A low-power consumption scene unmanned aerial vehicle (UAV) DOA estimation method, a computer storage medium, an electronic device and a system

Abstract

The application 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 fields of array signal processing and deep learning, and through construction of a fourth-order cumulative quantity matrix of received signals of a sparse array, non-Gaussian characteristics of signals in unmanned aerial vehicle DOA estimation are improved, hardware cost and complexity are reduced, DOA estimation precision is improved, a deep learning model is optimized in combination with pruning and quantization processing, and calculation complexity and storage requirements are reduced; the method, the computer storage medium, the electronic device and the system provided by the application are more suitable for accurate estimation of the direction of a signal source by an unmanned aerial vehicle.

Description

Technical Field

[0001] This invention relates to the technical fields of array signal processing and deep learning, and more specifically, to a method for estimating the DOA of a drone in a low-power scenario, a computer storage medium, an electronic device, and a system. Background Technology

[0002] In recent years, drones have experienced rapid development and have been widely applied in fields such as agricultural plant protection, police security, power line inspection, disaster relief, logistics transportation, and environmental monitoring due to their advantages of small size, high flexibility, and low cost. Direction of Atmosphere (DOA) estimation can accurately estimate the incident direction of drone signals, thereby improving the drone's positioning accuracy. This is crucial for drone navigation, obstacle avoidance, and path planning in these applications. However, these applications not only require drones to be highly efficient but also to minimize energy consumption while completing tasks, meaning that drones face limitations in terms of computing resources and energy consumption within embedded devices. Furthermore, drones in low-power applications have real-time and high-efficiency signal processing requirements. Due to the limited computing resources of drones, traditional array signal processing algorithms for DOA estimation, such as MUSIC, may be unsuitable due to their excessive computational complexity.

[0003] Higher-order statistical methods (such as those based on fourth-order cumulant matrices) are highly effective in suppressing Gaussian noise. In addition, deep learning technology has been introduced into the field of array signal processing, providing new solutions for DOA estimation of 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. Through large-scale data training, the robustness under non-Gaussian noise conditions and the accuracy of DOA estimation in low signal-to-noise ratio scenarios have been significantly improved. One existing technology discloses a deep learning-based DOA estimation method. This method constructs an array data model using received array data, obtains the output signal from the model, constructs a fourth-order cumulant of the output signal, obtains the fourth-order cumulant matrix, and performs vectorization processing. The processed result is then deduplicated, and a spatial smoothing algorithm is used to calculate the covariance matrix. Finally, the calculated covariance matrix is ​​input into a trained CNN model to obtain the DOA estimation result. This approach effectively improves the accuracy of the DOA estimation algorithm by combining sparse arrays, spatial smoothing algorithms, and deep neural network models to locate interference signals, while reducing computational load, hardware costs, and computational burden. However, such models typically rely on high-precision floating-point weights and activation functions, resulting in high computational complexity and storage requirements. Due to the computational resources and energy consumption limitations of UAVs, these deep learning models are difficult to deploy directly. Furthermore, existing deep learning methods generally neglect the hardware adaptability requirements in resource-constrained edge scenarios such as UAVs, lacking optimized designs for low-power applications.

[0004] Currently, research has explored several low-power optimization techniques to address the challenges faced by resource-constrained devices like drones. 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 values), but its application in array signal processing remains limited. Pruning and compression techniques remove redundant neurons or channels to reduce model complexity, but these methods are primarily focused on image processing and haven't yet been adequately designed for the specific task characteristics of array signal processing. Knowledge distillation transfers knowledge from high-precision models to low-precision models, improving the performance of lightweight models, but its application in complex signal processing scenarios such as DOA estimation remains insufficient. The stringent requirements of drone scenarios regarding real-time performance, power consumption, and algorithm robustness make it difficult for existing technologies to directly meet practical application requirements. Summary of the Invention

[0005] To address the issue that current methods for DOA estimation in UAVs using deep learning and array signal processing cannot meet the needs of UAVs in resource-constrained scenarios, this invention proposes a DOA estimation method and system for UAVs in low-power scenarios. Based on fourth-order cumulants, the non-Gaussian properties of the signal during UAV DOA estimation are improved. Furthermore, by combining optimized deep learning methods, computational complexity and storage requirements are reduced. Moreover, by incorporating sparse array design, hardware costs and complexity are reduced while improving DOA estimation accuracy. The method proposed in this invention is more suitable for accurate estimation of source directions by UAVs.

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

[0007] Firstly, this application proposes a method for estimating the DOA of a drone in low-power scenarios, including:

[0008] S1: Construct an array signal model using sparse array received data 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: Pruning and quantization of deep learning models;

[0012] S5: Use the covariance matrix to perform quantitative perception training on the deep learning model to obtain a trained deep learning model.

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

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

[0015]

[0016] Where Δ represents the reference spacing between array elements, and d m This represents the position of the m-th array element, where M represents the total number of array elements, and k represents the position of the m-th array element. m Indicates the element index, It represents the set of positive integers.

[0017] Based on the above technical means, by designing a sparse array, the virtual aperture is significantly expanded by utilizing non-uniform distribution, which can improve the resolution and accuracy of DOA estimation with a smaller number of physical array elements.

[0018] Preferably, there are K UAV signals in the spatial scene at an angle θ. k ={θ1, θ2, ..., θK The array reaches a sparse array, 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 time; where, The array receives the signal, A(Θ) = [a(θ1), ..., a(θ2)]. K )] represents the direction vector matrix, which is an M×K matrix; Indicates the signal source; This represents additive noise, which is non-Gaussian colored noise; each direction vector is defined as: Where, d m This 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 represented as:

[0020]

[0021] Where * denotes conjugation. Represents the Kronecker product, (·) H Let E{·} denote the matrix transpose, and let E{·} denote the expected computation.

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

[0023] S31: Vectorize the fourth-order cumulant matrix to obtain vector F, expressed as:

[0024] F = vec(C4) = Vu

[0025]

[0026] in, Let be the equivalent direction vector matrix composed of the virtual elements in the fourth-order cumulants. They are equivalent signal vectors. This represents the fourth-order cumulant of the information source;

[0027] Since there are repetitive signals in F, these repetitive signals are removed. The processing is based on the positional relationships of the distribution array, as shown in the following formula:

[0028]

[0029] Remove duplicates from F and sort it to construct a new vector F′;

[0030] S32: Reconstruct vector F′ to form a new vector F1, expressed as:

[0031] F1=(J T J)-1 J T F

[0032] Where J represents the screening matrix;

[0033] S33: Divide F1 into q overlapping submatrices. Each submatrix has q dummy elements, and the covariance matrix of the i-th subarray is defined as:

[0034] S34: Using spatial smoothing techniques, the equivalent covariance matrix is ​​divided into L sub-matrices, and the covariance matrices of the sub-matrices are averaged.

[0035]

[0036] Among them, R i R is the covariance matrix of the i-th submatrix. smooth This represents the covariance matrix after averaging the submatrices.

[0037] Based on the aforementioned techniques, the fourth-order cumulant matrix can extract non-Gaussian signal features and effectively suppress the influence of Gaussian noise, thereby significantly improving the robustness of DOA estimation in complex noise environments. Using spatial smoothing techniques to divide the submatrix into blocks and averaging their covariance matrices can effectively solve the correlation problem of array signals, while enhancing the positive definiteness of the matrix and eliminating the influence of array position offset.

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

[0039] The pruning process performed on deep learning models is called structured pruning, which includes the following steps:

[0040] Identify the transmission synaptic channels and neurons to be pruned, wherein the transmission synaptic channels are located between the input layer, hidden layer, and output layer;

[0041] Select pruning criteria, perform pruning operations, identify and remove redundant synaptic channels and neurons, satisfying the expression: W′={W j，：，：，： ,||W j，：，：，： ||F>T}, where W j，：，：，：Let ||·|| represent the j-th output channel, and ||·|| be the Frobenius norm.

[0042] The quantization process performed on 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 For the quantification range.

[0045] Based on the aforementioned techniques, pruning removes unimportant parameters (such as weights or neurons) from deep learning models to reduce computational and storage requirements. Structured pruning is more hardware-friendly, effectively reducing matrix dimensions and optimizing inference speed. Furthermore, by quantizing the weights and activation functions of deep learning models to low bit values ​​(such as binary or ternary values), the model's storage requirements and computational complexity can be significantly reduced.

[0046] Preferably, the covariance matrix R after averaging the submatrices is... i Represented as a matrix of magnitude and phase, it is expressed as:

[0047]

[0048] Where A represents the covariance matrix R i Amplitude factor, Represents the phase factor, since R i With conjugate symmetry, and retaining only the phase factor of the upper triangular region, design an eigenmatrix R. i Its upper triangle represents the imaginary part of the phase factor, and its lower triangle represents the real part of the phase factor. The diagonal elements are R. i The normalized values ​​of the diagonal elements are used to input more effective information into the training or prediction of deep learning models, as shown below:

[0049]

[0050] Vectorize it to obtain the feature vector r AP There are q×q elements, as shown below:

[0051] r Ap =vec(R i )

[0052] Transformed into a covariance magnitude eigenvector r AP The features are input into the deep learning model; the ReLU function is chosen as the activation function for each layer, and its expression is:

[0053]

[0054] In the process of quantization-aware training of a deep learning model using the covariance matrix, quantization error simulation is introduced. The training objective is to minimize the loss of the deep learning model after quantization, expressed as:

[0055]

[0056] The training objective of the pruned model is to minimize the impact of pruning on model performance loss, expressed as:

[0057]

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

[0059]

[0060] Upon completion of training, a well-trained deep learning model is obtained.

[0061] Based on the above technical means, by using quantization perception training to simulate the impact of inference quantization during the training process and to compensate for quantization errors, the loss of accuracy can be effectively reduced.

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

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

[0064] Secondly, this application proposes a computer storage medium storing an instruction program that executes the steps in the described low-power scenario UAV DOA estimation method.

[0065] Thirdly, this application proposes an electronic device comprising: a memory and a processor, wherein the memory stores a program for the low-power scenario UAV DOA estimation method, and when the program is executed by the processor, it implements the steps of the low-power scenario UAV DOA estimation method.

[0066] Fourthly, this application proposes a DOA estimation system for low-power UAVs, comprising:

[0067] The received signal calculation unit is used to construct an array signal model using sparse array received data and obtain the array received signal.

[0068] The fourth-order cumulant matrix construction unit constructs the fourth-order cumulant matrix of the array received signal;

[0069] The covariance matrix transformation unit is used to convert a fourth-order cumulant matrix into a covariance matrix;

[0070] The quantization pruning unit is used for pruning and quantizing deep learning models.

[0071] The deep learning model training unit uses the covariance matrix to perform quantitative perceptual training on the deep learning model to obtain a trained deep learning model.

[0072] The estimation unit uses a trained deep learning model to estimate the DOA of the drone.

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

[0074] This invention proposes a method for DOA estimation of unmanned aerial vehicles (UAVs) in low-power scenarios, a computer storage medium, electronic devices, and a system. By constructing a fourth-order cumulant matrix of the received signal in a sparse array, the non-Gaussianity of the signal during UAV DOA estimation is improved, reducing hardware costs and complexity while increasing DOA estimation accuracy. Furthermore, by combining pruning and quantization processing, the deep learning model is optimized, reducing computational complexity and storage requirements, making it more suitable for accurate estimation of the direction of the signal source by UAVs. Attached Figure Description

[0075] Figure 1 This is a flowchart illustrating the DOA estimation method for low-power UAVs in this embodiment of the invention.

[0076] Figure 2 This diagram illustrates the application of a low-power drone in an embodiment of the present invention.

[0077] Figure 3 A schematic diagram illustrating the composition of the deep learning model proposed in this embodiment of the invention;

[0078] Figure 4 This diagram illustrates the pruning process for deep learning models proposed in this embodiment of the invention.

[0079] Figure 5 This diagram illustrates the composition of the low-power scenario UAV DOA estimation system proposed in this embodiment of the invention. Detailed Implementation

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

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

[0082] It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings.

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

[0084] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0085] Example 1

[0086] This embodiment proposes a DOA estimation method for UAVs in low-power scenarios. The flowchart of this method can be found in [link to flowchart]. Figure 1 ,like Figure 1 As shown, it includes the following steps:

[0087] S1: Construct an array signal model using sparse array received data to obtain the array received signal;

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

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

[0090] S4: Pruning and quantization of deep learning models;

[0091] S5: Use the covariance matrix to perform quantitative perception training on the deep learning model to obtain a trained deep learning model.

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

[0093] In this embodiment, by constructing a fourth-order cumulant matrix of the received signal of the sparse array, the non-Gaussianity of the signal during UAV DOA estimation is improved, reducing hardware costs and complexity while improving DOA estimation accuracy. Furthermore, by combining pruning and quantization processing, the deep learning model is optimized, reducing computational complexity and storage requirements, making it more suitable for accurate estimation of the source direction by UAVs.

[0094] Example 2

[0095] Traditional uniform linear arrays (ULAs) suffer from insufficient resolution when the number of signal sources approaches or exceeds the number of array elements. Sparse arrays, however, significantly extend the virtual array aperture through non-uniform distribution, thereby improving the resolution and accuracy of DOA estimation with fewer physical array elements. In this embodiment, a sparse array is designed using an extended Cantor array, and the set of sparse array element positions ε is:

[0096]

[0097] Where Δ represents the reference spacing between array elements, and d m This represents the position of the m-th array element, where M represents the total number of array elements, and k represents the position of the m-th array element. m Indicates the element index, It represents the set of positive integers.

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

[0099] Figure 2 This diagram illustrates the application of a low-power drone in an embodiment of the present invention. Figure 2 In the diagram, 1 represents a drone, and there are K drone signals in the spatial scene at angles θ. k ={θ1, θ2, ..., θ K The array reaches a sparse array, 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 time; where, The array receives the signal, A(Θ) = [a(θ1), ..., a(θ2)]. K )] represents the direction vector matrix, which is an M×K matrix; Indicates the signal source; This represents additive noise, which is non-Gaussian colored noise; each direction vector is defined as: Where, d m This represents the position of the m-th element in the sparse array.

[0100] Traditional covariance matrices perform well under Gaussian white noise, but in real-world applications, noise is often non-Gaussian or colored, leading to a significant performance degradation of traditional methods. Higher-order cumulant matrices can extract non-Gaussian signal features and effectively suppress the influence of Gaussian noise, thus significantly improving the robustness of DOA estimation in complex noise environments. This embodiment constructs a fourth-order cumulant matrix for the array received signal, selecting it as the core method for feature extraction. Addressing the noise problem in the array signal model, the fourth-order cumulant matrix is ​​calculated to suppress Gaussian noise and extract non-Gaussian signal characteristics. The fourth-order cumulant vector matrix is ​​represented as:

[0101]

[0102] Where * denotes conjugation. Represents the Kronecker product, (·) H Let E{·} denote the matrix transpose, and E{·} denote the expected computation. In this embodiment, by introducing higher-order cumulants, an extended virtual array can be generated, whose direction matrix is ​​defined as: In sparse arrays, the correlation between signals can lead to the degradation of the covariance matrix. Spatial smoothing algorithms, by dividing the array into subarrays and averaging their covariance matrices, can effectively solve the correlation problem of array signals and enhance the positive definiteness of the matrix. Therefore, in this invention, to eliminate the influence of array position offset, spatial smoothing technology is used. 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 vector F, with the expression:

[0103] F = vec(C4) = Vu

[0104]

[0105] in, Let be the equivalent direction vector matrix composed of the virtual elements in the fourth-order cumulants. They are equivalent signal vectors. This represents the fourth-order cumulant of the information source;

[0106] Since there are repetitive signals in F, these repetitive signals are removed. The processing is based on the positional relationships of the distribution array, as shown in the following formula:

[0107]

[0108] Remove duplicates from F and sort it to construct a new vector F′;

[0109] S32: Reconstruct vector F′ to form a new vector F1, expressed as:

[0110] F1=(J T J) -1 J T F

[0111] Where J represents the screening matrix;

[0112] S33: Divide F1 into q overlapping submatrices. Each submatrix has q dummy elements, and the covariance matrix of the i-th subarray is defined as:

[0113] S34: Using spatial smoothing techniques, the equivalent covariance matrix is ​​divided into L sub-matrices, and the covariance matrices of the sub-matrices are averaged.

[0114]

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

[0116] In this embodiment, as Figure 3 As shown, the deep learning model employs a multi-layer fully connected network, which 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, which is composed of multiple neurons. Each neuron receives the outputs from all neurons in the previous layer, processes them through weighted summation and an activation function, and then generates its own output, which is passed 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 removes unimportant parameters (such as weights or neurons) from a neural network to reduce computational and storage requirements. It mainly includes unstructured pruning and structured pruning. Unstructured pruning selects weights based on their absolute values, aiming to retain larger weights and remove smaller ones. After pruning, the weight matrix becomes sparse, reducing both storage and computation. Structured pruning removes weights based on specific structures (such as entire channels or layers), aiming to remove specific output channels.

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

[0119] Here, ||·|| represents the Frobenius norm. Structured pruning is more hardware-friendly, effectively reducing matrix dimensions and optimizing inference speed.

[0120] In this embodiment, the pruning process performed on the deep learning model is structured pruning, which includes the following steps:

[0121] Identify the transmission synaptic channels and neurons to be pruned, wherein the transmission synaptic channels are located between the input layer, hidden layer, and output layer;

[0122] Select the pruning standard and perform the pruning operation, such as Figure 4 As shown, redundant synaptic channels and neurons are identified and removed, satisfying the expression: W′={W j，：，：，： ,||W j，：，：，： || F > T}, where ||·|| is the Frobenius norm.

[0123] Quantization methods can compress the model's weights and activation functions into low-bit values ​​(such as binary or ternary values), significantly reducing the model's storage requirements and computational complexity. The quantization process performed on deep learning models satisfies the expression:

[0124]

[0125] Where x represents the weight or activation value, s is the quantization scale, and Q... max For the quantization range (e.g., binary {-1, 1}).

[0126] However, direct quantization can lead to a decrease in model performance. Quantization-aware training (QAT) effectively reduces accuracy loss by simulating the effects of inference quantization during training and compensating for quantization errors. The covariance matrix R after averaging the submatrices is shown in the figure. i Represented as a matrix of magnitude and phase, it is expressed as:

[0127]

[0128] Where A represents the covariance matrix R i Amplitude factor, Represents the phase factor, since R i With conjugate symmetry, and retaining only the phase factor of the upper triangular region, design an eigenmatrix R. i Its upper triangle represents the imaginary part of the phase factor, and its lower triangle represents the real part of the phase factor. The diagonal elements are R. i The normalized values ​​of the diagonal elements are used to input more effective information into the training or prediction of deep learning models, as shown below:

[0129]

[0130] Vectorize it to obtain the feature vector r AP There are q×q elements, as shown below:

[0131] r AP =vec(R i ′)

[0132] Transformed into a covariance magnitude eigenvector r AP The features are input into the deep learning model; the ReLU function is chosen as the activation function for each layer, and its expression is:

[0133]

[0134] In the process of quantization-aware training of a deep learning model using the covariance matrix, quantization error simulation is introduced. The training objective is to minimize the loss of the deep learning model after quantization, expressed as:

[0135]

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

[0137]

[0138] By combining the loss of minimizing the deep learning model after quantization and the loss of minimizing the impact of pruning on model performance, the final training objective is:

[0139]

[0140] Upon completion of training, a well-trained deep learning model is obtained.

[0141] In this embodiment, before using the trained deep learning model for UAV DOA estimation, the method further includes: hardware deployment of the trained deep learning model using sparse matrix acceleration and pipeline optimization. See also Figure 2 The trained deep learning model is deployed on a hardware FPGA. Specifically, in hardware computing, sparse matrix multiplication can significantly reduce the number of operations and memory access requirements, thereby reducing computational complexity and energy consumption. Since the covariance matrix in DOA estimation itself has a certain sparsity, sparse matrix optimization can fully utilize this characteristic to achieve efficient use of hardware resources.

[0142] This embodiment transforms deep learning model operations into sparse matrix multiplication, optimizing hardware computational efficiency by calculating sparsity:

[0143]

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

[0145] Delay = Number of stages × Processing time per stage

[0146] Furthermore, memory bottlenecks are reduced by processing the covariance matrix through sub-blocks.

[0147] Example 3

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

[0149] Example 4

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

[0151] Example 5

[0152] like Figure 5As shown in the figure, this embodiment proposes a DOA estimation system for low-power UAVs, including: a received signal calculation unit, a fourth-order cumulant matrix construction unit, a covariance matrix transformation 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 using sparse array received data to obtain the 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 transformation 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-based 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's DOA.

[0153] The embodiments described are merely examples to clearly illustrate the present invention and are not intended to limit the implementation of the invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all possible implementations. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for estimating the DOA of a drone in low-power scenarios, characterized in that, include: S1: Construct an array signal model using sparse array received data to obtain the array received signal; S2: Construct the fourth-order cumulant matrix of the array received signal; S3: Convert the fourth-order cumulant matrix into a covariance matrix; S4: Pruning and quantization of deep learning models; The deep learning model employs a multi-layer fully connected network, consisting of an input layer, hidden layers, and an output layer. The input layer receives input data and passes it to the hidden layers, which are composed of multiple neurons. Each neuron receives the outputs from all neurons in the previous layer, processes them through weighted summation and an activation function, and then generates its own output, which is passed to the next layer. The output layer is responsible for generating the final network output. The deep learning model uses the ReLU function as its activation function, and the activation function expression is: ; The pruning process performed on deep learning models is called structured pruning, which includes the following steps: Identify the transmission synaptic channels and neurons to be pruned, wherein the transmission synaptic channels are located between the input layer, hidden layer, and output layer; The weight matrix of the convolutional layer Select a pruning criterion, perform the pruning operation, identify and remove redundant transmission synaptic channels and neurons, satisfying the expression: ,in, Let ||·|| represent the j-th output channel, and ||·|| be the Frobenius norm. The quantization process performed on the deep learning model satisfies the expression: in, Indicates weight or activation value. For quantitative scale, For quantification range; S5: Use the covariance matrix to perform quantitative perception training on the deep learning model to obtain a trained deep learning model. The covariance matrix Represented as a matrix of magnitude and phase, it is expressed as: Where A represents the covariance matrix Amplitude factor, Represents the phase factor, since With conjugate symmetry, and retaining only the phase factor of the upper triangular region, design an eigenmatrix. Its upper triangle represents the imaginary part of the phase factor, and its lower triangle represents the real part of the phase factor. The diagonal elements are... The normalized values ​​of the diagonal elements are used to input more effective information into the training or prediction of deep learning models, as shown below: Vectorize it to obtain the feature vector. ,have As shown below: Transform into covariance magnitude eigenvectors The features are input into the deep learning model; the ReLU function is chosen as the activation function for each layer, and its expression is: In the process of quantization-aware training of a deep learning model using the covariance matrix, quantization error simulation is introduced. The training objective is to minimize the loss of the deep learning model after quantization, expressed as: The training objective of the pruned model is to minimize the impact of pruning on model performance loss, expressed as: By combining the loss of minimizing the deep learning model after quantization and the loss of minimizing the impact of pruning on model performance, the final training objective is: Upon completion of training, a well-trained deep learning model is obtained. S6: Use the trained deep learning model for drone DOA estimation.

2. The DOA estimation method for low-power UAVs in claim 1, characterized in that, Designing a sparse array using an extended Cantor array, the set of element positions for the designed sparse array. ε for: in, Δ Indicates the reference for the spacing between array elements. Indicates the first The positions of the array elements, M represents the total number of array elements. k m Indicates the element index, It represents the set of positive integers.

3. The DOA estimation method for low-power UAVs in claim 1, characterized in that, let... Space Scene CCP K A drone signal at an angle ,..., Reaching a sparse array, The array signal model is as follows: , t Indicates the first t Each sampling time; among which, This indicates that the array receives signals. Let M represent the direction vector matrix. ; Indicates the signal source; This represents additive noise, which is non-Gaussian colored noise; each direction vector is defined as: ;in, In a sparse array, the first... The position of each array element.

4. The DOA estimation method for low-power UAVs in claim 3, characterized in that, The fourth-order cumulant matrix of the array received signal is represented as follows: in, Indicates conjugate. Indicates the Kronecker product. Indicates matrix transpose. This indicates the expected value calculation; The process of converting a fourth-order cumulant matrix into a covariance matrix includes the following steps: S31: Vectorize the fourth-order cumulant matrix to obtain a vector. F， The expression is: in, Let be the equivalent direction vector matrix composed of the virtual elements in the fourth-order cumulants. They are equivalent signal vectors. This represents the fourth-order cumulant of the information source; because There are duplicate signals. These duplicate signals are removed by processing them according to the positional relationship of the distribution array, as shown in the following formula: right Remove duplicates and sort the data to construct a new vector. ; S32: For vectors Reconstruct to form a new vector F 1. The expression is: in, J Represents the filtering matrix; S33: Will F 1 is divided into q overlapping submatrices. , i =1,2,3,...,q, each submatrix has q dummy elements, and the first... i The covariance matrix of each subarray is defined as: ; S34: Using spatial smoothing techniques, the equivalent covariance matrix is ​​divided into... L Each submatrix is ​​represented by a submatrix, and the covariance matrix of the submatrix is ​​averaged. in It is the first The covariance matrix of the submatrices, R smooth This represents the covariance matrix after averaging the submatrices.

5. The DOA estimation method for low-power UAVs in claim 1, characterized in that, Before using the trained deep learning model for drone DOA estimation, the following steps are required: hardware deployment of the trained deep learning model using sparse matrix acceleration and pipeline optimization methods.

6. A computer storage medium, characterized in that, The system stores an instruction program that executes the steps in the low-power scenario UAV DOA estimation method according to any one of claims 1 to 5.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a program for the low-power scenario UAV DOA estimation method according to any one of claims 1 to 5. When the program is executed by the processor, it implements the steps of the low-power scenario UAV DOA estimation method according to any one of claims 1 to 5.

8. A DOA estimation system for low-power drones, 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 5, including: The received signal calculation unit is used to construct an array signal model using sparse array received data and obtain the 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 transformation unit is used to convert a fourth-order cumulant matrix into a covariance matrix; The quantization pruning unit is used for pruning and quantizing deep learning models. The deep learning model training unit uses the covariance matrix to perform quantitative perceptual training on the deep learning model to obtain a trained deep learning model. The estimation unit uses a trained deep learning model to estimate the DOA of the drone.

Images