A deep reinforcement learning based direction of arrival estimation method and system
By combining the traditional MUSIC method with a deep augmentation network, the noise subspace is directly learned, which solves the limitations of traditional direction-of-arrival estimation methods under coherent signal source and model mismatch, and achieves high-resolution and robust direction-of-arrival estimation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI UNIV
- Filing Date
- 2022-11-01
- Publication Date
- 2026-05-29
AI Technical Summary
Traditional direction-of-arrival (DOA) estimation methods have limitations when dealing with coherent signal sources and model mismatches, and are computationally intensive and difficult to achieve high-resolution estimation.
By combining traditional MUSIC methods with deep augmentation networks, a deep augmentation network model is constructed. Utilizing CRNN and MLP layers, the noise subspace is directly learned, simplifying eigenvalue decomposition. The noise subspace is then used as training labels to improve estimation accuracy and robustness.
It achieves high-resolution performance estimation for multiple coherent signal sources, reduces computational complexity, and improves the robustness of estimation when there is a mismatch between the coherent source and the model.
Smart Images
Figure CN115687977B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of array signal processing, and more specifically, to a method and system for direction-of-arrival estimation based on depth enhancement. Background Technology
[0002] Direction of arrival (DoA) estimation technology has wide applications in fields such as radar, channel estimation, seismic exploration, and medical diagnosis.
[0003] In 1979, Schmidt proposed the Multiple Signal Classification (MUSIC) method. The traditional MUSIC method constructs spectral peaks based on the orthogonality of the signal and noise subspaces, and then performs peak search to estimate the direction of arrival (DOA). This method can estimate multiple targets simultaneously and has the advantage of high resolution compared to previous estimation methods. However, this method is difficult to implement, computationally intensive, and requires knowledge of noise characteristics and the number of signal sources, and it is almost impossible to estimate coherent sources, thus retaining significant limitations.
[0004] To overcome the limitations of the classic MUSIC method, improved versions have been proposed. For example, in 1984, Barbell proposed the root-finding MUSIC method, which improves resolution through polynomial root finding, but its drawback is that it is only applicable to linear arrays. Similarly, some researchers in China have also improved the classic MUSIC method. For instance, Zeng Hao et al. proposed using DFT to improve the MUSIC method for direction-of-arrival estimation, reducing the computational cost of traditional methods, but it still suffers from the same limitations as traditional methods.
[0005] With the continuous advancement of deep learning, researchers have begun to combine traditional model-based direction-of-arrival (DOA) estimation methods with neural networks. For example, in 2015, Xiong Xiao et al. used multilayer perceptrons to address the ODA problem in challenging environments, pioneering the integration of deep learning networks with traditional methods and proposing a classification problem for ODA estimation. In 2019, Chakrabarty S et al. proposed a supervised learning method based on convolutional neural networks for ODA estimation, where ODA estimation was formulated as a multi-class, multi-label classification problem. Kase Y et al. treated ODA estimation as a regression problem, applying deep neural networks to ODA estimation and evaluating its performance in the case of two equal-power, uncorrelated signals incident on a uniform linear array. The methods proposed above essentially implement model-independent ODA estimation using dense and convolutional neural networks. Because they are model-independent, these methods can handle array defects, but they involve high parameterization and lack the interpretability of model-based methods.
[0006] Hybrid model-based and data-driven methods for direction-of-arrival (DOA) estimation promise to address these shortcomings. For example, Elbir A M. proposed a multi-signal classification framework called Deep-MUSIC, which estimates discrete MUSIC spectra from the measured covariance matrix. The proposed Deep-MUSIC framework exhibits excellent estimation accuracy and low computational complexity, but like model-based methods, it uses the spatial spectrum as training labels. Barthelme A et al. proposed using a neural network to estimate the covariance matrix of the entire array from the sampled covariance matrices of subarrays. It models the DOA estimation problem as an end-to-end regression task, obtaining an estimated covariance matrix from incoherent subarrays via a neural network. This method enhances the robustness of the MUSIC approach; however, because the network is trained using the estimated covariance matrix as labels instead of the true covariance matrix, its neural network is not fully utilized. Summary of the Invention
[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and system for direction-of-arrival estimation based on depth enhancement.
[0008] According to one aspect of the present invention, a direction-of-arrival estimation method based on depth enhancement is provided, comprising:
[0009] Construct an array signal model;
[0010] Based on the array signal model, a dataset is established;
[0011] The dataset is preprocessed to divide it into training and test sets;
[0012] By combining the framework of traditional MUSIC methods with CRNN networks and MLP layers, a deep augmentation network model is constructed.
[0013] The network is trained based on the training set to obtain the optimal model of the deep augmentation network model;
[0014] The optimal model is used to estimate the direction of arrival.
[0015] Preferably, the array signal model is a far-field narrowband uniform linear array model.
[0016] Preferably, the dataset is selected from four types: coherent dataset, incoherent dataset, two closely spaced datasets, and datasets with mismatched models.
[0017] The covariance matrix was constructed and the true spectrum was calculated for each of the four datasets.
[0018] The data in the dataset and the processed data are divided into training set and test set according to a predetermined ratio, which are used to complete the training and testing work respectively.
[0019] Preferably, the deep augmentation network model incorporates pooling and normalization modules between its layers to optimize feature extraction and improve nonlinearity; it uses a noise subspace as a label to simplify the model.
[0020] Preferably, the process of input data in the deep augmentation network model includes:
[0021] Input data, the shape of which is (2m,T), where m is the number of array elements and T is the number of snapshots;
[0022] Transform the dimensions; the transformed shape is (T, 2m).
[0023] The data after dimensional transformation is normalized using a Batch Normalization (BN) layer.
[0024] After normalization, the data enters the CRNN layer to learn the noise subspace of the input data;
[0025] The data output from the CRNN layer is converted into complex values, and the transformation is performed in the complex space.
[0026] Stack the real and imaginary parts of the complex values;
[0027] Calculate the value of the incoming steering vector and the estimated value of the spatial spectrum of the noise subspace.
[0028]
[0029] Where a(θ) is the steering vector, E N For noise subspace;
[0030] The MLP layer is used to estimate d angles of arrival from the spatial spectrum.
[0031] Preferably, the CRNN layer includes:
[0032] The first convolutional layer includes a one-dimensional convolution with 64 channels, a kernel size of 5*5, a stride of 2, a padding mode of SAME, and an activation function of GeLU.
[0033] The first pooling layer includes one-dimensional max pooling, with a pooling size of 2, a step size of 2, and a fill mode of SAME.
[0034] The second convolutional layer includes a one-dimensional convolution, has 64 channels, a kernel size of 5*5, a stride of 2, a padding mode of SAME, and an activation function of GeLU.
[0035] The second pooling layer includes one-dimensional max pooling, with a pooling size of 2, a step size of 2, and a fill mode of SAME.
[0036] The GRU layer is a cyclic unit of an RNN network, with a size of 2m, where m is the number of elements.
[0037] Preferably, the MLP layer includes three fully connected dense layers and one linearly activated dense layer, wherein the number of neurons in the fully connected layer is 2m, the number of neurons in the linearly activated layer is d, which is the estimated angle of arrival of the number of output sources, and the activation function of the fully connected layer is set to ReLU.
[0038] According to a second aspect of the present invention, a direction-of-arrival estimation system based on depth enhancement is provided, comprising:
[0039] An array signal module, which constructs an array signal model;
[0040] The dataset module is responsible for creating datasets.
[0041] A preprocessing module preprocesses the dataset and divides it into training and test sets.
[0042] A deep augmentation network module, which combines the framework of the traditional MUSIC method with CRNN network and MLP structure to construct a deep augmentation network model;
[0043] The training module trains the network based on the preprocessed dataset to obtain the optimal model of the deep augmentation network model.
[0044] The test module uses the optimal model to estimate the direction of arrival.
[0045] According to a third aspect of the present invention, a terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, can be used to perform the methods described above, or to run the systems described above.
[0046] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can be used to perform the methods described above, or to run the system described above.
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] This invention discloses a depth-enhanced direction-of-arrival (DOA) estimation method and system, which uses a traditional MUSIC method as its architecture and combines data-driven and traditional MUSIC model-based approaches. This method overcomes the limitations of traditional MUSIC methods, enabling the detection of multiple coherent signal sources and exhibiting higher resolution. It involves a deep enhancement network that incorporates a convolutional recurrent neural network (CNN) to directly learn the noise subspace from the input data, thus improving feature extraction. Furthermore, it utilizes an MLP structure to estimate the DOA from the estimated spatial spectrum.
[0049] This invention presents a depth-enhanced direction-of-arrival (DOA) estimation method and system. To reduce network complexity, the eigenvalue decomposition of the traditional MUSIC method to construct a noise subspace is removed. This noise subspace is then directly learned through a CRNN network, and the training labels are estimated using the noise subspace, thus simplifying the structure. Furthermore, robustness is improved in cases of coherent source estimation and model mismatch. Attached Figure Description
[0050] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0051] Figure 1 This is a flowchart of a direction-of-arrival estimation method based on depth enhancement in one embodiment of the present invention;
[0052] Figure 2 This is a schematic diagram of the structure of a deep augmentation network in a preferred embodiment of the present invention;
[0053] Figure 3 This is a detailed schematic diagram of the CRNN layer in a preferred embodiment of the present invention. Detailed Implementation
[0054] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.
[0055] See Figure 1 This invention provides an embodiment of a direction-of-arrival estimation method based on depth enhancement, comprising:
[0056] S100, construct the array signal model;
[0057] S200, an array signal model built based on S100, establishes a dataset;
[0058] S300 preprocesses the dataset established in S200, dividing it into training and test sets;
[0059] S400 combines the framework of traditional MUSIC methods with CRNN networks and MLP structures to construct a deep augmentation network model;
[0060] S500: Based on the dataset preprocessed by S300, the network is trained to obtain the optimal model of the deep augmented network model.
[0061] S600, use the optimal model to estimate the direction of arrival.
[0062] In this embodiment, the present invention employs a regression learning-based direction-of-arrival estimation algorithm, which enhances the original MUSIC structure while overcoming the limitations of traditional model-based methods and simplifying the structure of traditional MUSIC models.
[0063] In one embodiment of the present invention, S100 is implemented to construct an array signal model, which is a far-field narrowband uniform linear array model.
[0064] In a preferred embodiment of the present invention, S200, the establishment and partitioning of the dataset, is implemented as follows:
[0065] First, the dataset is established. This invention discloses four types of dataset implementations: coherent dataset, incoherent dataset, two closely spaced datasets (with two signal sources and a small angular difference between them), and a model-mismatched dataset. The parameters involved in generating the datasets are as follows: the number of signal sources is set to 5 (specifically, when generating two closely spaced datasets, the number of signal sources is set to 2), the signal-to-noise ratio is 10dB, the number of array elements is 8, and the number of samples is 10000. The parameter settings in this embodiment make the results more apparent.
[0066] The coherent and incoherent datasets are set up to compare the performance differences between the present invention and the traditional MUSIC method, demonstrating that the present invention can overcome the limitation of traditional methods in estimating coherent sources. The use of two closely spaced sources shows that the present invention can estimate two closely spaced signals compared to the traditional MUSIC method, resulting in improved resolution. A model mismatch dataset is used, where each term of the steering vector is contaminated with Gaussian noise and does not match the corresponding angle, to analyze the robustness of the present invention.
[0067] In a preferred embodiment of the present invention, S300 is implemented to perform preprocessing. Specifically, the covariance matrix and the true spectrum are calculated based on the dataset, and then the training set and the test set are divided in a 1:9 ratio to complete the training and testing work.
[0068] In a preferred embodiment of the present invention, S400 is implemented to construct a deep augmentation network model, which is a network architecture based on the framework of the traditional MUSIC method and combined with CRNN network and MLP structure. Pooling and normalization modules are combined between layers to optimize feature extraction and improve nonlinearity. At the same time, the network model is further simplified by using the noise subspace as the label.
[0069] In a preferred embodiment, the initial learning rate of the network in this step is set to 0.001, the Adam optimizer is used, the number of iterations is 110, and the model is trained according to... Figure 2 As shown, the specific parameters for the order are as follows:
[0070] 1) The shape of the input data is (2m,T), where m is the number of array elements and T is the number of snapshots;
[0071] 2) Transform the dimensions; the transformed shape is (T, 2m).
[0072] 3) BN layer, perform normalization processing;
[0073] 4) The CRNN layer consists of two convolutional layers, two pooling layers, and one GRU layer. It learns the noise subspace of the input data and obtains features.
[0074] 5) Perform transformations in the complex space;
[0075] 6) Stacking the real and imaginary parts of the above complex values can more comprehensively process input data information, learn the noise subspace, and better extract the features of the input data.
[0076] 7) Calculate the spatial spectrum using the input steering vector and the estimated noise subspace, as shown in the following formula:
[0077]
[0078] Where a(θ) is the steering vector, E N This is the noise subspace.
[0079] 8) The estimated spatial spectrum is input into the MLP layer for angle of arrival estimation; the MLP layer specifically includes three fully connected dense layers and one linearly activated dense layer, where the number of neurons in the fully connected layer is 2m, where m is the number of array elements, and the number of neurons in the linearly activated layer is d, which is the estimated angle of arrival of the number of output sources.
[0080] 9) Estimate d angles of arrival from the spatial spectrum and output the estimation results.
[0081] In this embodiment, to reduce network complexity, the eigenvalue decomposition to construct a noise subspace in the traditional MUSIC method structure is removed. The noise subspace is then directly learned through the CRNN network, and the training labels are estimated using the noise subspace, thus simplifying the structure. Simultaneously, robustness to estimations of coherent sources and model mismatches is improved.
[0082] In another preferred embodiment, the specific network model within the CRNN network is as follows: Figure 3 The parameters shown are as follows:
[0083] (1) Convolutional layer 1, one-dimensional convolution, the number of channels in the convolutional layer is 64, the kernel size is 5*5, the stride is set to 2, the padding mode is set to SAME, and the activation function is set to GeLU;
[0084] (2) Pooling layer 1, one-dimensional max pooling, pooling size is 2, stride is set to 2, fill mode is set to SAME;
[0085] (3) Convolutional layer 2, one-dimensional convolution, the number of channels in the convolutional layer is 64, the kernel size is 5*5, the stride is set to 2, the padding mode is set to SAME, and the activation function is set to GeLU;
[0086] (4) Pooling layer 2, one-dimensional max pooling, pooling size is 2, step size is set to 2, and fill mode is set to SAME;
[0087] (5) GRU layer, which is the cyclic unit of the RNN network, with a size of 2m, where m is the number of array elements.
[0088] This CRNN layer can directly learn the noisy subspace from the input data, extracting features and capturing long-term dependencies better than a single CNN or RNN network. This is achieved by comparing its estimated DoA with the true DoA, and working with subsequent MLP layers to fine-tune the training of the noisy subspace and the ability to convert the music spectrum into DoA.
[0089] In another embodiment of the present invention, the performance of the method is evaluated by testing with a test set. The evaluation metric uses the root mean square periodicity error (RMSPE) to compare the estimated DoA angle with the true DoA angle. The calculation formula is as follows:
[0090]
[0091] Where mod(·) represents the modulo operation, ||·|| represents the norm, d is the number of sources, and θ is the true angle of arrival. For the estimated angle of arrival, Pd It is the set of all d×d permutations, and P represents the set of estimated angles. The lower the value of RMSPE, the higher the positioning accuracy and the better the performance of the system.
[0092] Based on the same inventive concept, in other embodiments of the present invention, a direction-of-arrival estimation system based on depth enhancement is provided, including an array signal module, a dataset module, a preprocessing module, a deep enhancement network module, a training module, and a testing module; the array signal module constructs an array signal model; the dataset module establishes a dataset; the preprocessing module preprocesses the dataset and divides it into training and testing sets; the deep enhancement network module combines the framework of the traditional MUSIC method with CRNN networks and MLP structures to construct a deep enhancement network model; the training module trains the network based on the preprocessed dataset to obtain the optimal model of the deep enhancement network model; the testing module uses the optimal model to perform direction-of-arrival estimation and uses evaluation metrics to evaluate the system performance.
[0093] Based on the same inventive concept, in other embodiments of the present invention, a terminal is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can be used to perform the above-described method or to run the above-described system.
[0094] Based on the same inventive concept, in other embodiments of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can be used to perform the above-described method or to run the above-described system.
[0095] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention. The above preferred features can be used in any combination without conflict.
Claims
1. A direction-of-arrival estimation method based on depth enhancement, characterized in that, include: Construct an array signal model; Based on the array signal model, a dataset is established; The dataset is preprocessed to divide it into training and test sets; By combining the framework of traditional MUSIC methods with CRNN networks and MLP layers, a deep augmentation network model is constructed. The network is trained based on the training set to obtain the optimal model of the deep augmentation network model; The optimal model is used for direction of arrival estimation; The deep augmentation network model incorporates pooling and normalization modules between its layers to optimize feature extraction and improve nonlinearity. It uses the noise subspace as a label to simplify the model; The process of input data in a deep augmented network model includes: The input data has a shape of (2m, T), where m is the number of array elements and T is the number of snapshots. Transform the dimensions, and the transformed shape is (T, 2m). The data after dimensional transformation is normalized using a Batch Normalization (BN) layer. After normalization, the data enters the CRNN layer to learn the noise subspace of the input data; The data output from the CRNN layer is converted into complex values, and the transformation is performed in the complex space. Stack the real and imaginary parts of the complex values; Calculate the value of the incoming steering vector and the estimated value of the spatial spectrum of the noise subspace. ； in, As the guiding vector, For noise subspace; The MLP layer is used to estimate d angles of arrival from the spatial spectrum.
2. The direction-of-arrival estimation method based on depth enhancement according to claim 1, characterized in that, The array signal model selected is a far-field narrowband uniform linear array model.
3. The direction-of-arrival estimation method based on depth enhancement according to claim 1, characterized in that, The dataset used consists of four types: coherent dataset, incoherent dataset, two closely spaced datasets, and datasets with mismatched models. The covariance matrix was constructed and the true spectrum was calculated for each of the four datasets. The data in the dataset and the processed data are divided into training set and test set according to a predetermined ratio, which are used to complete the training and testing work respectively.
4. The direction-of-arrival estimation method based on depth enhancement according to claim 1, characterized in that, The CRNN layer includes: The first convolutional layer comprises a one-dimensional convolution with 64 channels and a kernel size of 5.
5. Set step size to 2, fill mode to SAME, and activation function to GeLU; The first pooling layer includes one-dimensional max pooling, with a pooling size of 2, a step size of 2, and a fill mode of SAME. The second convolutional layer includes a one-dimensional convolution, has 64 channels, and a kernel size of 5.
5. Set step size to 2, fill mode to SAME, and activation function to GeLU; The second pooling layer includes one-dimensional max pooling, with a pooling size of 2, a step size of 2, and a fill mode of SAME. The GRU layer is a cyclic unit of an RNN network, with a size of 2m, where m is the number of elements.
5. The direction-of-arrival estimation method based on depth enhancement according to claim 1, characterized in that, The MLP layer includes three fully connected dense layers and one linearly activated dense layer. The number of neurons in the fully connected layers is 2m, and the number of neurons in the linearly activated layer is d, which is the estimated angle of arrival of the number of output sources. The activation function of the fully connected layers is set to ReLU.
6. A direction-of-arrival estimation system based on depth enhancement, used to implement the method according to any one of claims 1-5, characterized in that, include: An array signal module, which constructs an array signal model; The dataset module is responsible for creating datasets. The preprocessing module preprocesses the dataset, dividing it into training and test sets; A deep augmentation network module, which combines the framework of the traditional MUSIC method with CRNN network and MLP layer to construct a deep augmentation network model; The training module trains the network based on the preprocessed dataset to obtain the optimal model of the deep augmentation network model. The test module uses the optimal model to estimate the direction of arrival.
7. A terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it can be used to perform the method of any one of claims 1-5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, this program can be used to perform the method of any one of claims 1-5.