expression is: f(x) = max(0, x);

[0039] The pruning process for the deep - learning model is structured pruning, including the following process:

[0040] Determine the transfer synaptic channels and neurons to be pruned. The transfer synaptic channels are located between the input layer, the hidden layer, and the output layer;

[0041] Select a pruning criterion, perform pruning operations, identify and remove redundant synaptic channels and neurons that satisfy the expression: W′ = {W j，：，：，： , ||W j，：，：，： ||F > T}, where W j，：，：，： represents the j-th output channel, and ||·|| is the Frobenius norm.

[0042] The quantization process for the deep learning model satisfies the expression:

[0043]

[0044] where x represents the weight or activation value, s is the quantization scale, and Q max is the quantization range.

[0045] According to the above technical means, unimportant parameters (such as weights or neurons) in the deep learning model are removed through pruning to reduce computational and storage requirements. Structured pruning is more friendly in hardware implementation, can effectively reduce the matrix dimension, and optimize the inference speed. Additionally, by means of quantization processing, the weights and activation functions of the deep learning model are compressed into low-bit values (such as binary or ternary), which can significantly reduce the model's storage requirements and computational complexity.

[0046] Preferably, the covariance matrix R i after averaging the sub-matrices is represented in the matrix form of amplitude and phase, expressed as:

[0047]

[0048] where A represents the amplitude factor of the covariance matrix R i , represents the phase factor. Since R i has conjugate symmetry, only the upper triangular phase factor is retained. Design a feature matrix R i ′, whose upper triangle is the imaginary part of the phase factor, the lower triangle is the real part of the phase factor, and the diagonal elements are the normalized values of the diagonal elements of R i so that when more effective information is input into the deep learning model training or prediction, as follows:

[0049]

[0050] Vectorize it to obtain the eigenvector r AP , which has q×q elements, as follows:

[0051] r Ap = vec(R i )

[0052] Utilize the transformed covariance amplitude eigenvector rAP As input features, they are input into the deep learning model; the ReLU function is selected as the activation function for each layer, and the expression is:

[0053]

[0054] During the process of using the covariance matrix to perform quantization-aware training on the deep learning model, quantization error simulation is introduced, and the training objective is to minimize the loss of the deep learning model after quantization. The expression is:

[0055]

[0056] The training objective of the pruned model is to minimize the loss of the impact of pruning on the model performance. The expression is:

[0057]

[0058] Combining the minimization of the loss of the deep learning model after quantization and the minimization of the loss of the impact of pruning on the model performance, the final training objective is:

[0059]

[0060] When the training is completed, a trained deep learning model is obtained.

[0061] According to the above technical means, by using quantization-aware training to simulate the impact of inference quantization during the training process and compensating for the quantization error, the precision loss can be effectively reduced.

[0062] Preferably, before using the trained deep learning model for UAV DOA estimation, it further includes: performing hardware deployment on the trained deep learning model by using the methods of sparse matrix acceleration and pipeline optimization.

[0063] According to the above technical means, coefficient matrix acceleration can significantly reduce the number of operations and memory access requirements, reduce the computational complexity and energy consumption, and achieve efficient utilization of hardware resources.

[0064] In a second aspect, the present application proposes a computer storage medium storing an instruction program for executing the steps in the method for UAV DOA estimation in a low-power scenario.

[0065] In a third aspect, the present application proposes an electronic device, which includes: a memory and a processor. The memory stores the program of the method for UAV DOA estimation in a low-power scenario, and when the program is executed by the processor, the steps of the method for UAV DOA estimation in a low-power scenario are implemented.

[0066] In a fourth aspect, the present application proposes a UAV DOA estimation system in a low-power scenario, including:

[0067] A received signal calculation unit, configured to use a sparse array to receive data to construct an array signal model and obtain an array received signal;

[0068] A fourth-order cumulant matrix construction unit, configured to construct a fourth-order cumulant matrix of the array received signal;

[0069] A covariance matrix conversion unit, configured to convert the fourth-order cumulant matrix into a covariance matrix;

[0070] A quantization pruning unit, configured to perform pruning and quantization processing on a deep learning model;

[0071] A deep learning model training unit, configured to perform quantization-aware training on the deep learning model using the covariance matrix to obtain a trained deep learning model;

[0072] An estimation unit, configured to perform UAV DOA estimation using the trained deep learning model.

[0073] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0074] The present invention provides a UAV DOA estimation method, a computer storage medium, an electronic device and a system in a low-power scenario. By constructing a fourth-order cumulant matrix of the received signal of a sparse array, the non-Gaussian characteristics of the signal during UAV DOA estimation are improved, the hardware cost and complexity are reduced, while the DOA estimation accuracy is improved. And by combining pruning and quantization processing, the deep learning model is optimized, the computational complexity and storage requirements are reduced, and it is more suitable for accurate estimation of the source direction by UAVs. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It represents a schematic flowchart of the UAV DOA estimation method in the low-power scenario proposed in the embodiment of the present invention;

[0076] Figure 2 It represents a schematic diagram of the application of the UAV in the low-power scenario proposed in the embodiment of the present invention;

[0077] Figure 3 It represents a schematic diagram of the composition of the deep learning model proposed in the embodiment of the present invention;

[0078] Figure 4 It represents a schematic diagram of the pruning process of the deep learning model proposed in the embodiment of the present invention;

[0079] Figure 5 It represents a composition diagram of the UAV DOA estimation system in the low-power scenario proposed in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0080] The accompanying drawings are only for illustrative purposes and should not be construed as limiting the patent;

[0081] To better illustrate this embodiment, some parts of the accompanying drawings are omitted, enlarged or reduced, and do not represent the actual size;

[0082] For those skilled in the art, it is understandable that some well-known content descriptions in the accompanying drawings may be omitted.

[0083] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0084] The description of the positional relationship in the accompanying drawings is only for illustrative purposes and should not be construed as limiting the patent;

[0085] Embodiment 1

[0086] This embodiment proposes a method for estimating the DOA of an unmanned aerial vehicle (UAV) in a low-power scenario. The flowchart of this method is shown in Figure 1 , as Figure 1 shown, and includes the following steps:

[0087] S1: Use a sparse array to receive data to construct an array signal model to obtain the array received signal;

[0088] S2: Construct a fourth-order cumulant matrix of the array received signal;

[0089] S3: Convert the fourth-order cumulant matrix into a covariance matrix;

[0090] S4: Prune and quantize the deep learning model;

[0091] S5: Use the covariance matrix to perform quantization-aware training on the deep learning model to obtain a trained deep learning model;

[0092] S6: Use the trained deep learning model for UAV DOA estimation.

[0093] In this embodiment, by constructing a fourth-order cumulant matrix of the received signal of the sparse array, the non-Gaussian characteristics of the signal during UAV DOA estimation are improved, the hardware cost and complexity are reduced, and the DOA estimation accuracy is improved. Moreover, combined with pruning and quantization processing, the deep learning model is optimized, the computational complexity and storage requirements are reduced, and it is more suitable for accurate estimation of the source direction by the UAV.

[0094] Embodiment 2

[0095] Since the traditional Uniform Linear Array (ULA) faces the problem of insufficient resolution when the number of signal sources is close to or exceeds the number of array elements, while the sparse array significantly expands the virtual array aperture through non-uniform distribution, thus enabling the improvement of the resolution and accuracy of DOA estimation under the condition of a smaller number of physical array elements. In this embodiment, a sparse array is designed using the extended Cantor array, and the set ε of the positions of the array elements in the designed sparse array is:

[0096]

[0097] where Δ represents the reference element spacing, d m represents the position of the m-th element, M represents the total number of elements, k m represents the element index, represents the set of positive integers.

[0098] In this embodiment, the process of constructing an array signal model using the sparse array to receive data is as follows:

[0099] Figure 2 represents the schematic diagram of the application of the low-power scenario unmanned aerial vehicle proposed in the embodiment of the present invention, Figure 2 where 1 represents the unmanned aerial vehicle. Assume that there are a total of K unmanned aerial vehicle signals in the space scenario arriving at the sparse array at angles θ k ={θ 1 , θ 2 ,..., θ K}, k = 1, 2,..., K, and the array signal model is: X(t) = A(Θ)S(t) + N(t), t = 1, 2,..., T, where t represents the t-th sampling moment; among them, represents the array received signal, A(Θ) = [a(θ 1 ),..., a(θ K )] represents the direction vector matrix, which is an M×K matrix; represents the signal source; represents the additive noise, which is non-Gaussian colored noise; each direction vector is defined as: where d m represents the position of the m-th element in the sparse array.

[0100] The traditional covariance matrix performs well under Gaussian white noise. However, in practical application scenarios, the noise is usually non-Gaussian or colored noise, resulting in a significant decline in the performance of traditional methods. The high-order cumulant matrix can extract the features of non-Gaussian signals and effectively suppress the influence of Gaussian noise, thus significantly improving the robustness of DOA estimation in complex noise environments. In this embodiment, a fourth-order cumulant matrix of the array received signal is constructed, and the fourth-order cumulant matrix is selected as the core means for feature extraction. Aiming at the noise problem in the array signal model, the fourth-order cumulant matrix is calculated to suppress Gaussian noise and extract the characteristics of non-Gaussian signals. The fourth-order cumulant vector matrix is expressed as:

[0101]

[0102] where * represents conjugation, denotes the Kronecker product, (·) H represents the matrix transpose, and E{·} represents the expectation calculation. In this embodiment, by introducing high-order cumulants, an extended virtual array can be generated, and its direction matrix is defined as: In a sparse array, the correlation between signals may lead to the degradation problem of the covariance matrix. The spatial smoothing algorithm can effectively solve the correlation problem of array signals and enhance the positive definiteness of the matrix by dividing the sub-arrays and averaging their covariance matrices. Therefore, in the present invention, to eliminate the influence of array position offset, the spatial smoothing technique is adopted. The process of converting the fourth-order cumulant matrix into a covariance matrix includes the following steps: S31: Vectorize the fourth-order cumulant matrix to obtain a vector F, and the expression is:

[0103] F = vec(C 4 ) = Vu

[0104]

[0105] where is an equivalent direction vector matrix composed of virtual elements in the fourth-order cumulant, is an equivalent signal vector, represents the fourth-order cumulant of the signal source;

[0106] Since there are duplicate signals in F, remove the duplicate signals and process them according to the position relationship of the distributed array. The relationship is shown in the following formula:

[0107]

[0108] Remove the duplicates and sort F to construct a new vector F';

[0109] S32: Reconstruct the vector F' to form a new vector F 1 , and the expression is:

[0110] F 1 = (J T J) -1 J T F

[0111] Among them, J represents a screening matrix;

[0112] S33: Divide F 1 into q overlapping submatrices Each submatrix has q virtual elements. Define the covariance matrix of the i-th subarray as:

[0113] S34: Adopt spatial smoothing technology, divide the equivalent covariance matrix into L submatrices, and average the covariance matrices of the submatrices:

[0114]

[0115] Among them, R i is the covariance matrix of the i-th submatrix, and R smooth represents the covariance matrix after averaging the submatrices.

[0116] In this embodiment, as Figure 3 shown, the deep learning model adopts a multi-layer fully connected network. The multi-layer fully connected network consists of an input layer, a hidden layer, and an output layer. Among them, the input layer receives input data and passes it to the hidden layer. The hidden layer is composed of multiple neurons. Each neuron receives the outputs of all neurons from the previous layer, and after processing through weighted summation and an activation function, generates its own output and passes it to the next layer. The output layer is responsible for generating the final network output; In this embodiment, the deep learning model uses the ReLu function as the activation function, and the activation function expression is: f(x) = max(0, x).

[0117] Pruning is to remove unimportant parameters (such as weights or neurons) in the neural network to reduce computational and storage requirements, mainly including unstructured pruning and structured pruning. Unstructured pruning is based on the absolute value of the weights for selection. The goal of pruning is to retain larger weights and remove small weights. After pruning, the weight matrix becomes a sparse form, and both storage and calculation are reduced. Structured pruning is based on a specific structure (such as an entire channel or layer) for removal. The goal of structured pruning is to remove specific output channels:

[0118] W′ = {W j，：，：，： , ||W j，：，：，： || F > T},

[0119] Among them, ||·|| is the Frobenius norm. Structured pruning is more friendly in hardware implementation, which can effectively reduce the matrix dimension and optimize the inference speed.

[0120] In this embodiment, the pruning process performed on the deep learning model is structured pruning, including the following processes:

[0121] Determine the transfer synaptic channels and neurons to be pruned, where the transfer synaptic channels are located between the input layer, the hidden layer, and the output layer;

[0122] Select a pruning criterion and perform a pruning operation. As Figure 4 shown, identify and remove redundant transfer synaptic channels and neurons, satisfying the expression: W′ = {W j，：，：，： , ||W j，：，：，： || F > T}, where ||·|| is the Frobenius norm.

[0123] Since the quantization method can compress the weights and activation functions of the model into low-bit values (such as binary or ternary), it can significantly reduce the storage requirements and computational complexity of the model. The quantization process performed on the deep learning model satisfies the expression:

[0124]

[0125] where x represents the weight or activation value, s is the quantization scale, and Q max is the quantization range (such as binary {-1, 1}).

[0126] However, direct quantization may lead to a decrease in model performance. Quantization-aware training (QAT) can effectively reduce the accuracy loss by simulating the impact of inference quantization during the training process and compensating for the quantization error. Represent the covariance matrix R i after averaging the sub-matrices in matrix form of amplitude and phase, expressed as:

[0127]

[0128] where A represents the amplitude factor of the covariance matrix R i , represents the phase factor. Since R i has conjugate symmetry, only retain the phase factor of the upper triangle. Design a feature matrix R i ′, whose upper triangle is the imaginary part of this phase factor, the lower triangle is the real part of this phase factor, and the diagonal elements are the normalized values of the diagonal elements of R i so that when more effective information is input into the deep learning model training or prediction, as follows:

[0129]

[0130] Vectorize it to obtain the feature vector r AP , which has q×q elements and is as follows:

[0131] r AP = vec(R i ′)

[0132] Utilize the conversion to the covariance amplitude feature vector r AP As the input feature, input it into the deep learning model; select the ReLU function as the activation function for each layer, and the expression is:

[0133]

[0134] During the process of using the covariance matrix for quantization-aware training of the deep learning model, introduce quantization error simulation, and the training objective is to minimize the loss of the deep learning model after quantization, and the expression is:

[0135]

[0136] The training objective of the pruned model is to minimize the impact loss of pruning on the model performance, and the expression is:

[0137]

[0138] Combine minimizing the loss of the deep learning model after quantization and minimizing the impact loss of pruning on the model performance. The final training objective is:

[0139]

[0140] When the training is completed, a trained deep learning model is obtained.

[0141] In this embodiment, before using the trained deep learning model for UAV DOA estimation, it further includes: adopting the methods of sparse matrix acceleration and pipeline optimization to perform hardware deployment on the trained deep learning model. Refer to Figure 2 , deploy the trained deep learning model on the hardware FPGA. Specifically, in hardware computing, sparse matrix multiplication can significantly reduce the number of operations and memory access requirements, thereby reducing the computational complexity and energy consumption. Since the covariance matrix itself has a certain sparsity in DOA estimation, through sparse matrix optimization, this characteristic can be fully utilized to achieve efficient utilization of hardware resources.

[0142] In this embodiment, convert the deep learning model operations into sparse matrix multiplications, and optimize the hardware computing efficiency by calculating the sparsity:

[0143]

[0144] Then, for the FPGA to implement the pipeline structure, the total delay is defined as:

[0145] Delay = number of stages × processing time per stage

[0146] And the memory bottleneck is reduced by processing the covariance matrix with sub-blocks.

[0147] Embodiment 3

[0148] This embodiment provides a computer storage medium, which stores an instruction program for executing the steps in the low-power scenario UAV DOA estimation method described above.

[0149] Embodiment 4

[0150] This embodiment provides an electronic device, which includes: a memory and a processor. The memory stores the program of the low-power scenario UAV DOA estimation method. When the program is executed by the processor, the steps of the low-power scenario UAV DOA estimation method are implemented.

[0151] Embodiment 5

[0152] As Figure 5 shown, this embodiment provides a low-power scenario UAV DOA estimation system, which includes: a received signal calculation unit, a fourth-order cumulant matrix construction unit, a covariance matrix conversion unit, a quantization pruning unit, a deep learning model training unit, and an estimation unit. The received signal calculation unit is used to construct an array signal model by receiving data with a sparse array to obtain an array received signal. The fourth-order cumulant matrix construction unit is used to construct the fourth-order cumulant matrix of the array received signal. The covariance matrix conversion unit is used to convert the fourth-order cumulant matrix into a covariance matrix. The quantization pruning unit is used to prune and quantize the deep learning model. The deep learning model training unit uses the covariance matrix to perform quantization-aware training on the deep learning model to obtain a trained deep learning model. The estimation unit uses the trained deep learning model to estimate the UAV DOA.

[0153] The embodiments are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill 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 implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims

1. A method for estimating DOA of drones in low-power scenarios, characterized in that: include: S1: construct an array signal model using sparse array receiving data to obtain the array receiving signal; S2: construct the fourth-order cumulant matrix of the array receiving signal; S3: Convert the fourth-order cumulant matrix into a covariance matrix; S4: Prune and quantize the deep learning model; S5: Use the covariance matrix to perform quantization-aware training on the deep learning model to obtain a trained deep learning model; S6: Use the trained deep learning model for UAV DOA estimation.

2. The low-power scenario UAV DOA estimation method according to claim 1 is characterized in that: The sparse array is designed using the extended Cantor array, and the designed sparse array element position set ε is: Where Δ represents the array element spacing reference, d m represents the position of the mth array element, M represents the total number of array elements, k m represents the array element index, Represents the set of positive integers.

3. The low-power scenario UAV DOA estimation method according to claim 1 is characterized in that: There are K drone signals in the space scene with angle θ k ={θ1,θ2,...,θ K } arrives at the sparse array, k = 1, 2, ..., K, the array signal model is: X(t) = A(Θ)S(t) + N(t), t = 1, 2, ..., T, t represents the tth sampling time; where, represents the array receiving signal, A(θ)=[a(θ1),…,a(θ K )] represents the direction vector matrix, which is an M×K matrix; Indicates the signal source; represents additive noise, which is non-Gaussian colored noise; each direction vector is defined as: Among them, d m Represents the position of the mth array element in the sparse array.

4. The low-power scenario UAV DOA estimation method according to claim 3 is characterized in that: The fourth-order cumulant matrix of the array received signal is expressed as: Among them, * represents conjugation, represents the Kronecker product, (·) H represents matrix transpose, E{·} represents expected calculation; The process of converting a fourth-order cumulant matrix to a covariance matrix consists of the following steps: S31: Vectorize the fourth-order cumulant matrix to obtain a vector F, which is expressed as: F=vec(C4)=Vu in, is the equivalent direction vector matrix composed of virtual elements in the fourth-order cumulant, is the equivalent signal vector, Represents the fourth-order cumulant of the source; Since there are repeated signals in F, the repeated signals are removed and processed according to the position relationship of the distribution array. The relationship is shown in the following formula: Remove duplicates and sort F to construct a new vector F'; S32: Reconstruct the vector F′ to form a new vector F1, which is expressed as: F1=(J T J) -1 J T F Where J represents the screening matrix; S33: Divide F1 into q overlapping sub-matrices Each sub-matrix has q dummy elements, and the covariance matrix of the i-th sub-array is defined as: S34: Using spatial smoothing technology, the equivalent covariance matrix is ​​divided into L sub-matrices, and the covariance matrices of the sub-matrices are averaged: Among them, R i is the covariance matrix of the ith submatrix, R smooth represents the covariance matrix after averaging the sub-matrices.

5. The low-power scenario UAV DOA estimation method according to claim 4 is characterized in that: The deep learning model adopts a multi-layer fully connected network, which is composed of an input layer, a hidden layer and an output layer. The input layer receives input data and passes it to the hidden layer. The hidden layer is composed of multiple neurons. Each neuron receives the output of all neurons from the previous layer, and after weighted summation and activation function processing, generates its own output and passes it to the next layer. The output layer is responsible for generating the final network output; the deep learning model uses the ReLu function as the activation function, and the activation function expression is: f(x) = max(0, x); The pruning process for deep learning models is structured pruning. The process includes: Determining a transmission synaptic channel and a neuron to be pruned, wherein the transmission synaptic channel is located between an input layer, a hidden layer, and an output layer; The weight matrix in the convolutional layer Select the pruning criteria, perform pruning operations, identify and remove redundant transmission synaptic channels and neurons, and satisfy the expression: W'={W j,:,:,: ,||W j,:,:,: || F >T}, where W j,:,:,: represents the j-th output channel, and ||·|| is the Frobenius norm. The quantization of the deep learning model satisfies the expression: Among them, x represents the weight or activation value, s is the quantization scale, Q max The quantification range.

6. The low-power scenario UAV DOA estimation method according to claim 5 is characterized in that: The covariance matrix R after averaging the sub-matrices i Expressed in matrix form of magnitude and phase, it is expressed as: Where A represents the covariance matrix R i The amplitude factor of represents the phase factor, due to R i With conjugate symmetry, only the upper triangular phase factor is retained, and a characteristic matrix R is designed i ′, its upper triangle is the imaginary part of the phase factor, its lower triangle is the real part of the phase factor, and the diagonal elements are R i The normalized values ​​of the diagonal elements are as follows to input more effective information into the deep learning model training or prediction: Vectorize it and get the feature vector r AP , there are q×q elements, as shown below: r AP =more(R i ′) Use the transformation into covariance amplitude eigenvector r AP As input features, it is input into the deep learning model; the ReLU function is selected as the activation function of each layer, and the expression is: In the process of using the covariance matrix to perform quantization-aware training on the deep learning model, quantization error simulation is introduced. The training goal is to minimize the loss of the deep learning model after quantization. The expression is: The training goal of the pruned model is to minimize the impact of pruning on model performance, expressed as: Combining minimizing the loss of the deep learning model after quantization with minimizing the impact of pruning on model performance, the final training goal is: When the training is completed, a trained deep learning model is obtained.

7. The low-power scenario UAV DOA estimation method according to claim 1 is characterized in that: Before using the trained deep learning model for UAV DOA estimation, it also includes: using sparse matrix acceleration and pipeline optimization methods to deploy the trained deep learning model on hardware.

8. A computer storage medium, characterized in that The invention stores an instruction program for executing the steps in the low-power scenario UAV DOA estimation method according to any one of claims 1 to 7.

9. An electronic device, characterized in that: The electronic device comprises: a memory and a processor, wherein the memory stores a program of the low-power scenario UAV DOA estimation method as described in any one of claims 1 to 7, and when the program is executed by the processor, the steps of the low-power scenario UAV DOA estimation method as described in any one of claims 1 to 7 are implemented.

10. A low-power scenario UAV DOA estimation system, characterized in that: The system is used to implement the low-power scenario UAV DOA estimation method according to any one of claims 1 to 7, comprising: A received signal calculation unit, used to construct an array signal model using sparse array received data to obtain an array received signal; A fourth-order cumulant matrix construction unit, used for constructing a fourth-order cumulant matrix of an array receiving signal; A covariance matrix conversion unit, used for converting a fourth-order cumulant matrix into a covariance matrix; Quantization pruning unit, used to prune and quantize deep learning models; The deep learning model training unit uses the covariance matrix to perform quantitative perception training on the deep learning model to obtain a trained deep learning model; The estimation unit uses the trained deep learning model to estimate the UAV DOA.

Images