Denoising optimization DOA estimation method under non-ideal channel condition

Abstract

The invention relates to a denoising optimization DOA estimation method under a non-ideal channel condition, and the method comprises the steps: constructing a simulation data set under a simulation non-ideal channel condition, and constructing an improved denoising and classification network; and after training the improved denoising and classification network, acquiring test signal data, and inputting the test signal data into the trained model to obtain orientation estimation. According to the method, multiple complex factors are introduced into original signals, a training data set closer to an actual scene is constructed, and the problems of data scarcity and single type are solved; a noise adding-denoising scheme based on a diffusion model is provided, the data is enhanced and optimized, the influence of noise is suppressed, and the feature expression ability of the signal is improved; a multi-dimensional feature fusion complex attention mechanism deep learning framework is designed, efficient integration of multi-dimensional features is achieved, the problem of signal multi-dimensional feature splitting is solved, and therefore the robustness and precision of a model in a complex environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of calculation, estimation or counting, and in particular to a DOA estimation method with denoising optimization under non-ideal channel conditions. Background Art

[0002] Direction estimation, especially signal direction estimation in underwater scenes, is one of the important technologies in the field of signal processing. Usually, the spatial position inference of moving targets or sound sources relies on array signal processing technology, and the direction of the sound source or target is inferred by the received array signal. However, direction estimation in underwater scenes faces many challenges. First, the complexity of the underwater environment has a significant impact on signal propagation, such as scattering, attenuation, and various environmental interferences, which causes the spatial distribution characteristics of the signal to change and increases the accuracy of direction estimation; second, the signals between multiple targets often have high similarities, making it difficult to distinguish the directions of multiple targets in complex scenes, especially in multi-target tasks; finally, the collection of underwater signals faces high costs and is usually accompanied by limited data. These factors make how to improve the accuracy of direction estimation, especially in underwater scenes, under the conditions of severe noise interference and scarce data a key research issue.

[0003] More importantly, in actual application scenarios, there are various interference factors such as antenna position disturbance, gain / phase inconsistency, mutual coupling effect, nonlinear amplifier effect, etc., and the existing traditional direction estimation methods are usually based on the ideal antenna array model, namely the uniform linear array (ULA). Since the interference is ignored, in actual direction estimation applications, the performance of conventional direction estimation algorithms will be significantly affected and the generalization ability is poor.

[0004] To address these challenges, it is necessary to redesign the orientation estimation methods to adapt to underwater scenarios, especially to propose more robust solutions for scenarios with data imbalance, noise interference, and limited computing resources. Summary of the invention

[0005] The present invention solves the problems existing in the prior art and proposes a denoising and optimized DOA estimation method under non-ideal channel conditions.

[0006] The technical solution adopted by the present invention is a DOA estimation method for denoising optimization under non-ideal channel conditions, the method comprising the following steps:

[0007] S1 constructs a simulation data set under non-ideal channel conditions;

[0008] S2 builds an improved denoising and classification network;

[0009] S3 trains the improved denoising and classification network;

[0010] S4 obtains test signal data, inputs it into the trained model, and obtains a direction estimate.

[0011] Preferably, S1 comprises the following steps:

[0012] S1.1 Determine the DOA range and generate all non-overlapping DOA combinations based on the maximum number of targets, number of antennas, and resolution input;

[0013] S1.2 calculates the signal steering vector according to the generated DOA combination, the structure of the antenna array and the set resolution;

[0014] S1.3 applying different types of random perturbations to the antenna array to perform data enhancement, and constructing several training sets with different types of perturbations;

[0015] S1.4 Assign a unique position index to any data in the training set.

[0016] Preferably, the disturbance type includes one or more of position disturbance, gain disturbance, phase disturbance, mutual coupling disturbance and nonlinear effect disturbance.

[0017] Preferably, at least one training set does not include a signal augmented with random perturbations.

[0018] Preferably, the improved denoising and classification network comprises a diffusion-based denoising module and a complex attention mechanism module arranged in sequence;

[0019] The diffusion-based denoising module includes a forward denoising unit and a backward denoising unit.

[0020] Preferably, the backward denoising unit comprises N downsampling layers and N upsampling layers arranged in sequence, and the nth downsampling layer is jump-connected to the nth-to-last upsampling layer.

[0021] Preferably, the complex attention mechanism module includes a parallel complex operation unit and a channel attention unit, and the outputs of the two are aggregated and input into the self-attention unit.

[0022] Preferably, the complex operation unit comprises a complex convolution layer, a complex batch normalization layer and a complex activation function connected in sequence.

[0023] Preferably, a loss function of the improved denoising and classification network is constructed, and the loss function is associated with the error between the denoising output of the diffusion-based denoising module and the real signal, and the distance between the classification prediction distribution probability of the training sample and the real distribution.

[0024] The present invention relates to a denoising-optimized DOA estimation method under non-ideal channel conditions, which constructs a simulation data set simulating non-ideal channel conditions and builds an improved denoising and classification network. After training the improved denoising and classification network, test signal data is obtained and input into the trained model to obtain azimuth estimation.

[0025] The beneficial effects of the present invention are:

[0026] (1) Introducing multiple complex factors such as antenna mutual coupling effect, position disturbance, amplitude and phase disturbance, and nonlinear distortion into the original signal to build a training data set that is closer to the actual scene and solve the problems of data scarcity and single type;

[0027] (2) To address the problem of insufficient signal feature expression in low signal-to-noise ratio environments, a denoising scheme based on a diffusion model is proposed to enhance and optimize data, suppress the impact of noise, and improve the signal feature expression capability.

[0028] (3) To address the problem of multi-dimensional feature separation of the amplitude, phase, real part, and imaginary part of the reconstructed signal, a deep learning framework with a complex attention mechanism for multi-dimensional feature fusion was designed to achieve efficient integration of multi-dimensional features and solve the problem of multi-dimensional feature separation of the signal, thereby improving the robustness and accuracy of the model in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flow chart of the present invention;

[0030] Figure 2 A structural diagram of an improved denoising and classification network of the present invention;

[0031] Figure 3 This is a comparison chart of the experimental results of the present invention and the classic DOA algorithm. DETAILED DESCRIPTION

[0032] The present invention is further described in detail below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto.

[0033] A DOA estimation method with denoising optimization under non-ideal channel conditions, the method comprising the following steps:

[0034] S1 constructs a simulation data set under non-ideal channel conditions;

[0035] S2 builds an improved denoising and classification network;

[0036] S3 trains the improved denoising and classification network;

[0037] S4 obtains test signal data, inputs it into the trained model, and obtains a direction estimate.

[0038] The steps are described below in conjunction with specific implementation methods.

[0039] (1) Construct a simulation data set under non-ideal channel conditions;

[0040] In the present invention, the array model is ULA, which uses the characteristic of fixed phase difference on the array signal to add a phase difference to the signal to simulate the original data reaching the array, and then generates multiple training sets with different physical conditions by simulating factors such as antenna position disturbance, gain / phase inconsistency, mutual coupling effect and nonlinear effect.

[0041] Specifically, the following steps are included:

[0042] (1-1) Determine the DOA range and generate all non-overlapping DOA combinations based on the maximum number of targets, number of antennas, and resolution input;

[0043] Ensure that there is enough separation distance between each DOA to avoid signal confusion;

[0044] During implementation, signal parameter settings include:

[0045] DOA range: [-60°, 60°], simulating signals reaching the antenna array from different angles, with a resolution of 1°, and 121 angles are divided in a grid manner;

[0046] Number of targets: Randomly generate the number of targets, ranging from 1 to 3 targets, and generate all angle combinations according to the full permutation of the number of targets;

[0047] Array parameters: ULA array consisting of 16 radars;

[0048] Signal-to-noise ratio parameter: -10db;

[0049] Interference factor parameter settings include:

[0050] Position perturbation: The maximum standard deviation is set to 0.15 to simulate the physical position deviation of each antenna in the actual array; position perturbation affects the phase difference of the angle of arrival (DOA) of the signal on different antennas, simulating the impact of inconsistent antenna positions;

[0051] Gain perturbation: The maximum standard deviation is set to 0.5 to simulate the gain difference when the antenna receives the signal. Inconsistent antenna gain will cause the signal to show different amplitudes on different antennas, reflecting the actual situation of antenna equipment or channel attenuation.

[0052] Phase perturbation: The maximum standard deviation is set to 0.2, which is used to simulate the phase error of the signal when it is received by the antenna. Phase perturbation introduces random phase offset, causing the phase of the signal received by each antenna to deviate, simulating the phase inconsistency of the hardware device;

[0053] Mutual coupling effect: The maximum standard deviation is set to 0.06 to simulate the electromagnetic interference effect between antennas. The mutual coupling effect causes the signals in the antenna array to affect each other, especially between adjacent antennas. The interference will cause the received signal to be distorted, affecting the accuracy of the signal.

[0054] Nonlinear effect: The intensity is set to 1.0 to simulate nonlinear distortion in hardware devices. By using a nonlinear excitation function (tanh), the nonlinear effect transforms the signal nonlinearly to simulate the signal distortion caused by nonlinearity in actual devices.

[0055] This involves the data generation method and the maximum limit parameters of each disturbance.

[0056] (1-2) According to the generated DOA combination, the structure of the antenna array and the set resolution, the signal steering vector is calculated to simulate the phase offset of the signal on the antenna array under different DOAs. These vectors are used to construct the array signal of the training data;

[0057] The signal's steering vector satisfies,

[0058]

[0059] Among them, θ k is the DOA angle of the kth signal source, λ is the wavelength of the signal, d is the antenna spacing, which is usually set to half the wavelength, and n is the index of the antenna;

[0060] The array signal can be expressed as,

[0061]

[0062] Among them, s k (t) represents the transmission signal, y(t) represents the array signal;

[0063] (1-3) applying different types of random perturbations to the antenna array to perform data enhancement, and constructing several training sets with different perturbation types;

[0064] The disturbance type includes one or more of position disturbance, gain disturbance, phase disturbance, mutual coupling disturbance, and nonlinear effect disturbance.

[0065] Specifically, it includes:

[0066] (1-3-1) Set position disturbance: Generate a position deviation that follows a normal distribution for each antenna, with the maximum standard deviation set to 0.15, to simulate the impact of inconsistent antenna positions, satisfying:

[0067]

[0068] Among them, a(θ k ) is the steering vector with position disturbance, θ is the DOA angle, λ is the wavelength of the signal, d is the antenna spacing, n is the index of the antenna, d per is the physical position disturbance of the antenna;

[0069] (1-3-2) Set gain perturbation: Generate a gain value that follows a normal distribution for each antenna, with a maximum standard deviation of 0.5. The receiving gains of different antennas are inconsistent, which causes the signal to show different amplitudes on different antennas, satisfying:

[0070] s gain (t) = gain × y(t)

[0071] Where y(t) represents the array signal, gain is the gain disturbance of the signal;

[0072] (1-3-3) Set phase perturbation: Phase perturbation is used to simulate the phase error of the signal when it is received by the antenna. This perturbation will introduce random phase offset, causing the phase of the signal received by each antenna to deviate, simulating the phase inconsistency of the hardware device. Phase perturbation is introduced for each antenna. The generated phase offset follows a normal distribution with a maximum standard deviation of 0.2. It is used to simulate the phase error when the antenna receives the signal and satisfies:

[0073] s phase (t) = y(t) × exp(j·phase)

[0074] Where y(t) represents the array signal, and phase is the phase disturbance of the signal;

[0075] (1-3-4) Implementation of mutual coupling perturbation: The mutual coupling effect is simulated by the mutual coupling matrix. There is an electromagnetic interference problem between the antennas. The maximum mutual coupling effect is controlled by its maximum standard deviation of 0.06. The diagonal elements of the matrix are 1, indicating that the influence of each antenna on itself is 1. The non-diagonal elements represent the mutual coupling effect between different antennas. The value decays as the distance between the antennas increases, satisfying,

[0076] s mc (t) = M mc ×y(t)

[0077] Among them, M mc is the mutual coupling matrix, y(t) is the array signal, smc (t) is the signal after superimposing the mutual coupling effect;

[0078] (1-3-5) Introduction of nonlinear effects: The nonlinear effect simulates the nonlinear distortion in the hardware device through the nonlinear excitation function (tanh). The signal is transformed under the action of the nonlinear function.

[0079] s nl (t) = tanh(σ nonlinear ×y(t)

[0080] Among them, σ nonlinear is the parameter for controlling the intensity of nonlinear effect, y(t) is the array signal, s nl (t) is the signal after superimposing nonlinear effects;

[0081] (1-3-6) Signal normalization: The generated signals and various interferences are normalized; by standardizing the energy of the signals, the power of each signal is ensured to be within the same range, so as to maintain the consistency of different signals during the training process;

[0082] At least one of the training sets does not include a signal for data augmentation with random perturbations.

[0083] In practical applications, the training set includes:

[0084] Training set A: contains only array signals without any perturbations and nonlinear effects;

[0085] Training set B: contains the combination of array signal and position disturbance signal;

[0086] Training set C: contains the combination of array signal and gain perturbation signal;

[0087] Training set D: contains the combination of array signal and phase disturbance signal;

[0088] Training set E: contains the combination of array signals and mutual coupling disturbance signals;

[0089] Training set F: contains the combination of array signals and nonlinear effect signals;

[0090] Training set G: Contains a combination of position disturbance, gain disturbance, phase disturbance, mutual coupling disturbance and nonlinear effect signals (all disturbance factors exist at the same time).

[0091] (1-4) Assign a unique position index to any data in the training set;

[0092] The DOA estimation task is converted into a multi-classification task. For each DOA, a unique class index is assigned, from 0 to num_classes-1, where num_class represents the total number of classes. In this patent, the angle range of DOA is set to -60° to 60° with a resolution of 1°, which means that the angle range is divided into 121 classes, each class represents a specific DOA angle position, -60° corresponds to class index 0, -59° corresponds to class index 1, and so on, until 60° corresponds to class index 120. In this way, the DOA estimation problem can be expressed as a multi-classification task with 121 classes.

[0093] (2) Construct an improved denoising and classification network;

[0094] The improved denoising and classification network includes a diffusion-based denoising module and a multiple attention mechanism module arranged in sequence.

[0095] (2-1) The diffusion-based noise reduction module includes a forward noise addition unit and a backward noise removal unit;

[0096] Due to the interference of noise and non-ideal factors, traditional algorithms cannot distinguish between noise, interference and signals. Therefore, a signal denoising framework (DOA-Diffusion) based on diffusion model is designed. This framework utilizes the two processes of forward denoising (diffusion process) and backward recovery (reverse process) in the diffusion model.

[0097] Specifically, it includes:

[0098] (2-1-1) In the forward noise addition stage, we draw on the idea of ​​non-equilibrium thermodynamics to gradually add Gaussian noise to the original covariance matrix, so that it eventually approaches the white noise distribution. This process is modeled by the Markov chain, and the covariance matrix at the current moment only depends on the state at the previous moment. By introducing the time step parameter t and the noise scheduling parameter α t , gradually increase the intensity of Gaussian noise to simulate the diffusion process of the covariance matrix in the potential feature space. The goal of forward diffusion is to establish the joint distribution of noise and signal features by adding noise perturbations, providing a training basis for subsequent denoising and recovery. The process of forward denoising satisfies:

[0099]

[0100] Among them, x t-1 The signal at the current time step t-1 is gradually evolved from the initial signal x0. t The signal representing the current time step t represents the signal of the previous time step x t-1 Add noise, αt The noise control factor controls the degree of signal retention at the current time step, α t The larger the value, the more original signal components are retained. t is Gaussian noise that conforms to the standard normal distribution;

[0101] (2-1-2) In the backward recovery stage, the trained DOA denoising network is used to gradually restore the high-noise covariance matrix to the original noise-free matrix;

[0102] The backward denoising unit includes N downsampling layers and N upsampling layers arranged in sequence, and the nth downsampling layer is jump-connected to the nth upsampling layer from the end.

[0103] Specifically, the denoising network uses ResUnet, which includes 5 downscaling layers and 5 upscaling layers connected in sequence, and realizes the transmission and fusion of multi-dimensional features through longitudinal residual connections; in the downsampling stage, feature dimensionality reduction processing is performed. Through layer-by-layer convolution and pooling operations, the spatial resolution of the signal gradually decreases, while the number of channels increases, which helps to extract deep features in the signal. In the convolution process of each layer, the model adds the input features to the convolution features through residual connections to ensure that information is not lost while accelerating model training; in the upsampling stage, the spatial resolution of the feature map is gradually restored through transposed convolution or interpolation operations. Residual connections are also used in the upsampling process of each layer, and the high-resolution features extracted in the downsampling process are fused with the deep features restored in the upsampling process, so that the model can capture global information and local detail information at the same time, reduce information loss in the downsampling process, generate more accurate feature representation, thereby achieving the purpose of signal denoising and ensuring effective information transmission and fusion between multi-scale features.

[0104] At each time step t, the network predicts the noise component in the current covariance matrix and restores it to the covariance matrix of the previous moment using the conditional probability model. Specifically, the denoising network receives the current time step t and the noisy matrix as input, predicts the noise ∈, and thus calculates the estimated value of the noise-free matrix. By iterating this process, the noise-reduced covariance matrix is ​​finally restored. The backward recovery process satisfies:

[0105]

[0106] Among them, x t represents the data after t times of noise addition, x t-1 Represents the data after back diffusion, that is, the denoised data of the previous time step t-1, ∈ θ (x t ,t) represents the output of the noise prediction network, σt is the standard deviation of time step t, which is used to control the noise component of random sampling, It means that from the initial state x0 to x t The cumulative noise intensity added, z represents independent and identically distributed Gaussian noise.

[0107] In this module, the signal covariance matrix with interference and the ideal covariance matrix are input into diffusion for noise addition and denoising processing, so that the network has the ability to reconstruct the signal covariance matrix with interference, thereby achieving noise reduction.

[0108] (2-2) The complex attention mechanism module includes a parallel complex operation unit and a channel attention unit. The outputs of the two are aggregated and input into the self-attention unit to realize the extraction and fusion of multi-dimensional features of the signal.

[0109] Although attempts have been made to use signal features such as amplitude or phase as input for DOA estimation and beamforming, different signal features (such as the real part and the imaginary part) are often processed as independent input channels during feature processing. This method, to a certain extent, severs the mathematical logical relationship and physical consistency between signal features, and ignores their essential complementarity and inherent synergy. This feature separation makes it difficult for the network to establish a close association between features during feature learning, thereby affecting the model's comprehensive representation capability and reasoning accuracy for the target signal.

[0110] In order to solve the above problems, the present invention models and learns the mathematical logical relationship of multiple features such as the amplitude, phase, real part and imaginary part of the signal;

[0111] (2-2-1) Constructing complex number operation units

[0112] In order to ensure the mathematical logic relationship and physical consistency between signal features and simulate their essential complementarity and inherent synergy, a complex operation unit is proposed. The complex operation unit includes a complex convolution layer, a complex batch normalization layer and a complex activation function connected in sequence.

[0113] Complex convolution CC (complex-convolution) is the core component of the complex operation unit network, which is designed to effectively process and extract the spatial and phase features in the complex signal, and is used to process complex input data. Unlike real convolution, which only processes real features, complex convolution processes both the real and imaginary parts of the signal, and can capture the phase information and amplitude changes in the signal, thereby more comprehensively describing the signal characteristics; the input and convolution kernel of the complex convolution can be written as:

[0114] Input: V = V r +jV i , V r Represents the real part of the input signal, Vi Represents the imaginary part of the input signal;

[0115] Convolution: K = K r +jK i , K r represents the real part of the convolution kernel, K i Represents the imaginary part of the convolution kernel;

[0116] The complex convolution calculation formula is as follows:

[0117] K*V=(K t +jK i )*(V t +JV i )=(K t *V t -K i *V i )+j(K t *V i +K i *V t )

[0118] Complex Batch Normalization (CBN) effectively improves the training efficiency and stability of the network by normalizing the real and imaginary parts of the complex signal respectively, while retaining the phase information of the signal and enhancing the accuracy and robustness of DOA estimation, which can be expressed as:

[0119]

[0120] Among them, γ is a scaling parameter, which is used to adjust the scale and direction of the data to ensure that the normalized data adapts to the specific needs of the model. β is a translation parameter, which is used to adjust the position of the data so that its mean can be flexibly moved to improve the expressiveness and training effect of the model.

[0121] The complex activation function CAF (Complex Activation Function) is a key component for introducing nonlinear transformation. Different from the activation function in the real network, CAF needs to process complex input, retain the phase information of the signal, and enhance the expression ability of the network. In this embodiment, modReLU is used to meet the following requirements:

[0122]

[0123] Where z represents the input signal, θ z is the phase of z, and b is a learnable parameter. Since |z| is non-negative, parameter b allows the activation function to reach the dead zone.

[0124] The complex operation unit retains the intrinsic mathematical operation logic of the real and imaginary parts of the signal, realizing efficient feature extraction of complex signals.

[0125] (2-2-2) Constructing channel attention unit

[0126] For the amplitude and phase channel dimensions, the channel attention mechanism SE (Squeeze-and-Excitation) squeeze-excitation module is used, which is divided into three steps: squeeze compression operation, excitation excitation operation and reweight weight operation.

[0127] The compression operation performs global feature compression through the spatial dimension. Specifically, using Global Average Pooling, each two-dimensional feature map is mapped to a scalar, and the global response strength of the feature channel in the entire spatial range is extracted; not only does it greatly reduce the redundant information of the features, but it also enables each feature channel to effectively characterize the global statistical characteristics, thereby improving the model's ability to capture key information. Through this mechanism, the phase and amplitude features are more deeply understood with the support of the global receptive field, especially in complex signal scenarios (such as low signal-to-noise ratio environments), which can significantly enhance the model's expression and utilization of important features;

[0128] The excitation operation generates adaptive weights for each channel through a network consisting of two fully connected layers. These weights are limited to the range of [0,1] through the Sigmoid activation function to measure the importance of each channel and perform weighted fusion with the corresponding feature channels. Its working mechanism is similar to the "gate mechanism" in recurrent neural networks. It explicitly models the correlation between channels through learnable parameters w, thereby accurately identifying and strengthening key feature channels, while appropriately suppressing the contribution of secondary channels and optimizing the network's feature expression capabilities.

[0129] The weight operation regards the weight of the excitation output as the importance of each feature channel after feature selection, and then weights it to the previous features channel by channel through multiplication to complete the recalibration of the original features in the channel dimension.

[0130] (2-2-3) Fusion of multi-dimensional features based on self-attention unit

[0131] In order to achieve efficient fusion of multi-dimensional signal features, the self-attention mechanism is introduced to capture the long-range dependency and global correlation between features by calculating the internal correlation of the input features. Its core is to dynamically adjust the weights between features to highlight key features and suppress redundant information.

[0132] The attention mechanism satisfies:

[0133]

[0134] Among them, Q is Query, which means that the current feature "raises a question", K is Key, which is used to "match the question", and V is Value, which provides the actual information of the feature. The input features are mapped to Query, Key and Value representations through different linear transformations, and then the dot product of Query and Key is calculated and a weight matrix is ​​generated through the Softmax operation to measure the correlation between each feature and other features. Finally, the weight matrix is ​​used to perform weighted summation on the Value features to obtain the fused feature representation.

[0135] In the present invention, the real and imaginary features extracted by the complex convolutional network and the features enhanced by the amplitude-phase channel attention mechanism are first spliced ​​along the channel dimension through the Concat operation to form a multidimensional comprehensive feature representation. The spliced ​​feature matrix contains a variety of feature dimension information of the signal, and then the spliced ​​features are further aggregated and optimized based on the Query-Key-Value structure of the self-attention mechanism. By dynamically allocating weights, the self-attention mechanism establishes a strong correlation between multidimensional features, highlights important features, and optimizes global feature expression, thereby achieving deep fusion of multidimensional features. The network of the present invention effectively retains key information such as the amplitude, phase, real part and imaginary part of the signal, improves the expression ability and robustness of the network, and provides a more accurate feature basis for signal processing tasks.

[0136] (3) Training the improved denoising and classification network

[0137] A loss function of the improved denoising and classification network is constructed, and the loss function is associated with the error between the denoising output of the diffusion-based denoising module and the real signal, and the distance between the classification prediction distribution probability of the training sample and the real distribution.

[0138] Aiming at the problem that signal features are difficult to extract and classify accurately in a low signal-to-noise ratio environment, a deep learning framework for joint denoising and classification is designed. The denoising module restores the noisy signal to a clean signal through learning, minimizes the interference of noise on signal features, and the multiple attention mechanism module performs DOA classification on the signal.

[0139] To optimize these two subtasks, the denoising part uses the mean square error loss function (RMSE) in the joint loss function. The denoising effect is evaluated by measuring the error between the denoising network output and the true clean signal. The classification part uses the categorical cross-entropy loss function (Categorical Cross-Entropy Loss). The predicted probability of the DOA classification label is evaluated, and the denoising loss and classification loss are combined in a weighted manner. The weight coefficient controls the optimization proportion of the two subtasks respectively, ensuring that the model improves the classification accuracy while taking into account the denoising ability.

[0140] For the loss function of the denoising part, the root mean square error between the predicted value and the true value is calculated by RMSE, which is used to measure the average error between the predicted value and the true value. The expression is defined as follows:

[0141]

[0142] Where N is the number of samples, y i is the true value of the i-th sample, is the predicted value of the ith sample.

[0143] For the loss function of the classification part, the distance between the predicted distribution and the true distribution is measured by the multi-classification cross entropy loss function. When the probability predicted by the model is close to the true label, the loss value will be very small. The multi-classification cross entropy loss function satisfies:

[0144]

[0145] Among them, N is the number of samples, C is the number of categories, and y ij represents the true label of the i-th sample in the j-th category (one-hot encoding, only the correct category is 1, and other categories are 0), Represents the predicted probability of the i-th sample in the j-th category (obtained by the sigmoid function).

[0146] This framework establishes a collaborative relationship between the two subtasks by using the reconstructed signal output by the denoising network as the input of the classification network. The denoising network effectively eliminates the noise in the signal and improves the input quality of the classification network; the classification network accurately identifies the DOA of the signal, providing an efficient and robust solution for signal processing in complex signal environments.

[0147] Experiments show that the jointly trained model exhibits excellent signal reconstruction capability and classification accuracy under low signal-to-noise ratio conditions. The joint loss function satisfies:

[0148] L=α·L1+β·L2

[0149] Among them, α and β are weight coefficients for adjusting the two errors, which are used to balance the importance of denoising loss L1 and classification loss L2 in the total loss. In the specific implementation process, α is assigned to 0.8 and β is assigned to 0.2.

[0150] (4) Obtain test signal data and input it into the trained model to obtain a direction estimate.

[0151] like Figure 3 The figure shows the experimental effect comparison between the present invention and the classical DOA algorithm. 2,1 -SVD and CNN algorithms have RMSE close to Cramer-Rao Law Bound (CRLB), which is better than other algorithms such as MUSIC, ESPRIT, R-MUSIC and UnESPRIT. 2,1 -SVD, our method does not require tuning of any kind of parameters, which is a major advantage in practical applications. And compared to deep learning CNN, our method has a better average RMSE than CNN. The results also show that in high signal-to-noise ratio (SNR) cases, the RMSE of all grid-based methods has a "floor effect", that is, the error cannot be further reduced, and only grid-independent estimation methods (such as ESPRIT, R-MUSIC and UnESPRIT) can reach the Cramer-Rao lower bound. This performance is consistent for all grid-based methods and cannot be further improved unless a finer grid is used. In this experiment, although our method was not trained in high SNR scenarios, it was still able to predict sufficiently accurate angle estimates.

[0152] From the overall performance point of view, the performance of this method is significantly better than the classical algorithm at low signal-to-noise ratio levels, reflecting its superiority in complex environments. This advantage is attributed to the following points:

[0153] First, the denoising-denoising mechanism based on the diffusion model can effectively simulate the complex signal characteristics under low signal-to-noise ratio conditions, and restore key information through the denoising process, thereby improving the signal quality;

[0154] Secondly, the complex convolutional network (CCNN) is used to jointly model the real and imaginary parts of the signal, which fully retains the amplitude and phase information of the signal and enhances the feature extraction capability;

[0155] In addition, the importance of features is dynamically adjusted through the channel attention mechanism of amplitude and phase, and the self-attention mechanism (QKV) is introduced in the multi-dimensional feature fusion stage to achieve efficient aggregation of multi-dimensional features, thereby enhancing the representation ability of complex signals;

[0156] Finally, the classification network is combined with multi-classification cross entropy loss and coordinated with the denoising network to ensure the robustness and accuracy of the overall framework under various complex conditions.

[0157] These designs together provide a significant performance advantage for this method in complex environments.

[0158] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0159] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A denoising and optimized DOA estimation method under non-ideal channel conditions, characterized by: The method comprises the following steps: S1 constructs a simulation data set simulating non-ideal channel conditions; S2 builds an improved denoising and classification network; S3 training the improved denoising and classification network; S4 obtains test signal data, inputs it into the trained model, and obtains a direction estimate.

2. The method for DOA estimation with denoising optimization under non-ideal channel conditions according to claim 1, characterized in that: S1 includes the following steps: S1.1 Determine the DOA range and generate all non-overlapping DOA combinations based on the maximum number of targets, number of antennas and resolution input; S1.2 Calculate the signal steering vector based on the generated DOA combination, the antenna array structure and the set resolution; S1.3 applying different types of random perturbations to the antenna array to perform data augmentation, and constructing several training sets with different types of perturbations; S1.4 Assign a unique position index to any data in the training set.

3. The method for DOA estimation with denoising optimization under non-ideal channel conditions according to claim 2, characterized in that: The disturbance type includes one or more of position disturbance, gain disturbance, phase disturbance, mutual coupling disturbance, and nonlinear effect disturbance.

4. The method for DOA estimation with denoising optimization under non-ideal channel conditions according to claim 3, characterized in that: At least one of the training sets does not include a signal for data augmentation with random perturbations.

5. The method for DOA estimation with denoising optimization under non-ideal channel conditions according to claim 1, characterized in that: The improved denoising and classification network includes a diffusion-based denoising module and a complex attention mechanism module arranged in sequence; The diffusion-based denoising module includes a forward denoising unit and a backward denoising unit.

6. The method for DOA estimation with denoising optimization under non-ideal channel conditions according to claim 5, characterized in that: The backward denoising unit includes N downsampling layers and N upsampling layers arranged in sequence, and the nth downsampling layer is jump-connected to the nth upsampling layer from the end.

7. The method for DOA estimation with denoising optimization under non-ideal channel conditions according to claim 5, characterized in that: The complex attention mechanism module includes a parallel complex operation unit and a channel attention unit, and the outputs of the two are aggregated and input into the self-attention unit.

8. The method for DOA estimation with denoising optimization under non-ideal channel conditions according to claim 7, characterized in that: The complex operation unit includes a complex convolution layer, a complex batch normalization layer and a complex activation function connected in sequence.

9. The method for DOA estimation with denoising optimization under non-ideal channel conditions according to claim 7, characterized in that: A loss function of the improved denoising and classification network is constructed, and the loss function is associated with the error between the denoising output of the diffusion-based denoising module and the real signal, and the distance between the classification prediction distribution probability of the training sample and the real distribution.

Images