Underwater acoustic signal modulation identification method and system based on multi-resolution SSM fusion

Abstract

The invention discloses an underwater acoustic signal modulation identification method and system based on multi-resolution SSM fusion. The method comprises the following steps: preprocessing an underwater acoustic communication signal acquired by a receiving end; extracting low-layer time domain features of the preprocessed signals by using a one-dimensional residual network; the extracted low-layer time domain features are input into a constructed multi-resolution SSM fusion module for deep time sequence feature processing and fusion; performing global average pooling or attention weighted pooling on the fused feature sequence to obtain a global feature vector of the signal; and outputting a modulation type identification result through a full-connection classifier according to the global feature vector. According to the invention, through a multi-resolution state space model fusion structure, different time scale characteristics of underwater acoustic signals are effectively captured, a modulation identification method and system capable of realizing high-robustness and high-accuracy identification in a complex underwater acoustic channel are provided, and the stability and classification performance of the model are improved.

Description

Technical Field

[0001] This invention belongs to the field of digital signal processing and relates to underwater acoustic signal modulation recognition technology, specifically to an underwater acoustic signal modulation recognition method and system based on multi-resolution SSM fusion. Background Technology

[0002] In underwater acoustic communication, modulation identification is a crucial step in signal processing. However, the underwater acoustic channel environment is extremely complex, with severe multipath effects, time-varying characteristics, and high-intensity environmental noise, resulting in a typically low signal-to-noise ratio (SNR) at the receiver.

[0003] In existing technologies, methods for solving such problems mainly include those based on convolutional neural networks (CNNs). These methods utilize CNNs to extract local features of signals. For example, Chinese patent CN113114599A discloses a modulation recognition method based on a lightweight neural network. However, the inherent local receptive field of CNNs limits their ability to capture long-term temporal dependencies in signals. This leads to the loss of long-term correlated features at low signal-to-noise ratios, resulting in poor recognition accuracy. Recurrent neural network (RNN) models are prone to feature degradation, unstable convergence, or decreased classification accuracy when processing signals with long-term dependencies and low signal-to-noise ratios.

[0004] The state-space model (SSM) and structures such as Mamba proposed in recent years have shown good long-term dependency processing capabilities in sequence modeling, but when directly applied to underwater acoustic signals, they only process a single time scale and are difficult to take into account both short-term transients and long-term statistical characteristics.

[0005] Therefore, there is an urgent need for a new modulation recognition method that takes into account feature extraction at multiple time scales, stability, and high accuracy. Summary of the Invention

[0006] Purpose of the invention: To overcome the shortcomings of the existing technology, this invention provides a method and system for underwater acoustic signal modulation recognition based on multi-resolution SSM fusion. By using a multi-resolution state-space model fusion structure, it effectively captures the different time-scale features of underwater acoustic signals, providing a modulation recognition method and system that can achieve high robustness and high accuracy in complex underwater acoustic channels, thereby improving the stability and classification performance of the model.

[0007] Technical Solution: To achieve the above objectives, this invention provides a method for underwater acoustic signal modulation identification based on multi-resolution SSM fusion, comprising the following steps:

[0008] S1: Preprocess the underwater acoustic communication signal collected by the receiving end;

[0009] S2: Use a one-dimensional residual network to extract the low-level time-domain features of the preprocessed signal;

[0010] S3: Input the extracted low-level temporal features into the constructed multi-resolution SSM fusion module for deep temporal feature processing and fusion;

[0011] S4: Perform global average pooling or attention-weighted pooling on the fused feature sequence to obtain the global feature vector of the signal;

[0012] S5: Based on the global feature vector, output the modulation type recognition result through a fully connected classifier.

[0013] Furthermore, the preprocessing in step S1 includes:

[0014] Underwater acoustic communication signals contain I (in-phase) and Q (quadrature) components. First, the signal is normalized to eliminate the influence of sampling gain and power differences. Then, the signal is segmented into frames, and the length of each frame is unified to a fixed length. The complex sequence I / Q signal is preprocessed into a dual-channel real number sequence.

[0015] Furthermore, the extraction process in step S2 includes:

[0016] A1: The dimensions of the input signal are transformed through a dimension transformation operation (Permute);

[0017] A2: The transformed signal is passed through an initial convolutional layer, accompanied by batch normalization and ReLU activation function, for preliminary feature extraction and channel expansion;

[0018] A3: The initial features are downsampled using a max pooling layer to further reduce the sequence length and enhance the robustness of the features;

[0019] A4: Input the feature sequence processed in step A3 into the ResNet backbone network and output the feature map;

[0020] A5: The output feature map is transformed again through a dimension transformation operation (Permute) to swap the channel dimension and sequence length dimension, and then converted to the corresponding format.

[0021] Furthermore, the ResNet backbone network in step A4 consists of four stacked ResNet layers, specifically:

[0022] ResNet Layer 1 contains 3 residual blocks and has 64 output channels;

[0023] ResNet Layer 2 contains 4 residual blocks, has 128 output channels, and a stride of 2;

[0024] ResNet Layer 3 contains 6 residual blocks, has 256 output channels, and a stride of 2;

[0025] ResNet Layer 4 contains 3 residual blocks, with 512 output channels (i.e., d_model) and a stride of 2.

[0026] Furthermore, in step S3, the multi-resolution SSM fusion module employs three parallel Selective State-Space Model (SelectiveSSM) branches: a short-time SSM branch, a medium-time SSM branch, and a long-time SSM branch, wherein:

[0027] Short-time SSM branch: The convolution dilation stride is set to 1 and the state dimension is set to 8. It is used to extract the transient change features and local mutation information of the signal in a short time.

[0028] Mid-time SSM branch: The convolution dilation stride is set to 2 and the state dimension is set to 16, which is used to extract the temporal dependency features of the signal in the mid-time range.

[0029] The long-term SSM branch sets the convolution dilation stride to 4 and the state dimension to 32 to capture the long-term temporal correlation and overall statistical trend of the signal.

[0030] Furthermore, the execution process of the Selective State-Space Model (SelectiveSSM) branch in step S3 includes:

[0031] B1: Input feature X passes through the input projection layer, which is essentially a linear transformation layer that expands the feature dimension;

[0032] B2: The expanded features are split into two along the feature dimension by a feature segmentation (Chunk) operation, resulting in X1 and X2; X2 will be used as the gating signal for the final output.

[0033] B3: Transpose X1 in dimensions to accommodate subsequent causal one-dimensional dilated convolution (Causal Conv1D) operations;

[0034] B4: The causal one-dimensional dilated convolutional layer has a configurable dilation rate (Dil), which allows it to capture local contextual information at different time scales. The convolution result is then transposed again and passed through the SiLU activation function to generate intermediate features u;

[0035] B5: Feed the intermediate feature u into a key parameter to generate the projection layer X_Proj;

[0036] The X_Proj module employs a low-rank decomposition structure, specifically implemented as a linear layer (where r is typically 4) followed by another linear layer. This bottleneck structure significantly reduces the number of parameters and computational overhead, while also efficiently extracting information from u. The output of X_Proj is split into three time-varying parameters: Delta, B, and C (shape: [B, L, 1]) determine the intensity of the SSM state update after passing through the Softplus activation function; B (shape: [B, L, d_state]) serves as the mapping matrix from input to state; C (shape: [B, L, d_state]) serves as the mapping matrix from state to output; the X_Proj module also contains a fixed parameter A (derived from A_log), which is defined by nn.Parameter and is usually initialized as a linear sequence in logarithmic space, determining the inherent decay characteristics of the SSM;

[0037] B6: Intermediate feature u, activated by Softplus (Delta), B, C, and the fixed parameter A are fed into the selective scan module. The selective scan module calculates the hidden state h step by step based on the state update equation through a loop calculation. After the loop is completed, the complete scan result is obtained. ;

[0038] B7: Output of the selective scan module The initial separated X2 is gated through element-wise multiplication; this mechanism allows X2 to act as a selection gate, dynamically adjusted. The importance of features is highlighted, thereby enabling the filtering and enhancement of information;

[0039] B8: The gated features are passed through the output projection layer (Out_Proj, a linear transformation) to obtain the final output Y of the Selective State Space Model (SelectiveSSM) module (shape: [Batch, sequence length L, feature dimension D]).

[0040] Furthermore, the calculation process of the hidden state h in step B6 includes: firstly, using parameters... Discretize the continuous parameters to obtain and Then execute the state update equation:

[0041]

[0042] in, This represents element-wise multiplication. Enter the current time.

[0043] Furthermore, in step S3, the multi-resolution SSM fusion module uses an adaptive fusion network to perform multi-scale feature fusion and gating enhancement; the adaptive fusion network includes a fully connected layer, ReLU activation and Dropout layer, and introduces a gated linear unit (GLU) for residual gating.

[0044] The features processed by GLU are dimensionally adjusted by the output projection layer, and the output of the output projection layer is added to the residual of the output S of the short-time SSM branch.

[0045] The present invention also provides a hydroacoustic signal modulation identification system based on multi-resolution SSM fusion, comprising:

[0046] The preprocessing module preprocesses the underwater acoustic communication signals collected by the receiving end.

[0047] The time-domain feature extraction module uses a one-dimensional residual network to extract low-level time-domain features of the preprocessed signal;

[0048] The multi-scale feature fusion and gating enhancement module inputs the extracted low-level temporal features into the constructed multi-resolution SSM fusion module for deep temporal feature processing and fusion;

[0049] The classification output module performs global average pooling or attention-weighted pooling on the fused feature sequence to obtain the global feature vector of the signal and outputs the modulation type identification result.

[0050] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0051] (1) Multi-timescale joint modeling: Through the short, medium and long-term three-branch structure, it is possible to capture the rapid change characteristics and long-term trend characteristics of the signal at the same time, so as to achieve a balance between time resolution and global dependence.

[0052] (2) Dynamic parameter adaptation mechanism: The input-driven selective SSM structure is adopted to make the parameters adapt to the signal characteristics, which significantly improves the generalization ability and channel adaptability of the model.

[0053] (3) Low-rank bottleneck and stabilized state matrix: Low-rank mapping is introduced to reduce redundant computation, and at the same time, the numerical stability is improved by the diagonal negative exponential state matrix structure to avoid gradient explosion.

[0054] (4) LayerNorm and GLU gating mechanism can automatically learn the weight relationship between multi-resolution features, reduce feature conflict under low SNR and improve classification robustness.

[0055] (5) Streaming online recognition capability: The model provides step-by-step inference (step function), which can output recognition results in real time in the scenario of continuous signal reception, meeting the real-time requirements of underwater communication system.

[0056] (6) Significant performance improvement: The present invention improves the average recognition rate by about 12% compared with traditional CNN in the signal-to-noise ratio range of -10 dB to 18 dB. The performance improvement is particularly significant under low signal-to-noise ratio conditions of -2 dB and below. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 This is a schematic diagram of the overall architecture of the model in this invention;

[0059] Figure 3 This is a block diagram of the internal structure of the multi-resolution SSM fusion module;

[0060] Figure 4 This is a block diagram of the internal structure of the Selective State-Space Model (SelectiveSSM) module.

[0061] Figure 5 This paper compares the accuracy of the method of this invention with that of existing mainstream models under different signal-to-noise ratios. Detailed Implementation

[0062] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0063] Example 1:

[0064] like Figure 1 and Figure 2 As shown, this embodiment provides a method for underwater acoustic signal modulation identification based on multi-resolution SSM fusion, including the following steps:

[0065] S1: Preprocess the underwater acoustic communication signal collected by the receiving end;

[0066] Underwater acoustic communication signals contain I (in-phase) and Q (quadrature) components. First, the signal is normalized to eliminate the influence of sampling gain and power differences. Then, the signal is segmented into frames, and the length of each frame is unified to a fixed length. The I / Q signals of the complex sequence are preprocessed into a dual-channel real sequence.

[0067] In this embodiment, the preprocessed I / Q signal has the following data shape: [Batch, Sequence Length L, Input Channels C=2]

[0068] S2: Use a one-dimensional residual network (ResNet-34) to extract low-level temporal features of the preprocessed signal;

[0069] The extraction process in step S2 includes:

[0070] A1: To adapt to the input requirements of a one-dimensional convolutional layer, a dimension transformation operation (Permute) is used to transform the dimension of the input signal from [Batch, sequence length L, 2] to [Batch, 2, sequence length L].

[0071] A2: The transformed signal is passed through an initial convolutional layer (in this embodiment, a Conv1D with a kernel size of 7, a stride of 2, and padding of 3), accompanied by batch normalization and ReLU activation function, for preliminary feature extraction and channel expansion;

[0072] A3: The initial features are downsampled by a max pooling layer (in this embodiment, MaxPool1D with a kernel size of 3, a stride of 2, and a padding of 1) to further reduce the sequence length and enhance the robustness of the features;

[0073] A4: Input the feature sequence processed in step A3 into the ResNet backbone network and output the feature map;

[0074] The ResNet backbone consists of four stacked ResNet layers, specifically:

[0075] ResNet Layer 1 contains 3 residual blocks and has 64 output channels;

[0076] ResNet Layer 2 contains 4 residual blocks, has 128 output channels, and a stride of 2;

[0077] ResNet Layer 3 contains 6 residual blocks, has 256 output channels, and a stride of 2;

[0078] ResNet Layer 4 contains 3 residual blocks, with 512 output channels (i.e., d_model) and a stride of 2.

[0079] The ResNet backbone aims to extract hierarchical local features from the original signal and further compress the sequence length, ultimately outputting a feature map of shape [Batch, d_model, new_seq_len].

[0080] A5: In order to adapt to the input of the subsequent multi-resolution SSM fusion module, the feature map output by the ResNet backbone network is transformed again by the dimension transformation operation (Permute), which swaps the channel dimension and the sequence length dimension, and converts it into the format of [Batch, new_seq_len, d_model].

[0081] The ResNet-34 network provided in this embodiment employs a combination of multi-layer convolution, batch normalization, and residual structures, which can effectively preserve local detail features and alleviate the gradient vanishing problem. Through feature extraction in this stage, temporal features robust to multipath propagation and noise perturbations can be obtained.

[0082] S3: Input the extracted low-level temporal features into the constructed multi-resolution SSM fusion module for deep temporal feature processing and fusion;

[0083] like Figure 3 As shown, the multi-resolution SSM fusion module receives a deep feature sequence X (shape: [Batch, sequence length L, feature dimension D]) from the ResNet backbone, aiming to capture and effectively fuse information at different time scales in parallel. The core design of the multi-resolution SSM fusion module lies in employing three parallel Selective State Space Model (SelectiveSSM) branches: a short-time SSM branch, a medium-time SSM branch, and a long-time SSM branch, where:

[0084] Short-time SSM branch: The convolution dilation stride is set to 1 and the state dimension is set to 8. It is used to extract the transient change features and local mutation information of the signal in a short time.

[0085] Mid-time SSM branch: The convolution dilation stride is set to 2 and the state dimension is set to 16, which is used to extract the temporal dependency features of the signal in the mid-time range.

[0086] The long-term SSM branch sets the convolution dilation stride to 4 and the state dimension to 32 to capture the long-term temporal correlation and overall statistical trend of the signal.

[0087] The short-time SSM branch is configured with a smaller state dimension (e.g., d_state=8) and a shorter dilation rate (e.g., dilation=1). Its main function is to focus on capturing fast transient features and local contextual information in the signal, and its output is denoted as S. The medium-time SSM branch is configured with a medium state dimension (e.g., d_state=16) and a medium dilation rate (e.g., dilation=2), aiming to capture dependencies and patterns at medium time scales in the signal, and its output is denoted as M. The long-time SSM branch is configured with a larger state dimension (e.g., d_state=32) and a larger dilation rate (e.g., dilation=4), used to model long-term dependencies and macroscopic structures in the signal, and its output is denoted as L. The three SSM branches, through different configurations, allow the model to examine the input sequence from multiple time resolution perspectives, effectively compensating for the potential shortcomings of a single-time-scale SSM.

[0088] like Figure 4 As shown, the Selective State-Space Model (SelectiveSSM) branch receives input features X of shape [Batch, sequence length L, feature dimension D] and outputs features Y of the same shape. The operation of the SelectiveSSM module includes:

[0089] B1: Input feature X passes through the input projection layer, which is essentially a linear transformation layer that expands the feature dimension;

[0090] B2: The expanded features are split into two along the feature dimension by a feature segmentation (Chunk) operation, resulting in X1 and X2; X2 will be used as the gating signal for the final output.

[0091] B3: Transpose X1 in dimensions to accommodate subsequent causal one-dimensional dilated convolution (Causal Conv1D) operations;

[0092] B4: A causal one-dimensional dilated convolutional layer with a configurable dilation rate (Dil), which allows it to capture local contextual information at different time scales. The mathematical expression for this convolution operation is:

[0093]

[0094] in, Let be the convolution output at time t. Given the input sequence, For convolution kernel weights, The kernel size is [size]. The dilation coefficient; convolution result Perform a dimension transpose again, and then apply the SiLU activation function, which is defined as:

[0095]

[0096] After the above nonlinear transformation, intermediate features are generated. ;

[0097] B5: Feed the intermediate feature u into a key parameter to generate the projection layer X_Proj;

[0098] Each SSM branch adaptively generates state-space model parameters based on input features, including: time step parameters. Delta, B, and C are constrained to positive values ​​by the Softplus function to ensure the integrability of time evolution; the input transformation matrix B and the output transformation matrix C are used to control the propagation and reconstruction of the signal in the state space; the state matrix A is defined as a diagonal negative exponential form, where the parameter θ on the exponent is a trainable parameter, thereby ensuring the stability of the system.

[0099] The X_Proj module employs a low-rank decomposition structure, specifically implemented as a linear layer with r set to 4, followed by another linear layer. Here, r represents the dimensionality reduction ratio. By setting r=4, the model first compresses the d_model dimensional features to d_model / 4, processes them, and then maps them back to the original dimension. This bottleneck structure significantly reduces the number of parameters (by approximately 75%) without sacrificing performance; in fact, it provides a certain degree of regularization.

[0100] This bottleneck structure significantly reduces the number of parameters and computational overhead, while also effectively extracting information from intermediate features u. The output of X_Proj is split into three time-varying parameters: Delta, B, and C (shape: [B, L, 1]) determine the intensity of the SSM state update after passing through the Softplus activation function; B (shape: [B, L, d_state]) serves as the mapping matrix from input to state; C (shape: [B, L, d_state]) serves as the mapping matrix from state to output; the X_Proj module also contains a fixed parameter A (derived from A_log), which is defined through nn.Parameter and initialized to a diagonal negative exponential form, i.e. ,in As trainable logarithmic parameters, this design ensures that the real part of A is always negative, thus guaranteeing the mathematical stability of the system in long-term processing.

[0101] B6: Intermediate feature u, activated by Softplus B, C, and fixed parameter A are fed into the selective scan module. The selective scan module calculates the hidden state h step by step based on the state update equation through a loop calculation. After the loop is completed, the complete scan result is obtained. The specific calculation process is as follows: First, using the parameters... Discretize the continuous parameters to obtain and Then execute the state update equation:

[0102]

[0103] in, This represents element-wise multiplication. Enter the current time;

[0104] B7: Output of the selective scan module The initial separated X2 is gated through element-wise multiplication; this mechanism allows X2 to act as a selection gate, dynamically adjusted. The importance of features is considered to achieve information filtering and enhancement. The mathematical expression for this gating operation is:

[0105]

[0106] Where X2 is the gated branch segmented directly from the input features. The output of the scanning module is multiplied element-wise between the two.

[0107] B8: The gated features are passed through the output projection layer (Out_Proj, a linear transformation) to obtain the final output Y of the Selective State Space Model (SelectiveSSM) module (shape: [Batch, sequence length L, feature dimension D]).

[0108] In this embodiment, the SelectiveSSM module also provides a step method, which allows for online (streaming) processing of single-time-step inputs without changing the model architecture. This is of great significance for real-time system deployment. In the step method, the model caches the hidden state of the previous step. And calculate only based on the input u_t at the current time step. , , And updates and .

[0109] The outputs S, M, and L of the three SSM branches are concatenated (Cat) along the feature dimension to form a comprehensive feature sequence containing multi-scale information. The concatenated features are then normalized using LayerNorm to stabilize training and accelerate convergence. The normalized features are then fed into a fusion network. The multi-resolution SSM fusion module employs an adaptive fusion network for multi-scale feature fusion and gated enhancement. This adaptive fusion network includes fully connected layers, ReLU activation, and Dropout layers, and introduces a gated linear unit (GLU) for residual gating. The fully connected layers, ReLU activation, and Dropout layers are responsible for mapping the concatenated high-dimensional features back to the original D dimension and performing preliminary nonlinear processing. The fused features then pass through a gated linear unit (GLU). The GLU module learns a gating mechanism to adaptively filter and weight the fused features, thereby selectively enhancing important information and suppressing potential redundancy. The specific calculation formula for GLU is as follows:

[0110]

[0111] The input features are segmented into content streams. and gated flow , This is the sigmoid function. The GLU module is designed using a gating mechanism, specifically utilizing learned gating weights. For content Dynamic filtering is performed to dynamically adjust the contribution of multi-scale features, suppress noise interference, and thus enhance the complementarity of signals at different time scales.

[0112] Finally, the features processed by GLU are passed through an output projection layer (linear transformation) to adjust their dimension to D. The output of this output projection layer is then added to the residual of the output S of the short-time SSM branch ("+" operation).

[0113]

[0114] in, This is the final output of the multi-resolution fusion module. This represents the deep features after fusion, gating, and output projection. This represents the raw output of the short-term SSM branch.

[0115] This residual connection mechanism is crucial because it allows short-term feature information to still be directly transmitted to the module output in the form of a "highway" after multi-scale fusion processing, effectively avoiding information loss and contributing to the increase in model depth and training stability, ensuring that the model can make full use of features at different resolutions.

[0116] S4: Perform global average pooling or attention-weighted pooling on the fused feature sequence to obtain the global feature vector of the signal;

[0117] In this embodiment, the feature map output by the multi-resolution SSM fusion module (with shape [Batch, new_seq_len, d_model]) is then aggregated into a fixed-length feature vector (with shape [Batch, d_model]) by global average pooling.

[0118] S5: Finally, these aggregated feature vectors are fed into a fully connected layer (FC Classifier), which maps the features to the final classification space and outputs scores for nine different modulation types (modulation types include OFDM, 2FSK, 4FSK, 8FSK, BPSK, QPSK, 8PSK, 16QAM and 64QAM). After SoftMax activation, the probability of each modulation type is obtained, and the category with the highest probability is selected as the final modulation recognition result.

[0119] Example 2:

[0120] This embodiment provides an underwater acoustic signal modulation recognition system based on multi-resolution SSM fusion, including:

[0121] The preprocessing module preprocesses the underwater acoustic communication signals collected by the receiving end.

[0122] The time-domain feature extraction module uses a one-dimensional residual network to extract low-level time-domain features of the preprocessed signal;

[0123] The multi-scale feature fusion and gating enhancement module inputs the extracted low-level temporal features into the constructed multi-resolution SSM fusion module for deep temporal feature processing and fusion;

[0124] The classification output module performs global average pooling or attention-weighted pooling on the fused feature sequence to obtain the global feature vector of the signal and outputs the modulation type identification result.

[0125] Example 3:

[0126] To verify the effectiveness of the present invention, the following experiments and analyses were conducted in this embodiment:

[0127] 1. Experimental Environment and Dataset Setup

[0128] In this embodiment, the experiment uses the publicly available underwater acoustic channel dataset Watermark for verification. This dataset contains nine mainstream modulation types: OFDM, 2FSK, 4FSK, 8FSK, BPSK, QPSK, 8PSK, 16QAM, and 64QAM. Each signal sample is processed into... The feature matrix contains 1024 in-phase (I) sampling points and 1024 orthogonal (Q) sampling points.

[0129] The experiment simulated a complex noisy environment, with a signal-to-noise ratio (SNR) ranging from -10 dB to 18 dB in 2 dB increments. The dataset was strictly divided into training, validation, and test sets in an 8:1:1 ratio, with the training set containing 34,560 samples, and the validation and test sets each containing 4,320 samples. The samples in each set were evenly distributed across different SNRs.

[0130] 2. Experimental Results and Comparative Analysis

[0131] To evaluate the performance of the method of this invention, it was compared with four common modulation recognition methods, including traditional CNN, DenseNet, MobileNetV3, and ResNet18. Detailed comparative experimental results are shown in Table 1 and... Figure 5 As shown.

[0132] Table 1. Accuracy comparison (%) of the method of this invention with existing mainstream models at different signal-to-noise ratios.

[0133]

[0134] According to Table 1 and Figure 5 The experimental results show that, within a signal-to-noise ratio range of -10 dB to 18 dB, the present invention improves the average recognition rate by approximately 12% compared to traditional CNNs, with particularly significant performance improvements under low signal-to-noise ratio conditions of -2 dB and below. Therefore, the performance improvement effect of the present invention is significant.

Claims

1. A method for underwater acoustic signal modulation recognition based on multi-resolution SSM fusion, characterized in that, Includes the following steps: S1: Preprocess the underwater acoustic communication signal collected by the receiving end; S2: Use a one-dimensional residual network to extract the low-level time-domain features of the preprocessed signal; S3: Input the extracted low-level temporal features into the constructed multi-resolution SSM fusion module for deep temporal feature processing and fusion; S4: Perform global average pooling or attention-weighted pooling on the fused feature sequence to obtain the global feature vector of the signal; S5: Based on the global feature vector, output the modulation type recognition result through a fully connected classifier.

2. The underwater acoustic signal modulation recognition method based on multi-resolution SSM fusion according to claim 1, characterized in that, The preprocessing in step S1 includes: Underwater acoustic communication signals contain I and Q components. First, the signal is normalized to eliminate the influence of sampling gain and power differences. Then, the signal is segmented into frames, and the length of each frame is unified to a fixed length. The complex sequence I / Q signal is preprocessed into a dual-channel real number sequence.

3. The underwater acoustic signal modulation recognition method based on multi-resolution SSM fusion according to claim 2, characterized in that, The extraction process in step S2 includes: A1: Transform the dimensions of the input signal through a dimension transformation operation; A2: The transformed signal is passed through an initial convolutional layer, accompanied by batch normalization and ReLU activation function, for preliminary feature extraction and channel expansion; A3: The initial features are downsampled using a max pooling layer to further reduce the sequence length and enhance the robustness of the features; A4: Input the feature sequence processed in step A3 into the ResNet backbone network and output the feature map; A5: The output feature map is transformed again by swapping the channel dimension and sequence length dimension to convert it into the corresponding format.

4. The underwater acoustic signal modulation recognition method based on multi-resolution SSM fusion according to claim 3, characterized in that, The ResNet backbone network in step A4 consists of four stacked ResNet layers, specifically: ResNet Layer 1 contains 3 residual blocks and has 64 output channels; ResNet Layer 2 contains 4 residual blocks, has 128 output channels, and a stride of 2; ResNet Layer 3 contains 6 residual blocks, has 256 output channels, and a stride of 2; ResNet Layer 4 contains 3 residual blocks, has 512 output channels, and a stride of 2.

5. The underwater acoustic signal modulation recognition method based on multi-resolution SSM fusion according to claim 4, characterized in that, In step S3, the multi-resolution SSM fusion module employs three parallel selective state-space model branches: a short-time SSM branch, a medium-time SSM branch, and a long-time SSM branch, wherein: Short-time SSM branch: The convolution dilation stride is set to 1 and the state dimension is set to 8. It is used to extract the transient change features and local mutation information of the signal in a short time. Mid-time SSM branch: The convolution dilation stride is set to 2 and the state dimension is set to 16, which is used to extract the temporal dependency features of the signal in the mid-time range. The long-term SSM branch sets the convolution dilation stride to 4 and the state dimension to 32 to capture the long-term temporal correlation and overall statistical trend of the signal.

6. The underwater acoustic signal modulation recognition method based on multi-resolution SSM fusion according to claim 5, characterized in that, The operation process of the selective state-space model branch in step S3 includes: B1: Input feature X is expanded in dimension by passing it through the input projection layer; B2: The expanded features are split into two along the feature dimension by a feature segmentation operation to obtain X1 and X2; X2 will be used as the gating signal for the final output. B3: Transpose X1 in terms of dimensions to accommodate subsequent causal one-dimensional dilated convolution operations; B4: The convolution result is transposed again and then activated by the SiLU function to generate an intermediate feature u; B5: Input the intermediate feature u into the parameters to generate the projection layer X_Proj; The X_Proj module employs a low-rank decomposition structure, and the output of X_Proj is divided into three time-varying parameters: (Delta), B, and C, after passing through the Softplus activation function, determine the intensity of the SSM state update; B serves as the mapping matrix from input to state; C serves as the mapping matrix from state to output; the X_Proj module also contains a fixed parameter A, which determines the inherent decay characteristics of the SSM. B6: Intermediate feature u, activated by Softplus B, C, and fixed parameter A are fed into the selective scan module. The selective scan module calculates the hidden state h step by step based on the state update equation through a loop calculation. After the loop is completed, the complete scan result is obtained. ; B7: Output of the selective scan module The initial separated X2 is gated by element-wise multiplication; B8: The gated features are passed through the output projection layer to obtain the final output Y of the selective state-space model module.

7. The underwater acoustic signal modulation recognition method based on multi-resolution SSM fusion according to claim 6, characterized in that, The calculation process of the hidden state h in step B6 includes: firstly, using parameters Discretize the continuous parameters to obtain and Then execute the state update equation: ； in, This represents element-wise multiplication. Enter the current time.

8. The underwater acoustic signal modulation recognition method based on multi-resolution SSM fusion according to claim 7, characterized in that, In step S3, the multi-resolution SSM fusion module uses an adaptive fusion network to perform multi-scale feature fusion and gating enhancement. The adaptive fusion network includes fully connected layers, ReLU activation and Dropout layers, and introduces gated linear units (GLUs) for residual gating.

9. The underwater acoustic signal modulation recognition method based on multi-resolution SSM fusion according to claim 8, characterized in that, In step S3, the features processed by GLU are dimensionally adjusted through the output projection layer, and the output of the output projection layer is added to the residual of the output S of the short-time SSM branch.

10. A hydroacoustic signal modulation recognition system based on multi-resolution SSM fusion, characterized in that, For implementing the method of claim 1, the system comprises: The preprocessing module preprocesses the underwater acoustic communication signals collected by the receiving end. The time-domain feature extraction module uses a one-dimensional residual network to extract low-level time-domain features of the preprocessed signal; The multi-scale feature fusion and gating enhancement module inputs the extracted low-level temporal features into the constructed multi-resolution SSM fusion module for deep temporal feature processing and fusion; The classification output module performs global average pooling or attention-weighted pooling on the fused feature sequence to obtain the global feature vector of the signal and outputs the modulation type identification result.

Images