Direction of arrival estimation method of underwater acoustic signal based on improved Mamba network

Abstract

The present invention belongs to the field of underwater acoustic array signal processing, and specifically relates to a method for estimating the direction of arrival of underwater acoustic signals based on an improved Mamba network. The method comprises the following steps: 1) establishing a hydrophone array model, independently sampling source signals, calculating the signal's covariance matrix, separating the real and imaginary parts of the covariance matrix, and then normalizing the spliced ​​covariance matrix to serve as a dataset training model; 2) building a lightweight module based on the Mamba network, compressing the covariance matrix, adjusting the dimension, obtaining information correlation, mixing signals of all time steps, and outputting a characteristic spectrum of the signal; 3) processing the characteristic spectrum using a CBMA module, adjusting the dimension through a fully connected layer, and then using a sigmoid function to output an estimated value of the corresponding direction of arrival; and 4) cascading the models, training with a dataset, adjusting parameters, and using the trained model to estimate the direction of arrival of unknown signals. This method reduces computational complexity and is more adaptable to underwater environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater acoustic array signal processing, and in particular relates to an underwater acoustic signal direction-of-arrival estimation method based on an improved Mamba network. Background Art

[0002] Underwater acoustic communication uses water as a transmission medium to transmit information via sound waves. With relatively low attenuation and strong penetration, it is currently the only way to achieve long-distance underwater communication. It is widely used in underwater exploration, marine resource development, and other fields. Hydrophone arrays, capable of receiving and transmitting acoustic information from underwater equipment, are crucial for communication and positioning functions.

[0003] To further improve the positioning accuracy and communication quality of underwater equipment, direction-of-arrival (DOA) estimation technology is widely used. This technology uses a hydrophone array to analyze the angle of arrival of sound source signals to accurately estimate the target's position. However, due to the high noise interference and significant multipath effects of underwater acoustic channels, and the fact that underwater nodes are often battery-powered, their energy is limited. Conventional DOA estimation methods significantly degrade in performance, thus affecting the accuracy of underwater target positioning.

[0004] Therefore, a low-complexity and high-accuracy DOA estimation method is needed to achieve high-resolution underwater acoustic signal direction of arrival estimation. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide an underwater acoustic array signal direction of arrival (DOA) estimation method based on the Mamba network and CBAM module, which estimates the angle signal through regression and uses deep learning methods to perform comprehensive feature extraction of signal angle information, thereby solving the problems of poor versatility, low accuracy and high complexity of traditional methods.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a method for estimating the direction of arrival of underwater acoustic signals based on an improved Mamba network, comprising the following steps:

[0007] Step 1: Establish a hydrophone array model, sample the source signal independently, calculate the signal covariance matrix, separate the real and imaginary parts of the covariance matrix, and concatenate them in the vertical dimension to obtain a matrix of size 2M×M. Normalize the concatenated covariance matrix and use it as the dataset for training the model.

[0008] Step 2: Build a lightweight module based on the Mamba network to compress the normalized covariance matrix, adjust the dimension, obtain the correlation of information at different times, dynamically control the output weights through the gating mechanism, mix the signals of all time steps, capture the deep-level features of the signals, and output the characteristic spectrum of the signals;

[0009] Step 3: Process the feature spectrogram using the CBMA module to capture the correlation between deep features, pay attention to important features, adjust the dimension through the fully connected layer, and use the Sigmoid function to output the estimated value of the corresponding direction of arrival;

[0010] Step 4: Cascade the models of steps 2 and 3, and train them using the dataset in step 1. Adjust the parameters during the training process. After the model training is completed, use the trained model to estimate the direction of arrival of the unknown signal.

[0011] In the above-mentioned method for estimating the direction of arrival of underwater acoustic signals based on the improved Mamba network, in step 1, a hydrophone array model is established, and independent sampling of the source signal includes:

[0012] Step 1-1: Deploy a uniform linear hydrophone array with M elements vertically underwater. Each element corresponds to a single channel, and the element spacing is d and is less than half the signal wavelength.

[0013] Step 1-2: Calculate the delay difference based on the path difference D between the mth array element and the previous array element ,in, represents the speed of sound, ,in Represents the incident angle of all possible incoming wave directions;

[0014] Step 1-3: Based on the delay difference, the phase difference of the mth array element is obtained as: ,in, represents the wavelength of the carrier, represents the imaginary unit, represents the frequency of the signal, Indicates The incident angle in the direction of the incoming wave, e represents a natural constant;

[0015] Step 1-4: The signal received by the mth array element under the influence of underwater noise is expressed as: ,in, is the noise received by the mth array element, Indicates the current incoming wave direction. represents all possible directions of incoming waves in space, Indicates the A narrowband signal source;

[0016] Step 1-5: Represent the signal in step 1-4 using a matrix:

[0017] ,in, represents the signal matrix, represents the noise matrix, , represents the flow matrix of the signal, and T represents the time period.

[0018] The above-mentioned underwater acoustic signal direction of arrival estimation method based on the improved Mamba network calculates the formula of the signal covariance matrix as follows: ,in, represents the covariance matrix of the signal, represents the input matrix, It represents the conjugate transpose of the matrix. For a finite length signal, can be caused by Snapshot sampling at the moment To express it, the expression can be written as:

[0019] .

[0020] In the above-mentioned underwater acoustic signal direction of arrival estimation method based on the improved Mamba network, step 1 further includes:

[0021] Perform min-max normalization on the spliced ​​results. ,in, represents the signal after normalization, S represents the signal before normalization, and Indicates the maximum and minimum values ​​of the signal within the carrier period.

[0022] In the above-mentioned underwater acoustic signal direction of arrival estimation method based on the improved Mamba network, step 2 includes:

[0023] Step 2-1: Perform a 2M×1 one-dimensional convolution operation on the data in step 1, expanding the 2M×M matrix into a 1×M sequence. The expression for the one-dimensional convolution is: ,in, represents the input of the first convolutional layer, and Respectively represent the weight matrix and bias term of the t-th convolution kernel in the first convolution layer, Represents the output of the t-th convolution kernel in the first convolution layer, represents the activation function;

[0024] Step 2-2: Use RELU as the activation function of the one-dimensional convolutional layer. The expression of the RELU function is: ,in, represents the input signal;

[0025] Step 2-3: The matrix signal is flattened and then input into the Mamba model, which includes the state space model and the gating unit to synchronously process the signal and dynamically control the output weight through the gating mechanism;

[0026] Step 2-4: Use the Sigmoid function to normalize the gate value to the [0,1] interval. The expression of the Sigmoid function is:

[0027] ;

[0028] Steps 2-5: After repeating the above steps for the entire sequence, the information at all moments is merged. The information at each moment contains the state information of the previous moment, and the deep-level features of the signal are extracted.

[0029] In the above-mentioned underwater acoustic signal direction of arrival estimation method based on the improved Mamba network, steps 2-3 include:

[0030] Step a: Adjust the dimension of the matrix signal through a linear layer to obtain a one-dimensional sequence of length 32×M. Use layer normalization to reduce the offset of the internal variables of the model. The expression of layer normalization is as follows: ,in, represents the normalized output of the layer, represents the normalized input of the layer, and Represent the mean and standard deviation of the output of each neuron in the linear layer during the entire process, is a small constant;

[0031] Step b: The output sequence is expressed through the state space model, state update equation and output equation as follows: , where V is the state transfer matrix, used to control the change of state over time, B is the input mapping matrix, which maps the input signal to the state space, F is the output mapping matrix, which maps the state to the output space, and G is the direct transfer matrix, which is used to map the output signal directly to the output;

[0032] Step c: For the information at each moment, introduce the time mixing method, for the input ,in represents the batch size, represents the sequence length, represents the hidden dimension, and the temporal mixing is expressed as ,in, represents element-wise multiplication, represents a learnable time parameter, Represents a movement operation in time;

[0033] Step d: Integrated time step The time mixing output can be expressed as: ,The time step adopts an exponential decay method, the size can be learned, and is dynamically allocated by the optimizer to adaptively adjust each time step.

[0034] In the above-mentioned underwater acoustic signal direction of arrival estimation method based on the improved Mamba network, in step 3, the CBMA module includes a channel attention mechanism and a spatial attention mechanism, which extracts the input features in two stages by sequentially inferring the attention maps in the channel and spatial dimensions.

[0035] In the above-mentioned underwater acoustic signal direction of arrival estimation method based on the improved Mamba network, step 3 includes:

[0036] Step 3-1: The signal passes through the channel attention module, which uses the channel relationship between features to generate a channel attention map to capture the correlation between different channels, expressed as: ,in, represents the sigmoid activation function, and Represent the dimensionality reduction convolution and dimensionality increase convolution operations respectively, and Represents average pooling and maximum pooling operations respectively;

[0037] Step 3-2: The signal passes through the spatial attention module, which uses the spatial relationship between features to capture the importance of different spatial positions in the features, thereby extracting features in a targeted manner, expressed as: ,in, Indicates that a convolution operation is performed with a convolution kernel of size 7×7. Indicates that the results of average pooling and maximum pooling are spliced ​​in the channel dimension;

[0038] Step 3-3: Process the signal through three fully connected layers. The expression of the fully connected layer is: ,in, represents the input of the fully connected layer, represents the output of the fully connected layer, represents the weight matrix between two hidden layers, represents the weight matrix between the second hidden layer and the output layer, and represents the bias term, and represents the activation function;

[0039] Step 3-4: Use the Sigmoid function as the output layer to convert the arrival angle into a continuous distribution between [0, 1], and then perform denormalization to obtain the arrival direction estimation result.

[0040] In the above-mentioned underwater acoustic signal direction of arrival estimation method based on the improved Mamba network, step 4 includes:

[0041] Step 4-1: Label the signals according to different angles and use them as training sets to train the network model;

[0042] Step 4-2: During the training process, the mini-batch gradient descent method is used to train the neural network. The training set is divided into several groups of data. The neural network learns information from one group of data in each iteration.

[0043] Step 4-3: Select mean square error as the loss function and minimize the loss function as the training goal. The expression of mean square error MSE is:

[0044] ,in represents the number of samples, and Represent the true value and estimated value of the sample respectively;

[0045] Step 4-4: Iterate the network model multiple times, optimize the network model parameters based on the training objectives in step 4-3, save the model parameters as model weights, input the unlabeled signal received by the hydrophone array, and the model outputs the angle information of the wave to accurately estimate the direction of arrival.

[0046] Compared with traditional methods, the present invention's method for estimating the direction of arrival of underwater acoustic signals based on an improved Mamba network has the following advantages: Traditional signal subspace-based methods and beamforming methods require long spectral peak searches, have high algorithmic complexity, and are difficult to apply to energy-constrained underwater environments. The present invention's method uses deep learning to extract signal context features through a Mamba network and utilizes the attention mechanism of the CBAM module to focus on important features, reducing computational complexity. The trained model consumes only a small amount of computing resources during inference, making it more adaptable to underwater environments.

[0047] Traditional DOA estimation methods based on compressed sensing rely on grid resolution for accuracy, making it difficult to achieve super-resolution DOA estimation. This paper uses a neural network to construct a regression model, and the DOA estimation results output by the network are continuous and have higher resolution.

[0048] Underwater environments are subject to strong noise interference. Traditional methods are susceptible to noise interference under high signal-to-noise ratio conditions, resulting in large deviations between the estimated results and the true value. The proposed method, leveraging the powerful feature extraction capabilities of neural networks and the Mamba architecture, can accurately extract signal angle features under strong interference conditions with greater robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is an implementation scene graph in an embodiment of the present invention;

[0050] Figure 2 is a schematic diagram of a hydrophone array receiving signals in an embodiment of the present invention;

[0051] Figure 3 Schematic diagram of the network structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is described below in conjunction with specific implementation methods and drawings.

[0053] In an underwater wireless sensor network, a linear hydrophone array is used to receive underwater acoustic signals. The array received signals are modeled, and the covariance matrix of the array signals is calculated. After concatenation and preprocessing, the signals are used as network input. An improved deep learning model based on the Mamba network is designed to process the signal samples. This network, composed of multiple improved Mamba blocks, leverages Mamba's powerful feature extraction capabilities and low complexity to extract correlations between data. A state-space model and gating mechanism are used to process the signals and extract multi-dimensional features. The CBAM module then uses channel-by-channel and spatial attention mechanisms to capture key features and reduce computational complexity. The DOA estimation problem is transformed into a regression problem, and a sigmoid function is used to output continuous estimates of the corresponding angles, improving network resolution.

[0054] like Figure 1-3 As shown, a method for estimating the direction of arrival of underwater acoustic signals based on an improved Mamba network includes the following steps:

[0055] Step 1: Establish a hydrophone array model, sample the source signal independently, calculate the signal covariance matrix, separate the real and imaginary parts of the covariance matrix, concatenate them in the vertical dimension to obtain a 2M×M matrix, and normalize the processed covariance matrix to obtain the data training model;

[0056] Step 2: Build a lightweight module based on the Mamba network to compress the signal matrix, adjust the dimension, obtain the correlation of information at different times, dynamically control the output weights through the gating mechanism, mix the signals of all time steps, and capture the deep-level characteristics of the signal;

[0057] Step 3: The output of step 2 is processed using the CBMA module to capture the correlation between features and focus on important features. After adjusting the dimension through the fully connected layer, the sigmoid function is used to output the estimated value of the corresponding direction of arrival.

[0058] Step 4: Adjust the parameters to train the data training model in step 1, and use the trained model to estimate the direction of arrival of unknown signals.

[0059] Specifically, a method for estimating the direction of arrival of underwater acoustic signals based on an improved Mamba network is described. Figure 1 The application scenario of the present invention is in an underwater wireless sensor network, using a hydrophone array to receive signals sent by a target device and estimate the direction of arrival based on the angle of the incoming wave.

[0060] Reference Figure 2 , the embodiment process of the array receiving signal in the present invention is as follows:

[0061] Step 1: Build a hydrophone array model and obtain array reception data.

[0062] In underwater application scenarios, for ease of deployment, a uniform linear hydrophone array model is often used to receive the signal sent by the signal source. The number of array elements used is The hydrophone array has each array element corresponding to a single channel, and the array element spacing is and is smaller than half the wavelength of the signal. We assume that , ,…… direction, in the far field there is Individual and An irrelevant narrowband signal source is incident on the array element. There is a path difference between the array element and the previous array element , which can be expressed as:

[0063] ,

[0064] in Represents the incident angle of all possible incoming wave directions.

[0065] Based on this, the delay difference can be calculated: ,in Indicates the speed of sound.

[0066] Rule No. The phase difference of each array element can be expressed as:

[0067] ,

[0068] in, represents the wavelength of the carrier, represents the imaginary unit, represents the frequency of the signal, Indicates The incident angle in the direction of the incoming wave, e represents a natural constant. Considering the influence of underwater noise, such as Figure 2 As shown, yes The sub-element in Represents the Mth element in the array, then the The signal received by each array element can be expressed as:

[0069] ,

[0070] in, Indicates the current incoming wave direction. represents all possible directions of incoming waves in space, Indicates the A narrowband signal source, For the The receiving noise of each array element. Usually, for intuitive expression, the above formula is expressed in the form of a matrix:

[0071] ,

[0072] in, represents the signal matrix, represents the noise matrix, In the actual process, it is necessary to collect The sampling snapshot within a period of time, at which time the received signal Receive signal Contains original information such as signal waveform, modulation method, frequency, bandwidth, angle, etc.

[0073] Step 2: Calculate the covariance matrix of the array received signal and preprocess the covariance matrix.

[0074] For the received signal in step 1 , calculate its covariance matrix, the covariance matrix can be expressed as:

[0075] ,

[0076] in, represents the covariance matrix of the signal, represents the input matrix, It represents the conjugate transpose of the matrix. In practical applications, since the received signal is of finite length, can be caused by Snapshot sampling at the moment To express it, the expression can be written as:

[0077] ,

[0078] The above formula This is the complex covariance matrix. Compared to the original signal, it contains fewer dimensions of information and is easier to process. Is a complex matrix. In order to avoid feature loss and retain the integrity of the signal, Separate into real part and the imaginary part , whose expression is:

[0079] ,

[0080] ,

[0081] For an array of M elements, we get The real part matrix of and the imaginary part matrix , and concatenate them in the vertical dimension to obtain a matrix of size 2M×M. The matrix is ​​normalized and the calculation method is:

[0082] ,

[0083] in, represents the normalized signal. represents the signal before normalization, and Represents the maximum and minimum values ​​of the carrier signal, and the data obtained is used to train the network model.

[0084] Reference Figure 3 The embodiment process of Mamba network construction and training in the present invention is as follows:

[0085] Step 1: Build a network model and input the signal.

[0086] Specifically, the input data must first undergo a 2M×1 one-dimensional convolution operation to convert the size The matrix is ​​expanded into a 1×M sequence. The expression of one-dimensional convolution is:

[0087] ,

[0088] in, represents the input of the first convolutional layer, and Represent the weight matrix and bias term of the t-th convolution kernel in the first convolution layer, respectively. Represents the output of the tth convolution kernel in the first convolution layer, Represents the activation function. For the one-dimensional convolution layer, RELU is used as the activation function to reduce computational overhead. The expression of the RELU function is:

[0089] ,

[0090] in, Indicates the input signal.

[0091] The matrix signals are flattened and fed into the Mamba model, which includes a state-space model (SSM) and a gating unit to process the signals synchronously. The Mamba model provides an efficient method for processing sequence data by utilizing selective state space and structured state space components.

[0092] Specifically, the matrix signal is resized through a linear layer to obtain a one-dimensional sequence of length 32×M, and layer normalization is performed to reduce the offset of the internal variables of the model. The expression of layer normalization is as follows:

[0093] ,

[0094] in represents the normalized output of the layer, represents the normalized input of the layer, and Represent the mean and standard deviation of the output of each neuron in the linear layer during the entire process, is a small constant used to prevent the denominator from being zero.

[0095] The normalized output is processed using the Dropout layer to discard the output of some neurons to prevent the network from overfitting. The Dropout expression can be expressed as:

[0096] ,

[0097] Among them, x represents the signal received by the dropout layer, It is a binary mask, which has only two expressions: 0 and 1. 0 means discard, and 1 means keep. represents the dropout probability of each neuron.

[0098] The output sequence is passed through the state space model, and its basic form can be represented by the following state update equation and output equation:

[0099] ,

[0100] Where V is the state transition matrix, used to control the change of state over time. B is the input mapping matrix, mapping the input signal to the state space. F is the output mapping matrix, mapping the state to the output space. G is the direct transfer matrix, used to directly map the output signal to the output. Parameters A, B, and F are learnable and updated in real time by the network. The direct transfer matrix G is eliminated to reduce computational complexity and the number of parameters.

[0101] For the information at each moment, a time mixing method is introduced. ,in represents the batch size, represents the sequence length, represents the hidden dimension, and the time mixture can be expressed as

[0102] ,

[0103] represents element-wise multiplication, represents a learnable time parameter, Represents a movement operation in time. Considering the time step The time mixing output can be expressed as:

[0104] , e represents the natural number base, which is the natural number base e commonly used in mathematics, not a symbol;

[0105] The time step size is exponentially decayed. The size is learnable and dynamically allocated by the optimizer, adaptively adjusting each time step.

[0106] At each time step, the output weight is dynamically controlled by the gating mechanism. The gate value is normalized to the [0,1] interval using the Sigmoid function. The expression of the Sigmoid function is:

[0107] ,

[0108] Where e represents a natural constant. After repeating the above steps for the entire sequence, the information at all moments is merged. The information at each moment contains the state information of the previous moment, and the deep-level features of the signal are extracted. SiLU is introduced as an activation function to enhance the nonlinear representation ability of the model. The expression of SiLU is:

[0109] ,

[0110] After that, the signal is input into the CBAM module. The CBAM module mainly consists of a channel attention mechanism and a spatial attention mechanism. Through the channel attention module, the channel attention map is generated by utilizing the channel relationship between features and capturing the correlation between different channels. It can be expressed as:

[0111] ,

[0112] in, represents the sigmoid activation function, and Represent the dimensionality reduction convolution and dimensionality increase convolution operations respectively, and Represent the average pooling and maximum pooling operations respectively.

[0113] After that, the spatial attention module uses the spatial relationship between features to capture the importance of different spatial positions in the features, thereby extracting features in a targeted manner. This module can be expressed as:

[0114] ,

[0115] in, Indicates that a convolution operation is performed with a convolution kernel of size 7×7. This concatenates the results of average pooling and maximum pooling in the channel dimension. This allows for deep feature extraction of the signal without introducing additional computation, helping to reduce computational complexity.

[0116] Finally, the signal is processed through three fully connected layers. The expression of the fully connected layer is as follows

[0117] ,

[0118] in, represents the input of the fully connected layer, represents the output of the fully connected layer, represents the weight matrix between two hidden layers, represents the weight matrix between the second hidden layer and the output layer, and represents the bias term, and Represents the activation function.

[0119] The Sigmoid function is used as the output layer to convert the arrival angle into a continuous distribution between [0, 1], improving the network resolution. The arrival direction estimation result is then obtained through denormalization.

[0120] Step 2: Train the network and fine-tune it, saving the best model weights.

[0121] Specifically, the signals are labeled according to different angles and used as training sets for the network model. During training, the neural network is trained using the mini-batch gradient descent method. The training set is divided into several groups of data, and the neural network learns information from one group of data in each iteration. The mean squared error (MSE) is used as the loss function, and the training objective is to minimize the loss function. The expression of MSE is as follows:

[0122] ,

[0123] in, represents the number of samples, and represent the true value and estimated value of the sample respectively.

[0124] The network model is iterated multiple times, and the weights that most accurately estimate the direction of arrival are saved as the optimal model weights. After saving the trained model, the unlabeled signal received by the hydrophone array is input. The model outputs the angle information of the wave, which accurately estimates the direction of arrival.

[0125] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for a person skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to replace some of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions claimed to be protected by the present invention.

Claims

1. A method for estimating the direction of arrival of underwater acoustic signals based on an improved Mamba network, characterized in that: The following steps are involved: Step 1: Establish a hydrophone array model, independently sample the source signal, calculate the signal covariance matrix, separate the real and imaginary parts of the covariance matrix, and splice it in the vertical dimension to obtain a matrix of size 2M×M. The spliced ​​covariance matrix is ​​normalized as a data set training model. Establishing a hydrophone array model and independently sampling the source signal includes: Step 1-1: Deploy a uniform linear hydrophone array with M elements vertically underwater. Each element corresponds to a single channel, and the element spacing is d and is less than half the signal wavelength. Step 1-2: Calculate the delay difference based on the path difference D between the mth array element and the previous array element ,in, represents the speed of sound, ,in Represents the incident angle of all possible incoming wave directions; Step 1-3: Based on the delay difference, the phase difference of the mth array element is obtained as: ,in, represents the wavelength of the carrier, represents the imaginary unit, represents the frequency of the signal, Indicates The incident angle in the direction of the incoming wave, e represents a natural constant; Step 1-4: The signal received by the mth array element under the influence of underwater noise is expressed as: ,in, is the noise received by the mth array element, Indicates the current incoming wave direction. represents all possible directions of incoming waves in space, Indicates the A narrowband signal source; Step 1-5: Represent the signal in step 1-4 using a matrix: ,in, represents the signal matrix, represents the noise matrix, , represents the flow matrix of the signal, T represents the time period; The formula for calculating the covariance matrix of the signal is: ,in, represents the covariance matrix of the signal, represents the input matrix, It represents the conjugate transpose of the matrix. For a finite length signal, can be caused by Snapshot sampling at the moment To express it, the expression can be written as: , in order to avoid feature loss and preserve signal integrity, Separate into real part and the imaginary part , whose expression is: ， ， For an array of M elements, we get The real part matrix of and the imaginary part matrix , and concatenate them in the vertical dimension to obtain a matrix of size 2M×M; The normalization process includes performing min-max normalization on the spliced ​​result. ,in, represents the signal after normalization, S represents the signal before normalization, and Indicates the maximum and minimum values ​​of the signal within the carrier period; Step 2: Build a lightweight module based on the Mamba network, compress the normalized covariance matrix, adjust the dimension, obtain the correlation of information at different times, dynamically control the output weight through the gating mechanism, mix the signals of all time steps, capture the deep features of the signal, and output the characteristic spectrum of the signal. The mixing of the signals of all time steps includes: for the information at each moment, introduce the time mixing method, for the input ,in represents the batch size, represents the sequence length, represents the hidden dimension, and the temporal mixing is expressed as ,in, represents element-wise multiplication, represents a learnable time parameter, Represents the movement operation in time, the integrated time step The time mixing output can be expressed as: ,The time step adopts an exponential decay method, the size can be learned, and is dynamically allocated by the optimizer to adaptively adjust each time step; Step 3: Process the feature spectrogram using the CBAM module to capture the correlation between deep features, pay attention to important features, adjust the dimension through the fully connected layer, and use the Sigmoid function to output the estimated value of the corresponding direction of arrival; Step 4: Cascade the models of steps 2 and 3, and train them using the dataset in step 1. Adjust the parameters during the training process. After the model training is completed, use the trained model to estimate the direction of arrival of the unknown signal.

2. The underwater acoustic signal direction of arrival estimation method based on the improved Mamba network according to claim 1 is characterized in that: The step 2 includes: Step 2-1: Perform a 2M×1 one-dimensional convolution operation on the data in step 1, expanding the 2M×M matrix into a 1×M sequence. The expression for the one-dimensional convolution is: ,in, represents the input of the first convolutional layer, and Respectively represent the weight matrix and bias term of the t-th convolution kernel in the first convolution layer, Represents the output of the t-th convolution kernel in the first convolution layer, represents the activation function; Step 2-2: Use RELU as the activation function of the one-dimensional convolutional layer. The expression of the RELU function is: ,in, represents the input signal; Step 2-3: The matrix signal is flattened and then input into the Mamba model, which includes the state space model and the gating unit to synchronously process the signal and dynamically control the output weight through the gating mechanism; Step 2-4: Use the Sigmoid function to normalize the gate value to the [0,1] interval. The expression of the Sigmoid function is: ； Steps 2-5: After repeating the above steps for the entire sequence, the information at all moments is merged. The information at each moment contains the state information of the previous moment, and the deep-level features of the signal are extracted.

3. The underwater acoustic signal direction of arrival estimation method based on the improved Mamba network according to claim 2 is characterized in that: Steps 2-3 include: Step a: Adjust the dimension of the matrix signal through a linear layer to obtain a one-dimensional sequence of length 32×M. Use layer normalization to reduce the offset of the internal variables of the model. The expression of layer normalization is as follows: ,in, represents the normalized output of the layer, represents the normalized input of the layer, and Represent the mean and standard deviation of the output of each neuron in the linear layer during the entire process, is a small constant; Step b: The output sequence is expressed through the state space model, state update equation and output equation as follows: , where V is the state transfer matrix, used to control the change of state over time, B is the input mapping matrix, which maps the input signal to the state space, F is the output mapping matrix, which maps the state to the output space, and G is the direct transfer matrix, which is used to map the output signal directly to the output; Step c: For the information at each moment, introduce the time mixing method, for the input ,in represents the batch size, represents the sequence length, represents the hidden dimension, and the temporal mixing is expressed as ,in, represents element-wise multiplication, represents a learnable time parameter, Represents a movement operation in time; Step d: Integrated time step The time mixing output can be expressed as: ,The time step adopts an exponential decay method, the size can be learned, and is dynamically allocated by the optimizer to adaptively adjust each time step.

4. The underwater acoustic signal direction of arrival estimation method based on the improved Mamba network according to claim 1 is characterized in that: In step 3, the CBAM module includes a channel attention mechanism and a spatial attention mechanism, which extracts input features in two stages by sequentially inferring attention maps in the channel and spatial dimensions.

5. The underwater acoustic signal direction of arrival estimation method based on the improved Mamba network according to claim 4 is characterized in that: The step 3 comprises: Step 3-1: The signal passes through the channel attention module, which uses the channel relationship between features to generate a channel attention map to capture the correlation between different channels, expressed as: ,in, represents the sigmoid activation function, and Represent the dimensionality reduction convolution and dimensionality increase convolution operations respectively, and Represents average pooling and maximum pooling operations respectively; Step 3-2: The signal passes through the spatial attention module, which uses the spatial relationship between features to capture the importance of different spatial positions in the features, thereby extracting features in a targeted manner, expressed as: ,in, Indicates that a convolution operation is performed with a convolution kernel of size 7×7. Indicates that the results of average pooling and maximum pooling are spliced ​​in the channel dimension; Step 3-3: Process the signal through three fully connected layers. The expression of the fully connected layer is: ,in, represents the input of the fully connected layer, represents the output of the fully connected layer, represents the weight matrix between two hidden layers, represents the weight matrix between the second hidden layer and the output layer, and represents the bias term, and represents the activation function; Step 3-4: Use the Sigmoid function as the output layer to convert the arrival angle into a continuous distribution between [0, 1], and then perform denormalization to obtain the arrival direction estimation result.

6. The underwater acoustic signal direction of arrival estimation method based on the improved Mamba network according to claim 1 is characterized in that: The step 4 comprises: Step 4-1: Label the signals according to different angles and use them as training sets to train the network model; Step 4-2: During the training process, the mini-batch gradient descent method is used to train the neural network. The training set is divided into several groups of data. The neural network learns information from one group of data in each iteration. Step 4-3: Select mean square error as the loss function and minimize the loss function as the training goal. The expression of mean square error MSE is: ,in represents the number of samples, and Represent the true value and estimated value of the sample respectively; Step 4-4: Iterate the network model multiple times, optimize the network model parameters with the training objectives in step 4-3, save the model parameters, input the unlabeled signal received by the hydrophone array, and the model outputs the angle information of the wave to accurately estimate the direction of arrival.

Images