Underwater acoustic communication signal demodulation method based on enhanced double-branch convolutional neural network

Abstract

The present application belongs to the technical field of underwater acoustic communication signal demodulation, and particularly relates to an underwater acoustic communication signal demodulation method based on an enhanced double-branch convolutional neural network. The received signal is processed into a data set composed of signal segments and their corresponding labels, serving as an input signal; the enhanced double-branch convolutional neural network extracts key features of the input signal in the time domain and the frequency domain respectively, and outputs time domain features and frequency domain features; the time domain features and the frequency domain features are fused to generate fusion features, and after nonlinear transformation and feature mapping of the fusion features, the class probability is output through a Softmax function to obtain a classification result. Compared with the traditional single time domain or frequency domain processing method, the present application extracts key features of the signal in the time domain and the frequency domain respectively through a time domain branch model and a frequency domain branch model, and realizes deep fusion of the two through a cross attention mechanism, so as to fully exploit time-frequency complementary information and effectively improve the demodulation accuracy of digital modulation signals.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic communication signal demodulation technology, specifically relating to an underwater acoustic communication signal demodulation method based on an enhanced dual-branch convolutional neural network. Background Technology

[0002] Underwater acoustic communication is a key technology in fields such as marine exploration, military communications, and environmental monitoring. Due to the complexity of the underwater environment, signal propagation is affected by multipath effects, frequency-selective fading, Doppler effects, and noise interference, making demodulation extremely challenging. Multipath effects can cause severe inter-symbol interference (ISI), and the time-varying characteristics of underwater channels make fixed-parameter demodulation methods difficult to adapt to complex environments. Furthermore, underwater acoustic communication has a low data rate, and channel estimation errors and synchronization errors can both affect demodulation performance.

[0003] Existing demodulation methods are mainly divided into coherent demodulation and non-coherent demodulation. Coherent demodulation has high accuracy when the channel is stable, but it relies on accurate channel estimation and synchronization and is easily affected by underwater acoustic channel fluctuations. Although non-coherent demodulation is more robust, it still has performance bottlenecks in low signal-to-noise ratio or high-speed mobile scenarios. In addition, as underwater acoustic communication develops towards high data rates and complex modulation schemes, traditional demodulation methods have limitations in terms of computational complexity and real-time performance, making it difficult to meet practical needs.

[0004] In recent years, frequency domain processing methods have gradually attracted attention. The Short-Time Fourier Transform (STFT) can convert a time-domain signal into a frequency-domain representation, generating a complex matrix containing time-frequency information. The real and imaginary parts correspond to cosine and sine components, respectively, which helps analyze the frequency distribution of the signal in different time segments. Further calculation of the amplitude and phase spectra, followed by logarithmic transformation and normalization, can improve the dynamic range, computational stability, and reduce the impact of different amplitude variations, thereby enhancing the ability to extract signal features. However, how to efficiently fuse time-domain and frequency-domain features and fully extract time-frequency information remains a challenge in underwater acoustic signal demodulation. Therefore, there is an urgent need for a highly versatile and accurate demodulation method to adapt to complex underwater channels, improve demodulation accuracy, and possess noise resistance and real-time processing capabilities to meet the practical application requirements of modern underwater acoustic communication. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method for demodulating underwater acoustic communication signals based on an enhanced dual-branch convolutional neural network.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: a method for demodulating underwater acoustic communication signals based on an enhanced dual-branch convolutional neural network, comprising the following steps:

[0007] Dataset construction involves processing the received signal into a dataset consisting of signal segments and their corresponding labels, which serves as the input signal.

[0008] Feature extraction enhances the dual-branch convolutional neural network to extract key features of the input signal in both the time and frequency domains, and outputs time-domain and frequency-domain features respectively;

[0009] The time-domain features and frequency-domain features are fused to generate fused features. After nonlinear transformation and feature mapping of the fused features, the class probabilities are output through the Softmax function to obtain the classification result.

[0010] Preferably, the received signal is converted into time-frequency information representation by applying short-time Fourier transform, and then the phase spectrum and amplitude spectrum are extracted and used as two channels to construct input vectors; the received signal is cut into multiple small segments by the sliding window method, each segment is used as an independent sample, and each sample is assigned a label according to its modulation mode, and the processed signal segments and corresponding labels are combined to form a dataset.

[0011] Preferably, the preprocessing of the received signal first maps the discrete bit stream into symbols, with the number of symbol categories corresponding to the modulation order; then, the baseband signal is formed through a pulse shaping filter; and finally, the baseband signal is converted into a frequency band signal using carrier modulation.

[0012] Preferably, the received signal is a digitally modulated signal obtained by adding additive white Gaussian noise to the frequency band signal.

[0013] Preferably, the extracted amplitude spectrum is subjected to logarithmic compression and normalization, and the logarithmically compressed amplitude spectrum is scaled to a fixed interval [0, 1].

[0014] Preferably, the temporal feature extraction first processes the input signal through multi-scale convolutional layers to extract temporal features of different time windows, then uses a one-dimensional max pooling layer to downsample the convolutional output, and then introduces a temporal attention mechanism to weight the pooled features and dynamically adjust the signal weights of each time window.

[0015] Preferably, the convolution kernels in the multi-scale convolutional layer include, but are not limited to, 3x1 and 5x1.

[0016] Preferably, frequency domain feature extraction is performed by first using a composite convolutional layer, which combines multiple convolutional kernels to capture features of different frequency components; then, an adaptive convolutional layer is used to further process the features to adjust the convolution operation according to the characteristics of the input signal; next, a two-dimensional max pooling layer is used to downsample the convolution output; finally, a weighted attention mechanism is introduced to perform weighted processing on the pooled frequency domain features based on specific weights.

[0017] Preferably, a cross-attention mechanism is used to fuse time-domain features and frequency-domain features, calculate weights, enable information exchange between time-domain and frequency-domain features, and dynamically adjust importance weights.

[0018] Preferably, the fused features are passed to the fully connected layer for nonlinear transformation and feature mapping.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] 1. A demodulation method for underwater acoustic communication signals based on enhanced dual-branch convolutional neural networks. Compared with traditional single time-domain or frequency-domain processing methods, this application adopts composite convolution, adaptive convolution, multi-scale feature extraction and other means. By using time-domain branching models and frequency-domain branching models, key features of the signal are extracted in the time domain and frequency domain respectively. The cross-attention mechanism is used to achieve deep fusion of the two, thereby fully mining the complementary information of time and frequency, which can effectively improve the demodulation accuracy of digital modulation signals.

[0021] 2. By introducing a dynamic weighting mechanism and fully connected feature mapping, the model can adaptively adjust the contribution weights of different signal features, further enhancing its ability to distinguish complex modulation modes.

[0022] 3. With the help of the Softmax classification layer, the model can effectively reduce the impact of signal distortion, improve demodulation accuracy, and ensure strong anti-interference and generalization capabilities in complex channel environments, thereby significantly improving the overall accuracy of signal demodulation. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall model structure of the present invention;

[0024] Figure 2 This is a flowchart of the underwater acoustic communication system of the present invention. Detailed Implementation

[0025] To facilitate understanding of the present invention, it will be described in more detail below with reference to the accompanying drawings and specific embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention.

[0026] Combination Figure 1-2 Understanding a method for demodulating underwater acoustic communication signals based on an enhanced dual-branch convolutional neural network, comprising the following steps:

[0027] S1: Generate different types of digitally modulated signals using MATLAB and add Gaussian white noise to simulate an underwater environment. Apply a short-time Fourier transform to the signals to convert them into time-frequency information representation, extract the phase and amplitude spectra, and perform logarithmic compression and normalization to obtain compressed time-frequency information. Divide the signals into multiple small segments using a sliding window method, with each segment serving as an independent sample. Assign a label to each sample based on its modulation scheme, and combine the processed signal segments and their corresponding labels to form a dataset.

[0028] S2: The temporal branching model is designed by first processing the input signal through multi-scale convolutional layers to extract temporal features from different time windows. Next, a one-dimensional max-pooling layer is used to downsample the convolutional output, reducing feature dimensionality and enhancing key temporal information. Subsequently, a temporal attention mechanism is introduced to weight the pooled features, dynamically adjusting the signal weights for each time window.

[0029] S3: The frequency domain branching model is designed. First, frequency domain features are extracted through a composite convolutional layer, which combines multiple convolutional kernels to capture features of different frequency components. Then, an adaptive convolutional layer is used to further process the features, adjusting the convolution operation according to the characteristics of the input signal. Next, a two-dimensional max-pooling layer is used to downsample the convolutional output, reducing the feature dimensionality. Finally, a weighted attention mechanism is introduced to weight the pooled frequency domain features based on specific weights.

[0030] S4: The output features from the time and frequency domain branches are fused using a cross-attention mechanism to generate a comprehensive feature representation. This fused feature is then passed to a fully connected layer for nonlinear transformation and feature mapping. Finally, the class probability is output through the Softmax function in the classification layer to obtain the classification result, which is the original symbol.

[0031] Step S1, the specific steps are as follows:

[0032] S1-1: Determine the modulation parameters, such as the number of symbols, modulation order, carrier frequency, and symbol spacing. Different modulation schemes can all be represented using a unified modulation model. During generation, the discrete bitstream is first mapped to symbols, and the number of symbol categories and modulation order are determined accordingly. Correspondingly, the baseband signal is then formed through a pulse shaping filter, and finally, the baseband signal is converted into a frequency band signal using carrier modulation.

[0033] In the unified digital modulation model, the baseband signal can be represented as:

[0034] ;

[0035] in, Represents a signed number; Indicates the first A number symbol, The ordinal number of the digit symbol. For time, the size of the symbol set corresponds to the modulation order. ; This is the impulse response of a pulse shaping filter, used to limit bandwidth and reduce inter-symbol interference; It is the symbol interval, that is, the duration of each symbol.

[0036] Frequency band modulation is achieved by multiplying the baseband signal and the carrier signal. The carrier signal is defined as:

[0037] ;

[0038] Therefore, the modulated frequency band signal is:

[0039] ;

[0040] in, For carrier frequency; This indicates taking the real part, ensuring that the transmitted signal is a real number; The spectrum of the baseband signal is shifted to be near the carrier frequency.

[0041] To simulate common noise interference in underwater environments, additive white Gaussian noise (AWGN) is added to the frequency band signal. The model is as follows:

[0042] ;

[0043] in, To receive signals; It is Gaussian white noise with a mean of 0 and a variance of 1. It follows a normal distribution.

[0044] S1-2: In order to capture the changes in signal over time and frequency, the Short Time Fourier Transform (STFT) is often used to transform the received signal. The spectrum is then transformed to the time-frequency domain. The complex matrix obtained through STFT contains both amplitude and phase information, providing rich information for subsequent feature extraction and neural network processing. To further adapt to the requirements of neural networks, the amplitude spectrum is typically subjected to logarithmic compression and normalization.

[0045] STFT converts a signal into a time-frequency representation after localization, and its mathematical definition is:

[0046] ;

[0047] in, In time and frequency Complex representation of numbers on; It is a window function used to localize signals; Indicates the center time of the window function. From The amplitude spectrum and phase spectrum are extracted separately using the following formulas:

[0048] ;

[0049] ;

[0050] in, express The real part; express The imaginary part; Indicates at time and frequency points Upper amplitude spectrum; This indicates the phase information at that point.

[0051] To reduce the dynamic range of the amplitude spectrum, logarithmic compression is used:

[0052] ;

[0053] The formula is passed The amplitude values ​​are compressed to avoid taking the logarithm of zero, while making smaller amplitude information more apparent. Furthermore, a normalization operation is used to scale the logarithmically compressed amplitude spectrum to a fixed interval [0,1], as shown in the formula:

[0054] ;

[0055] in, and These are the minimum and maximum values ​​at all time-frequency points, respectively.

[0056] The normalized logarithmic magnitude spectrum and phase spectrum are used as two channels respectively. The input vector is constructed. The final input tensor has the following shape: ,in This indicates the number of time frames obtained from STFT. 2 indicates frequency resolution, and 2 indicates two feature channels.

[0057] S1-3: To construct a dataset for neural network training, long signals typically need to be segmented into multiple short-time signal segments. The sliding window method ensures that each segment contains sufficient local feature information while maintaining a sufficient number of samples for model training. Therefore, this method uses the sliding window method to segment long signals into multiple fixed-length segments and assigns a label to each segment according to the modulation scheme, forming a dataset for easy input into the network model.

[0058] Step S2, the specific steps are as follows:

[0059] S2-1: In time-domain signal processing, convolutional neural networks are widely used to extract local features from raw signals. Through multi-scale convolution operations, features can be extracted at different time scales, capturing local patterns in the signal. Using convolution kernels of different sizes allows the model to adapt to features in different time windows, improving its ability to learn signals.

[0060] This paper selects two different convolutional kernel sizes, 3x1 and 5x1, to extract features at different time scales. Here, 3 and 5 represent the length of the convolutional kernel in the time dimension, and 1 represents its size in the feature dimension. The 3x1 convolutional kernel captures features within a shorter time window, while the 5x1 convolutional kernel captures features within a longer time window. Assume the... Each convolutional kernel has a size of 3 or 5, and the convolution operation can be represented as:

[0061] ;

[0062] Among them is Input signal, It is the i-th convolutional kernel, with a size of 3 or 5. Indicates the size of the convolution kernel. This is the output feature map after convolution. The above feature maps are stacked into a multidimensional matrix. This matrix contains time-domain features at different scales:

[0063] ;

[0064] S2-2: Pooling layers in convolutional neural networks are used for downsampling, which reduces feature map size and computational complexity while preserving important local information. Max pooling is often used to retain the largest activation value in each local region, thereby highlighting the most significant signal features. Based on the output after multi-scale convolution, this paper uses a pooling layer with a size of [missing information]. Pooling window and stride The max pooling operation outputs:

[0065] ;

[0066] This operation reduces the dimensionality of the original temporal features and makes the network invariant to local translations.

[0067] S2-3: The attention mechanism dynamically weights the importance of different time windows, enabling the network to focus on the signal components most useful for demodulation. The temporal attention mechanism designed in this paper calculates the attention weight for each time window using a trainable function, and then weights the pooled features. Let... Indicates the first The pooled feature vectors are then subjected to attention scoring by a fully connected layer. Obtain the score for each window, then normalize and calculate the attention weights:

[0068] ;

[0069] For the first The feature scoring of each window is represented by a linear transformation in this paper: , This represents the normalized attention weights, followed by a weighted average of the features for each time window:

[0070] ;

[0071] Based on the above representation, the weighted output This indicates that the model focuses on the temporal features of the most important moments.

[0072] Features processed by multi-scale convolution, max pooling, and temporal attention mechanisms This constitutes the final feature representation of the time-domain branch, where This represents the number of time windows remaining after pooling. These features will be used as input to subsequent fusion network layers for further classification.

[0073] Step S3, the specific steps are as follows:

[0074] S3-1: After the frequency domain signal undergoes a short-time Fourier transform, it is usually represented as a complex matrix. It means that among them Represents time frame, Frequency is represented. To capture information from different frequency components, a composite convolutional layer is used to convolve this frequency domain representation using multiple convolutional kernels of different sizes. Multiple convolutional kernels can extract local frequency domain features at different scales, thus giving the model a richer spectral representation capability.

[0075] For frequency domain input data In implementation, it is treated as a joint representation of real and imaginary parts, for the th convolution kernels The compound convolution operation is as follows:

[0076] ;

[0077] in, For the input frequency domain matrix, For the first Each convolutional kernel is located at... The weight, For the first The feature maps output by each convolutional kernel are concatenated along the channel direction to form the output of a composite convolutional layer.

[0078] ;

[0079] in This indicates the number of convolutional kernels used. The concatenated output has multiple feature channels and can be used for subsequent adaptive convolution processing.

[0080] S3-2: The adaptive convolutional layer designed in this paper dynamically adjusts the convolution kernel parameters according to the characteristics of the input signal, thereby achieving more flexible feature extraction. By introducing a trainable parameter adjustment factor, the convolution operation becomes locally adaptive to different signals. Based on the output of the compound convolution... The adaptive convolutional layer further processes it, and the formula can be expressed as:

[0081] ;

[0082] As an adaptive filter, its parameters are learned through backpropagation, which can be viewed as a dynamic adjustment of a traditional convolutional kernel. The output of the adaptive convolutional layer is used to adjust the window size. and step size Max pooling, the pooling operation is as follows:

[0083] ;

[0084] in Indicates the first Each pooling window covers a region. After pooling, the size of the feature map is reduced while retaining the most salient local frequency domain features. This represents the output after pooling.

[0085] S3-3: The weighted attention mechanism primarily targets the signal distribution along the frequency axis, calculating weights based on the significance of each frequency component. This allows the model to highlight frequency regions that are more critical to the task while suppressing interference from irrelevant frequency bands and optimizing feature extraction. This part utilizes fully connected layers. Calculate attention score Then normalize to obtain the weights. :

[0086] ;

[0087] ;

[0088] Then, the pooled output is weighted:

[0089] ;

[0090] After the above steps, the output of the frequency domain branch model is the feature processed by compound convolution, adaptive convolution, max pooling, and weighted attention mechanisms. This output serves as the input to the subsequent neural network, used to complete the signal demodulation task.

[0091] Step S4, the specific steps are as follows:

[0092] S4-1: A cross-attention mechanism is employed to fuse time-domain and frequency-domain features. This module calculates weights, enabling information exchange between time-domain and frequency-domain features and dynamically adjusting their importance weights. This allows for the construction of a more comprehensive feature representation, improving feature expressive power. For time-domain branch features... Frequency domain branch Use linear transformations to calculate Query, Key, and Value respectively:

[0093] ;

[0094] ;

[0095] in, This refers to the query and key mapping of time-domain features to frequency-domain features. This refers to the query and key mapping of frequency domain features to time domain features. For value vector mapping, It is the latent space dimension of attention computation.

[0096] The cross-attention score is derived from the attention of time-domain features to frequency-domain features, and the attention of frequency-domain features to time-domain features:

[0097] ;

[0098] ;

[0099] in, This represents the attention weight of time-domain features on frequency-domain features. The normalization factor represents the attention weight of frequency domain features on time domain features. Used to stabilize the gradient. Then, weighted features are calculated:

[0100] ;

[0101] in, and Representing the enhancement information obtained from the frequency domain branch of the time-domain features, and the enhancement information obtained from the time domain branch of the frequency-domain features, respectively, the final fusion of time-domain and frequency-domain features:

[0102] ;

[0103] The enhanced features generated by cross-attention are concatenated to the original features to ensure complete information representation, and then transformed through a fully connected layer:

[0104] ;

[0105] in, These are the weights of the fully connected layer. It is a bias term. It is a non-linear activation function. It is the feature dimension after fusion. It's the hidden layer dimension, and finally, the features. After passing through a classification head, the class probability is output:

[0106] ;

[0107] in, For the final classification layer weights, For classification bias, Given the number of categories, Softmax calculates the category probabilities:

[0108] ;

[0109] The classification results are as follows:

[0110] .

[0111] Using the above method, after the modulated signal is input into the model, the predicted symbol will be automatically output, thus achieving signal demodulation.

[0112] This underwater acoustic communication signal demodulation method, based on an enhanced dual-branch convolutional neural network, employs composite convolution, adaptive convolution, and multi-scale feature extraction techniques to extract key signal features in both the time and frequency domains, compared to traditional single time-domain or frequency-domain processing methods. A cross-attention mechanism is then used to achieve deep fusion of these two features, fully leveraging complementary time-frequency information and effectively improving the demodulation accuracy of digitally modulated signals. By introducing a dynamic weighting mechanism and fully connected feature mapping, the model can adaptively adjust the contribution weights of different signal features, further enhancing its ability to distinguish complex modulation modes. With the help of the Softmax classification layer, the model effectively reduces the impact of signal distortion, improves demodulation accuracy, and ensures strong anti-interference and generalization capabilities even in complex channel environments, thereby significantly improving the overall accuracy of signal demodulation.

[0113] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for demodulating underwater acoustic communication signals based on an enhanced dual-branch convolutional neural network, characterized in that, The steps are as follows: Dataset construction involves processing the received signal into a dataset consisting of signal segments and their corresponding labels, which serves as the input signal. Feature extraction enhances the dual-branch convolutional neural network to extract key features of the input signal in both the time and frequency domains, and outputs time-domain and frequency-domain features respectively; The time-domain features and frequency-domain features are fused to generate fused features. After nonlinear transformation and feature mapping of the fused features, the class probabilities are output through the Softmax function to obtain the classification result. The received signal is converted into time-frequency information representation by applying short-time Fourier transform, and then the phase spectrum and amplitude spectrum are extracted and used as two channels to construct the input vector. The received signal is then cut into multiple small segments by the sliding window method, with each segment as an independent sample. Each sample is assigned a label according to its modulation mode, and the processed signal segments and corresponding labels are combined to form a dataset. Temporal feature extraction first processes the input signal through multi-scale convolutional layers to extract temporal features of different time windows. Then, a one-dimensional max pooling layer is used to downsample the convolutional output. Next, a temporal attention mechanism is introduced to weight the pooled features and dynamically adjust the signal weights of each time window. Frequency domain feature extraction is first performed using a composite convolutional layer, which combines multiple convolutional kernels to capture features of different frequency components. Subsequently, an adaptive convolutional layer is used to further process the features to adjust the convolution operation according to the characteristics of the input signal. Next, a two-dimensional max pooling layer is used to downsample the convolution output. Finally, a weighted attention mechanism is introduced to perform weighted processing on the pooled frequency domain features based on specific weights.

2. The underwater acoustic communication signal demodulation method based on an enhanced dual-branch convolutional neural network according to claim 1, characterized in that, The preprocessing of the received signal first maps the discrete bit stream into symbols, and the number of symbol categories corresponds to the modulation order; The baseband signal is then formed using a pulse shaping filter, and finally converted into a frequency band signal using carrier modulation.

3. The underwater acoustic communication signal demodulation method based on an enhanced dual-branch convolutional neural network according to claim 2, characterized in that, The received signal is a digitally modulated signal obtained by adding additive white Gaussian noise to the frequency band signal.

4. The underwater acoustic communication signal demodulation method based on an enhanced dual-branch convolutional neural network according to claim 1, characterized in that, The extracted amplitude spectrum is subjected to logarithmic compression and normalization, and the logarithmically compressed amplitude spectrum is scaled to a fixed interval [0, 1].

5. The underwater acoustic communication signal demodulation method based on an enhanced dual-branch convolutional neural network according to claim 1, characterized in that, Convolutional kernels in multi-scale convolutional layers include, but are not limited to, 3x1 and 5x1.

6. The underwater acoustic communication signal demodulation method based on an enhanced dual-branch convolutional neural network according to claim 1, characterized in that, A cross-attention mechanism is used to fuse time-domain and frequency-domain features. Calculate weights to enable information exchange between time-domain and frequency-domain features, and dynamically adjust importance weights.

7. The underwater acoustic communication signal demodulation method based on an enhanced dual-branch convolutional neural network according to claim 1, characterized in that, The fused features are passed to the fully connected layer for nonlinear transformation and feature mapping.

Images