Non-ideal array DOA estimation method based on multi-scale weight distribution and transfer learning and electronic equipment

Abstract

The invention relates to a non-ideal array DOA estimation method based on multi-scale weight distribution and transfer learning and electronic equipment, and the method comprises the steps: constructing a fine tuning data set under a non-ideal array condition, the fine tuning data set comprising array signal data simulating various array physical defects; constructing an improved denoising and classification network, and training the improved denoising and classification network by adopting a transfer learning strategy; inputting a test signal in a non-ideal array environment into the trained improved denoising and classification network, and performing DOA estimation; the electronic equipment is realized based on the method. According to the method, the calculation overhead is reduced, the model is more robust to the interference of array errors and noise, the feature expression ability of signals is improved, the robustness and precision of the model in a complex environment are improved, transfer learning enables the model to obtain better performance in a low signal-to-noise ratio environment through a parameter sharing mode, and the method is suitable for large-scale popularization and application. And new non-ideal array data can be quickly adapted, and the retraining time and computing resources are reduced.

Description

A method for estimating the DOA of a nonideal array based on multi-scale weight allocation and transfer learning, and an electronic device thereof. Technical Field

[0001] This invention relates to the field of signal processing, and in particular to a non-ideal array DOA estimation method and electronic device based on multi-scale weight allocation and transfer learning related to direction of arrival (DOA) estimation. Background Technology

[0002] In modern sonar detection, DOA estimation is widely used to infer the direction of signal origin. Traditional DOA estimation methods, such as MUSIC and ESPRIT, typically rely on ideal antenna array models (e.g., uniform linear arrays like ULA), assuming that the array elements have consistent positions, amplitudes, and phases. However, in practical applications, due to array installation errors, inconsistencies in antenna elements, and environmental factors, arrays often suffer from positional offsets and gain / phase inconsistencies. These non-ideal array conditions significantly reduce the estimation accuracy of traditional methods.

[0003] Furthermore, with the increasing complexity of wireless communication environments, signal propagation is affected by various factors such as multipath effects, interference, and noise, which further degrades the performance of traditional DOA estimation algorithms under low signal-to-noise ratio (SNR) conditions. Therefore, designing a DOA estimation method that can adapt to non-ideal array conditions and maintain high estimation accuracy under low SNR environments has become a significant challenge in the field of signal processing.

[0004] Although deep learning techniques have demonstrated good performance in DOA estimation, especially in robustness under low signal-to-noise ratio conditions, existing deep learning methods often ignore or simplify non-ideal array factors. This leads to the model failing to fully learn array error characteristics, thus affecting its generalization ability in complex real-world environments. Therefore, a new approach is needed to improve the model's adaptability to non-ideal array conditions. Summary of the Invention

[0005] This invention proposes a non-ideal array DOA estimation method and electronic device based on multi-scale weight allocation and transfer learning. It can not only effectively improve the DOA estimation accuracy in low signal-to-noise ratio environments, but also maintain strong robustness and generalization ability in the presence of array errors (such as array element position offset, amplitude and phase offset, etc.).

[0006] The technical solution adopted in this invention is a non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning, the method comprising the following steps:

[0007] S1 Constructs a fine-tuning dataset under non-ideal array conditions, the fine-tuning dataset containing array signal data simulating various array physical defects;

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

[0009] S3 uses a transfer learning strategy to train the improved denoising and classification network;

[0010] S4 inputs the test signal under a non-ideal array environment into the trained improved denoising and classification network to perform DOA estimation.

[0011] Preferably, in S1, the simulated array physical defects include one or more of antenna position perturbation, gain perturbation, phase perturbation, mutual coupling perturbation, and nonlinear effect perturbation; the fine-tuning dataset includes multiple sub-training sets composed of different combinations of defects.

[0012] Preferably, in S2, the improved denoising and classification network includes a denoising module based on a diffusion model and a complex attention mechanism module;

[0013] The complex attention mechanism module includes a complex operation unit and a multi-scale convolutional channel attention unit. The complex operation unit is used to extract joint features from the real and imaginary parts of the input signal, and the multi-scale convolutional channel attention unit is used to perform cross-channel interaction and weight recalibration on the channel features.

[0014] Preferably, the multi-scale convolutional channel attention unit includes a global average pooling layer, a multi-scale convolutional module, an activation function, and an output layer.

[0015] Preferably, the multi-scale convolution module achieves multi-scale feature extraction through parallel one-dimensional convolutional layers, wherein the kernel sizes of the one-dimensional convolutional layers are different from each other.

[0016] Preferably, the multi-scale convolutional channel attention unit further includes a self-attention unit for performing feature fusion on the outputs of the complex operation unit and the multi-scale convolutional channel attention unit.

[0017] Preferably, in S3, the network is pre-trained on an ideal or general dataset to obtain a base model, and then the base model is adjusted using the fine-tuning dataset constructed in S1.

[0018] Preferably, the adjustment involves freezing the network parameters of the feature extraction part of the base model, which consists of a noise reduction module and a complex attention mechanism module, and updating only the parameters of the fully connected layers used for DOA classification in the network.

[0019] Preferably, the total loss function for training the network is correlated with the loss that measures the error between the denoised output and the real signal, and the loss that measures the accuracy of DOA classification prediction.

[0020] An electronic device includes 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 implements the aforementioned non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning.

[0021] This invention relates to a non-ideal array DOA estimation method and electronic device based on multi-scale weight allocation and transfer learning. The method involves constructing a fine-tuning dataset under non-ideal array conditions, which includes array signal data simulating various array physical defects; constructing an improved denoising and classification network, and training the improved denoising and classification network using a transfer learning strategy; inputting test signals from the non-ideal array environment into the trained improved denoising and classification network to perform DOA estimation; and implementing an electronic device based on the method.

[0022] The beneficial effects of this invention are as follows:

[0023] (1) The multi-scale convolutional channel attention mechanism enhances feature extraction capability, reduces computational overhead, and makes the model more robust to array errors and noise interference;

[0024] (2) To address the problem of insufficient signal feature expression capability under low signal-to-noise ratio environment, a noise addition-denoising scheme based on diffusion model is proposed to enhance and optimize the data, suppress the influence of noise, and improve the signal feature expression capability.

[0025] (3) To address the problem of fragmented multidimensional features of the reconstructed signal, including amplitude, phase, real part, and imaginary part, a deep learning framework with complex attention mechanism for multidimensional feature fusion was designed to achieve efficient integration of multidimensional features, thus solving the problem of fragmented multidimensional features of the signal and improving the robustness and accuracy of the model in complex environments.

[0026] (4) Since the scale and quality of training data are limited under non-ideal array conditions, transfer learning can enable the model to achieve better performance in low signal-to-noise ratio environments by sharing parameters, and can quickly adapt to new non-ideal array data, reducing the time and computational resources for retraining. Attached Figure Description

[0027] Figure 1 is a flowchart of the method of the present invention;

[0028] Figure 2 is a structural diagram of the improved denoising and classification network of the present invention;

[0029] Figure 3 is a structural diagram of the feature extraction part in the multi-scale convolutional channel attention unit of the present invention;

[0030] Figure 4 is a schematic diagram of transfer learning in this invention;

[0031] Figure 5 shows the DOA estimation performance of the non-ideal array random angle in the experimental example of the present invention;

[0032] Figure 6 shows the DOA estimation error of the random angle of the non-ideal array in the experimental example of the present invention;

[0033] Figure 7 shows the RMSE performance of each algorithm under different signal-to-noise ratio conditions;

[0034] Figure 8 shows the RMSE performance of each algorithm under different snapshot count conditions;

[0035] Figure 9 shows the RMSE performance of each algorithm at different angle intervals. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to embodiments, but the scope of protection of the present invention is not limited thereto.

[0037] This invention relates to a method for estimating the DOA of a non-ideal array based on multi-scale weight allocation and transfer learning, the method comprising the following steps:

[0038] (1) Construct a fine-tuning dataset under non-ideal array conditions, wherein the fine-tuning dataset contains array signal data simulating various array physical defects;

[0039] (2) Construct an improved denoising and classification network;

[0040] (3) The improved denoising and classification network is trained using a transfer learning strategy;

[0041] (4) Input the test signal under non-ideal array environment into the trained improved denoising and classification network to perform DOA estimation.

[0042] The steps are explained below with reference to specific implementation methods.

[0043] (1) Construct a fine-tuning dataset under non-ideal array conditions, wherein the fine-tuning dataset contains array signal data simulating various array physical defects;

[0044] The simulated array physical defects include one or more of the following: antenna position perturbation, gain perturbation, phase perturbation, mutual coupling perturbation, and nonlinear effect perturbation; the fine-tuning dataset includes multiple sub-training sets composed of different combinations of defects.

[0045] In this invention, the array model is ULA. By utilizing the characteristic that there is a fixed phase difference in the array signal, a phase difference is added to the signal to simulate the original data arriving at the array. Then, by simulating factors such as antenna position perturbation, gain / phase inconsistency, mutual coupling effect and nonlinear effect, multiple training sets with different physical conditions are generated.

[0046] After generating the signal and various interferences, normalization processing is performed; by standardizing the energy of the signal, the power of each signal is ensured to be within the same range, so as to maintain the consistency of different signals during training.

[0047] At least one training set does not include signals that have been augmented with random perturbations.

[0048] In practical applications, the training set includes:

[0049] Training set A: Contains only array signals, without any disturbances or nonlinear effects;

[0050] Training set B: contains a combination of array signals and position perturbation signals;

[0051] Training set C: contains a combination of array signals and gain perturbation signals;

[0052] Training set D: contains a combination of array signals and phase perturbation signals;

[0053] Training set E: contains a combination of array signals and mutually coupled perturbation signals;

[0054] Training set F: contains a combination of array signals and nonlinear effect signals;

[0055] Training set G: contains a combination of position perturbation, gain perturbation, phase perturbation, mutual coupling perturbation and nonlinear effect signals (all perturbation factors exist simultaneously).

[0056] Assign a unique orientation index to any data point in the training set for subsequent classification tasks.

[0057] (2) Construct an improved denoising and classification network;

[0058] The improved denoising and classification network includes a denoising module based on a diffusion model and a complex attention mechanism module;

[0059] Among them, the denoising module based on the diffusion model can refer to the denoising module that includes a forward noise addition unit and a backward noise removal unit;

[0060] The complex attention mechanism module includes a complex operation unit and a multi-scale convolutional channel attention unit. The complex operation unit is used to extract joint features from the real and imaginary parts of the input signal, including sequentially connected complex-convolution layers, complex batch normalization layers, and complex activation functions. The multi-scale convolutional channel attention unit is used to perform cross-channel interaction and weight recalibration of channel features.

[0061] The multi-scale convolutional channel attention unit includes a global average pooling layer, a multi-scale convolutional module, an activation function, and an output layer. For the amplitude and phase channel dimensions, multi-scale convolutional channel attention (MSCCA) is used, which consists of three steps: GAP (global average pooling operation), multi-scale convolution extraction, and weight adjustment.

[0062] The multi-scale convolution module achieves multi-scale feature extraction through parallel one-dimensional convolutional layers, and the kernel sizes of the one-dimensional convolutional layers are different from each other.

[0063] Global average pooling (GAP) performs a global aggregation operation on the spatial dimension of each channel to obtain a scalar that characterizes the overall statistical properties of that channel, expressed as:

[0064]

[0065] in, The compression function is essentially a global average pooling operation, representing the original channel feature map. Feature maps compressed to 1×1 pixels in width and height ;

[0066] The main improvement of multi-scale convolution extraction for channel attention SE lies in replacing the original compressed excitation with parallel multi-scale one-dimensional convolution for cross-channel interaction. This enhances the feature joint between channels while reducing the model's complexity. Parallel multi-scale one-dimensional convolution is introduced during the channel attention calculation process. Cross-channel interaction is achieved by using one-dimensional convolution with different receptive fields;

[0067] Specifically, this embodiment uses three different sizes of convolution kernels. , and To enhance the capture of features at different scales, the calculation is as follows:

[0068]

[0069] vector Input to activation function By introducing nonlinear factors, the corresponding weighting factors are finally obtained. ,

[0070]

[0071] By using small convolutional kernels to capture local details and large convolutional kernels to capture global semantics, feature maps of different scales are stacked in the channel direction;

[0072] The weighting operation treats the weights of the multi-scale convolution output as the importance of each feature channel after feature selection, and assigns weight factors accordingly. With each channel feature map Multiplication is used to generate a weighted feature map, i.e.

[0073]

[0074] in, This represents the newly generated feature map. This indicates the feature map of each channel. Multiply by the weight of the corresponding channel That is, weighting the features of each channel to enhance the representation of the target features.

[0075] The multi-scale convolutional channel attention unit further includes a self-attention unit for performing feature fusion on the outputs of the complex operation unit and the multi-scale convolutional channel attention unit.

[0076] In this invention, the real and imaginary features extracted by a complex convolutional network, along with features enhanced by a multi-scale convolutional channel attention mechanism, are first concatenated along the channel dimension using a Concat operation to form a multi-dimensional comprehensive feature representation. The concatenated feature matrix contains information from multiple feature dimensions of the signal. Subsequently, a Query-Key-Value structure based on a self-attention mechanism further aggregates and optimizes the concatenated features. By dynamically allocating weights, the self-attention mechanism establishes strong correlations between multi-dimensional features, highlighting important features and optimizing global feature expression, thereby achieving deep fusion of multi-dimensional features. The network of this invention effectively preserves key information such as the amplitude, phase, real and imaginary parts of the signal, improving the network's expressive power and robustness, and providing a more accurate feature foundation for signal processing tasks.

[0077] (3) The improved denoising and classification network is trained using a transfer learning strategy;

[0078] The network is pre-trained on an ideal or general dataset to obtain a base model, which is then tuned using the fine-tuning dataset constructed in (1).

[0079] The adjustment involves freezing the network parameters of the feature extraction part of the base model, which consists of a noise reduction module and a complex attention mechanism module, and updating only the parameters of the fully connected layers used for DOA classification in the network.

[0080] In this invention, the weights of the fully connected layers are fine-tuned. During the fine-tuning process, the parameters of the feature extraction module (including PLR, MSCRC64, MSCRC128, etc.) of the pre-trained model remain fixed to preserve the general feature representation obtained from the source domain. Only the weights of the fully connected layers in the direction prediction module are updated. These parameters are adjusted during the backpropagation of the target domain data, thereby avoiding overfitting or "forgetting" problems caused by large-scale updates. Due to the significant reduction in training parameters, the training speed and stability during the fine-tuning stage are improved, enabling the method to achieve good model adaptation even with limited data.

[0081] The total loss function for training the network is correlated with the loss that measures the error between the denoised output and the real signal, and the loss that measures the accuracy of DOA classification prediction.

[0082] For the loss function in the denoising part, the root mean square error (RMSE) is calculated using RMSE to measure the average error between the predicted and true values. The expression is defined as follows.

[0083]

[0084] in, It is the sample size. It is the first The true value of each sample It is the first The predicted value for each sample.

[0085] For the loss function in the classification part, the multi-class cross-entropy loss function measures the distance between the predicted distribution and the true distribution. When the probability predicted by the model is close to the true label, the loss value will be small. The multi-class cross-entropy loss function satisfies:

[0086]

[0087] Where N is the number of samples and C is the number of categories. This represents the true label of the i-th sample in the j-th category (one-hot encoded, with 1 for the correct category and 0 for the other categories). This represents the predicted probability of the i-th sample in the j-th category (obtained through the sigmoid function).

[0088] This framework establishes a collaborative relationship between the two sub-tasks by using the reconstructed signal output by the denoising network as the input to the classification network. The denoising network effectively eliminates noise in the signal, improving the input quality of the classification network; while the classification network accurately identifies the DOA of the signal, providing an efficient and robust solution for signal processing in complex signal environments.

[0089] 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:

[0090] 2

[0091] Here, α and β are weighting coefficients that adjust the two types of errors, used to balance the importance of denoising loss L1 and classification loss L2 in the total loss. In the specific implementation, α is assigned a value of 0.8 and β is assigned a value of 0.2.

[0092] (4) Input the test signal under non-ideal array environment into the trained improved denoising and classification network to perform DOA estimation.

[0093] To test the performance of the network of this invention, a test was conducted on a DOA estimation scenario at random angles; a fixed angle difference between the two signal sources was set as follows: The signal-to-noise ratio is 0 dB, and the number of snapshots is 600. The DOA of the first signal source increases from -60° to 57° (in 1° increments), and the second signal source increases the DOA based on this. Furthermore, the standard deviations of position offset, amplitude offset, and phase offset were all set to 0.1 to simulate irrational array conditions. As shown in Figure 5, the vertical axis represents the angle, and the horizontal axis represents the sample index. The solid line represents the true angle, while the hollow marked points reflect the estimated values ​​of the DOA of the incoming wave direction by each method. As shown in Figure 6, the vertical axis shows the error between the estimated angle and the true value, and the horizontal axis is still the sample index; the closer the error value is to 0, the better the estimation effect of the algorithm. Compared with other comparative methods, the method of this invention (MS-WANet-T) has a more uniform overall error distribution, avoids the occurrence of large error points, and exhibits superior performance.

[0094] Figure 7 shows the RMSE performance under different signal-to-noise ratio conditions compared with traditional methods and models without transfer mechanisms. Especially in the high SNR range, other methods show performance bottlenecks, while MS-WANet-T, with its pre-training fine-tuning strategy, demonstrates stronger robustness and generalization ability across the entire SNR range.

[0095] As shown in Figure 8, in terms of RMSE performance under different snapshot counts, the un-fine-tuned deep learning model, while exhibiting some stability, suffers from performance bottlenecks due to its inability to adapt to non-ideal array characteristics. In some cases, it even performs worse than the traditional algorithm UnESPRIT, demonstrating insufficient robustness. In contrast, the transfer learning model MS-WANet-T consistently maintains the lowest RMSE across the entire snapshot range, demonstrating stronger stability and generalization ability, effectively validating the advantages of transfer strategies in complex array environments.

[0096] Figure 9 shows the RMSE performance at different angular intervals. Traditional methods show a significant performance drop at small angles, with limited resolution. Deep learning methods are generally more stable, but the untuned model still has a performance bottleneck under array perturbations. MS-WANet-T significantly improves small-angle resolution through transfer learning, maintaining the lowest RMSE across the entire angular range, demonstrating stronger robustness and generalization.

[0097] Overall, the fine-tuned network of this invention (MS-WANet-T) significantly outperforms the classic algorithm. Its superior stability compared to the untuned MS-WANet demonstrates its excellence in complex environments. First, the diffusion-based noise-denoising mechanism effectively simulates complex signal features under low signal-to-noise ratio conditions and recovers key information through the denoising process, improving signal quality. Second, the joint modeling of the real and imaginary parts of the signal using a complex convolutional network fully preserves the amplitude and phase information, enhancing feature extraction capabilities. The multi-scale convolutional channel attention mechanism significantly reduces the number of parameters and computational overhead while effectively preserving and strengthening the direct correspondence and dependency between channels, thus improving overall performance. Furthermore, the introduction of a self-attention mechanism in the multi-dimensional feature fusion stage enables efficient aggregation of multi-dimensional features, enhancing the representation of complex signals. Additionally, the addition of irrational array data to the fine-tuning dataset enhances the model's adaptability under various non-ideal array conditions. The classification network, combined with multi-class cross-entropy loss, works in synergy with the denoising network to ensure the robustness and accuracy of the overall framework under various complex conditions.

[0098] The present invention also relates to an electronic device, including 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 implements the aforementioned non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning.

[0099] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0100] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for estimating the DOA of a non-ideal array based on multi-scale weight allocation and transfer learning, characterized in that: The method includes the following steps: S1 Constructing a fine-tuning dataset under non-ideal array conditions, the fine-tuning dataset containing array signal data simulating various array physical defects; S2 Constructing an improved denoising and classification network; S3 Training the improved denoising and classification network using a transfer learning strategy; S4 Inputting the test signal under the non-ideal array conditions into the trained improved denoising and classification network to perform DOA estimation.

2. The non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning according to claim 1, characterized in that: In S1, the simulated array physical defects include one or more of the following: antenna position perturbation, gain perturbation, phase perturbation, mutual coupling perturbation, and nonlinear effect perturbation; the fine-tuning dataset includes multiple sub-training sets composed of different combinations of defects.

3. The non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning according to claim 1, characterized in that: In S2, the improved denoising and classification network includes a denoising module based on a diffusion model and a complex attention mechanism module. The complex attention mechanism module includes a complex operation unit and a multi-scale convolutional channel attention unit. The complex operation unit is used to extract joint features from the real and imaginary parts of the input signal, and the multi-scale convolutional channel attention unit is used to perform cross-channel interaction and weight recalibration on the channel features.

4. The non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning according to claim 3, characterized in that: The multi-scale convolutional channel attention unit includes a global average pooling layer, a multi-scale convolutional module, an activation function, and an output layer.

5. The non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning according to claim 4, characterized in that: The multi-scale convolution module achieves multi-scale feature extraction through parallel one-dimensional convolutional layers, and the kernel sizes of the one-dimensional convolutional layers are different from each other.

6. The non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning according to claim 4, characterized in that: The multi-scale convolutional channel attention unit further includes a self-attention unit for performing feature fusion on the outputs of the complex operation unit and the multi-scale convolutional channel attention unit.

7. The non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning according to claim 1, characterized in that: In S3, the network is pre-trained on an ideal or general dataset to obtain a base model, which is then adjusted using the fine-tuning dataset constructed in S1.

8. The non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning according to claim 7, characterized in that: The adjustment involves freezing the network parameters of the feature extraction part of the base model, which consists of a noise reduction module and a complex attention mechanism module, and updating only the parameters of the fully connected layers used for DOA classification in the network.

9. The non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning according to claim 1, characterized in that: The total loss function for training the network is correlated with the loss that measures the error between the denoised output and the real signal, and the loss that measures the accuracy of DOA classification prediction.

10. An electronic device 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 implements the non-ideal array DOA estimation method based on multi-scale weight allocation and transfer learning as described in any one of claims 1 to 9.

Images