Non-cooperative underwater acoustic communication signal detection and modulation identification integrated method

Abstract

The invention discloses a non-cooperative underwater acoustic communication signal detection and modulation identification integrated method, which comprises the following steps: preprocessing an obtained broadband underwater acoustic signal to obtain a time domain signal sequence; obtaining an initial complex number time-frequency spectrum matrix according to the time-domain signal sequence, and further obtaining an initial time-frequency energy spectrum matrix; giving an initial judgment by segmenting the detection network; when the primary judgment result is that the suspected communication signal exists, performing frequency shift and resampling on the signal according to the center frequency estimation value of the suspected communication signal to obtain a low-sampling-rate signal; inputting the low-sampling-rate signal into a coding network based on a pre-training audio model to extract time sequence features; and inputting the time sequence characteristics into a classification head, outputting a modulation mode judgment result of the communication signal, and giving a final judgment result about whether the communication signal exists or not. According to the method, underwater acoustic communication signal detection and modulation mode identification integrated processing with high robustness and high precision can be realized in a complex underwater acoustic noise environment and a non-cooperative scene.

Description

Technical Field

[0001] This invention belongs to the field of signal processing and relates to underwater acoustic communication signal detection and modulation recognition technology, specifically to an integrated method for non-cooperative underwater acoustic communication signal detection and modulation recognition. Background Technology

[0002] With the rapid development of applications such as marine observation, seabed resource exploration, underwater unmanned platforms, and marine military applications, underwater acoustic communication systems are being used more and more widely in complex marine environments. In non-cooperative scenarios, accurate and robust detection of underwater acoustic communication signals and identification of modulation schemes are fundamental for subsequent communication system analysis, protocol inversion, and spectrum management. In such cases, the intercepting party usually cannot know in advance the carrier frequency, bandwidth, and modulation scheme of the transmitted signal. Therefore, it is necessary to automatically detect and identify the modulation scheme of unknown underwater acoustic communication signals under complex underwater acoustic channel conditions such as strong noise, multipath delay, Doppler spread, and interference signals.

[0003] Most existing modulation identification technologies are based on an idealized prior assumption: that the time position and frequency band of the signal to be identified are known prior to it, and that the signal has been accurately intercepted. Due to this heavy reliance on this prior assumption, existing identification methods often require a separate signal detection module in practical applications. This simple cascaded architecture of "detect first, then identify" has significant drawbacks: on the one hand, complex and non-stationary marine environmental noise easily leads to a large number of false alarm segments generated by the front-end detector; on the other hand, existing modulation classifiers typically only perform closed-set classification for known modulation schemes, lacking the ability to identify and reject interference segments. Once the front end misclassifies a noise segment as a signal and sends it to the back end, the classifier is forced to incorrectly classify it as a certain modulation scheme, resulting in a persistently high false alarm rate for the entire system. Currently, there is a severe lack of an integrated processing mechanism in the field that can automatically extract the signal frequency band from the original broadband signal and complete both "detection and identification."

[0004] In recent years, large-scale pre-trained audio models have demonstrated excellent transfer learning capabilities in tasks such as speech recognition and audio event detection, learning rich time-frequency structures and auditory patterns from large amounts of general audio data. However, there are still few existing technologies that systematically introduce audio pre-trained models into underwater acoustic communication signal processing, especially combining them with multi-resolution time-frequency detection networks to form an integrated detection and modulation recognition solution. A complete technical solution that simultaneously considers detection accuracy, modulation recognition robustness, and engineering deployment feasibility in complex noise environments is still lacking. Summary of the Invention

[0005] Purpose of the invention: In order to solve the technical problem that existing underwater acoustic communication signal identification methods rely heavily on prior knowledge of signal start and end times and frequency bands, and are prone to high false alarm rates, this invention provides an integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification, so as to achieve autonomous detection, automatic estimation of center frequency and highly robust modulation identification of broadband underwater acoustic signals in unknown frequency bands.

[0006] Technical Solution: To achieve the above objectives, this invention provides an integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification, comprising the following steps:

[0007] S1: The acquired broadband underwater acoustic signal is preprocessed to obtain a time-domain signal sequence;

[0008] S2: Perform a short-time Fourier transform on the time-domain signal sequence to obtain the initial complex time-frequency spectrum matrix, and then obtain the initial time-frequency energy spectrum matrix;

[0009] S3: The initial time-frequency energy spectrum matrix is ​​zero-padded and then input into the time-frequency segmentation detection network based on the U-Net structure. The segmentation detection network provides an initial judgment on whether there is a suspected communication signal.

[0010] S4: When the initial decision result of step S3 is that there is a suspected communication signal, the signal is frequency shifted and resampled according to the estimated center frequency value of the suspected communication signal to obtain a low sampling rate signal.

[0011] S5: Input the low sampling rate signal into the coding network based on the pre-trained audio model to extract temporal features;

[0012] S6: Input the temporal features extracted in step S5 into the classification head, which includes a residual convolutional layer, a channel attention module, and a classification layer. Output the modulation mode determination result of the communication signal and give the final determination result of whether the communication signal exists based on the modulation mode identification result.

[0013] Further, step S1 includes: acquiring the broadband underwater acoustic signal x collected by the underwater acoustic receiving device. raw [n], sampling rate f s The original broadband signal is subjected to low-pass filtering, amplitude normalization, and fixed-time-length clipping to obtain the time-domain signal sequence x[n].

[0014] Further, in step S2, the total length of the time-domain signal sequence x[n] is set to S, and a short-time Fourier transform with window length L and step size H is used to obtain the initial complex time-frequency spectrum matrix Z(u,v):

[0015]

[0016] Where L is an even number, representing the window length; u∈{0,1,…,T-1} is the initial time index. ‎ represents the total number of time indices; v∈{0,1,…,F-1} represents the frequency index, F=L / 2+1 represents the total number of frequency indices; w[n] represents the window function, which can be a Hanning window, a Hamming window, or a Blackman window, etc.

[0017] Further processing of Z(u, v) yields the initial time-frequency energy spectrum matrix P(u, v):

[0018] .

[0019] Furthermore, in step S3, the time-frequency segmentation detection network includes a convolutional feature extraction head, an encoder, a decoder, and a signal presence detection branch. The signal presence detection branch is connected to the end of the encoder. The encoder contains K-level downsampling paths, the decoder contains K-level upsampling paths, and the signal presence detection branch is an architecture consisting of a pooling layer and a fully connected layer cascaded together.

[0020] Furthermore, the process of making an initial judgment on the existence of a suspected communication signal through the time-frequency segmentation detection network in step S3 includes:

[0021] The initial time-frequency energy spectrum matrix P(u, v) is zero-padded in both the time and frequency dimensions to obtain matrix X. Matrix X is then input into the encoder, and the encoder outputs the tensor E. (K) A global pooling operation is performed to obtain a global feature vector g. The global feature vector g is then input into a fully connected layer 1, and finally, a normalized probability value p is output through a Sigmoid activation function. When p is greater than or equal to a preset threshold β, a suspected communication signal is determined to exist, triggering subsequent decoder operations. When p is less than the preset threshold β, a suspected communication signal is determined to exist, and forward propagation is terminated.

[0022] Furthermore, the decoder operation in step S3 includes:

[0023] When the decoder operation is triggered, E (K) Enter the K-level upsampling path; E (K) Input a 2D convolutional layer and a Softmax activation function to get a matrix as output. (t,f), where (t,f)∈(0,1) represents the confidence level that an underwater acoustic communication signal exists at the time-frequency unit (t,f); finally, the time and frequency dimensions added by zero padding are removed, thus outputting a matrix consistent with T×F. (t, f).

[0024] Furthermore, the method for obtaining the center frequency estimate in step S4 includes:

[0025] Based on matrix (t, f) is further used to obtain matrix B(t, f), and the detection probability matrix is ​​then calculated according to a preset threshold θ. Binarizing (t, f) yields matrix B(t, f); specifically, given a threshold θ∈(0, 1), the following rules are applied to... Binarize (t,f):

[0026]

[0027] The matrix B(t, f) is traversed, and elements with a value of 1 are extracted. For any element (t, f) with a value of 1 in B(t, f), the surrounding eight elements (t±1, f), (t, f±1), (t±1, f±1), etc., with a value of 1 are merged into the same connected region, until all elements with a value of 1 are included in the corresponding connected region. After dividing into several independent connected regions, the number of elements contained in each connected region is counted, and the connected region with the largest number of elements is selected as the effective time-frequency main region of the target signal. Subsequently, the set of frequency indices corresponding to all elements in the effective time-frequency main region is extracted. The mathematical average of all frequency indices in the set of frequency indices is calculated to obtain the average frequency index f. idx Calculate the center frequency f of the suspected underwater acoustic communication signal. c The specific formula is as follows:

[0028]

[0029] Among them, f s Let x[n] be the sampling rate of the signal;

[0030] For samples identified as containing suspected underwater acoustic communication signals, according to f c The signal is frequency-shifted to adjust the center frequency to f0, resulting in y[n]; y[n] is then sampled at a rate of... Resampling is performed to obtain a low-sampling-rate sequence y'[n].

[0031] Furthermore, in step S5, the encoding network based on the pre-trained audio model includes a convolutional front end, a multi-head self-attention module, and a feedforward network module, and a low-rank adaptation module is inserted into the multi-head self-attention module and the feedforward network module for local parameter fine-tuning.

[0032] Furthermore, in step S6, the classification head is composed of a feature extraction unit, a global pooling layer, a normalization layer, and a fully connected classification layer cascaded together. The feature extraction unit includes a one-dimensional residual convolutional layer and a channel attention module.

[0033] Beneficial Effects: This invention starts with broadband raw underwater acoustic signals and sequentially performs multi-resolution time-frequency feature extraction, signal presence detection, time-frequency segmentation detection, frequency band estimation and down-conversion, audio pre-training coding, and modulation scheme identification, realizing an integrated processing flow of "detection-frequency shifting-resampling-coding-recognition". Compared with existing technologies, this invention has the following advantages:

[0034] 1. By combining multi-resolution short-time Fourier transform with U-Net time-frequency segmentation detection network, time-frequency mask can be automatically learned from broadband raw underwater acoustic signals to achieve signal detection under broadband conditions.

[0035] 2. By tightly coupling frequency conversion and resampling operations with the detection results, an integrated processing link of "detection-frequency conversion-resampling-encoding-recognition" is constructed, enabling the modulation mode recognition network to process only the low-frequency segment containing the target signal, thereby reducing the task difficulty and improving the recognition accuracy.

[0036] 3. By introducing a large-scale audio pre-trained coding network and fine-tuning it using a low-rank adaptation method, it can still maintain strong generalization ability in underwater acoustic scenarios with limited training samples, effectively solving the problem of overfitting of deep models under small sample conditions. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a schematic diagram of the time-frequency segmentation detection network.

[0039] Figure 3 This is a schematic diagram of a binary classification confusion matrix for a signal detection task.

[0040] Figure 4 A schematic diagram of the confusion matrix for modulation scheme identification tasks. Detailed Implementation

[0041] 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.

[0042] Example 1:

[0043] like Figure 1 As shown, this embodiment provides an integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification, including the following steps:

[0044] S1: The acquired broadband underwater acoustic signal is preprocessed to obtain a time-domain signal sequence;

[0045] Acquire broadband underwater acoustic signal x from underwater acoustic receiving equipment raw [n], sampling rate f s The original broadband signal is subjected to low-pass filtering, amplitude normalization, and fixed-time-length clipping to obtain the time-domain signal sequence x[n].

[0046] In this embodiment, amplitude normalization is performed in the following way:

[0047]

[0048] In the formula, x filt [n] represents the low-pass filtered output signal, |·| represents the absolute value operation, and max(·) represents the maximum value operation, in order to eliminate the differences in signal amplitude scale caused by different acquisition conditions.

[0049] The passband of the low-pass filter is 1 Hz to 24 kHz, and the signal sampling rate is f. s The signal frequency is 48 kHz, the signal duration is 3 seconds, and the signal length S = 48000 × 3 = 144000.

[0050] S2: Perform a short-time Fourier transform on the time-domain signal sequence to obtain the initial complex time-frequency spectrum matrix, and then obtain the initial time-frequency energy spectrum matrix;

[0051] Let the total length of the time-domain signal sequence x[n] be S. Using a short-time Fourier transform with a window length L and a step size H, we obtain the initial complex time-frequency spectrum matrix Z(u,v):

[0052]

[0053] Where L is an even number, representing the window length; u∈{0,1,…,T-1} is the initial time index. , where is the total number of time indices; v∈{0,1,…,F-1} is the frequency index, F=L / 2+1 is the total number of frequency indices; w[n] is the window function, which can be Hanning window, Hamming window or Blackman window, etc.;

[0054] Further processing of Z(u, v) yields the initial time-frequency energy spectrum matrix P(u, v):

[0055]

[0056] In this embodiment, the window length L is 1024 points, the step H is 128 points, the total number of time indices T=1118, the total number of frequency indices F=513 at this scale, and the Hanning window function is used for transformation.

[0057] S3: The initial time-frequency energy spectrum matrix is ​​zero-padded and then input into the time-frequency segmentation detection network based on the U-Net structure. The segmentation detection network provides an initial judgment on whether there is a suspected communication signal.

[0058] like Figure 2 As shown, the time-frequency segmentation detection network includes a convolutional feature extraction head, an encoder, a decoder, and a signal presence detection branch. The signal presence detection branch is connected to the end of the encoder. The encoder contains K-level downsampling paths, and the decoder contains K-level upsampling paths. The signal presence detection branch is an architecture consisting of a pooling layer and a fully connected layer cascaded together.

[0059] The process of making an initial determination on the existence of suspected communication signals using a time-frequency segmentation detection network includes:

[0060] The initial time-frequency energy spectrum matrix P(u, v) is zero-padded in both the time and frequency dimensions to obtain matrix X, where the dimensions of both the time and frequency dimensions of X are 2. K The smallest integer that is divisible by and not less than the dimension corresponding to P(u,v), where K is the set number of network downsampling levels;

[0061] Input matrix X into the encoder, and output tensor E from the encoder. (K) A global pooling operation is performed to obtain a global feature vector g. The global feature vector g is then input into a fully connected layer 1, and finally, a normalized probability value p is output through a Sigmoid activation function. When p is greater than or equal to a preset threshold β, a suspected communication signal is determined to exist, triggering subsequent decoder operations. When p is less than the preset threshold β, a suspected communication signal is determined to exist, and forward propagation is terminated.

[0062] When the decoder operation is triggered, E (K) Enter the K-level upsampling path; E (K) Input a 2D convolutional layer and a Softmax activation function to get a matrix as output. (t,f), where (t,f)∈(0,1) represents the confidence level that an underwater acoustic communication signal exists at the time-frequency unit (t,f); finally, the time and frequency dimensions added by zero padding are removed, thus outputting a matrix consistent with T×F. (t, f).

[0063] In this embodiment, the downsampling and upsampling levels of the time-frequency segmentation detection network are K=2, and the specific parameter configurations of each module of the network are as follows:

[0064] The feature extraction header consists of two 2D convolutional layers with kernel size of 3×3, stride of 1, and padding of 1, followed by a ReLU activation layer; the first downsampling block pooling has a stride of 2, followed by two 3×3 convolutional layers and a ReLU activation layer; the second downsampling block pooling has a stride of 2, followed by two 3×3 convolutional layers and a ReLU activation layer, yielding the output tensor E. (2) Subsequently, the signal presence detection branch extracts E. (2) Global average pooling compresses the spatial dimension to obtain a global feature vector g of length 64. Here, fully connected layer 1 consists of a 64×32 weight matrix, a 32-bit bias vector, a ReLU activation function, and a 32×1 weight matrix with scalar bias parameters. After processing by fully connected layer 1, the global feature vector is activated by a Sigmoid function, outputting the existence probability p. The weight matrix in fully connected layer 1 is randomly initialized using a normal distribution with a mean of 0 and a variance of 0.01, and all bias parameters are initialized to 0.

[0065] In this embodiment, a decision threshold β=0.5 is selected. When p≥0.5, a suspected communication signal is detected, triggering the decoder; otherwise, the operation is terminated, and a noise (no suspected communication signal) decision is directly output. If the decoder is triggered, the module enters the first upsampling block. This module first uses a transposed convolutional layer with a kernel size of 2×2 and a stride of 2, followed by two layers of 2D convolution with a kernel size of 3×3, a stride of 1, and padding of 1, and ReLU activation. Then, the module enters the second upsampling block, which first uses a transposed convolutional layer with a kernel size of 2×2 and a stride of 2, followed by two layers of 2D convolution with a kernel size of 3×3, a stride of 1, and padding of 1, and ReLU activation. Then, a 1×1 2D convolutional layer is used to map the number of channels to 2, corresponding to the noise class and the signal class. After the Softmax activation function, a 2×1120×516 probability tensor is obtained. The single probability channel representing the signal class is extracted, and the previous edge zero-padding area is removed, finally restoring a 1118×513 detection probability matrix. (t, f).

[0066] This embodiment requires training the time-frequency segmentation detection network. During the training phase, a training sample set containing communication signals and background noise is constructed, and each sample is labeled with a corresponding binary time-frequency mask matrix M(t, f), where M(t, f) = 1 indicates the presence of an underwater acoustic communication signal at the time-frequency unit (t, f), and M(t, f) = 0 indicates the absence of an underwater acoustic communication signal. The binary time-frequency mask matrix serves as the supervision signal. During training, the time-frequency segmentation detection network outputs a detection probability matrix. (t, f), and construct a segmentation loss function based on the difference between the detection probability matrix and the labeled binary time-frequency mask matrix. The segmentation loss function adopts pointwise binary cross-entropy loss, which is defined as:

[0067]

[0068] Here, T and F represent the time dimension and frequency dimension, respectively. Simultaneously, for the overall existence decision branch, a sample-level supervision label y is constructed. exist ∈{0,1}, where y exist =1 indicates that underwater acoustic communication signals are present in the sample, y exist =1 indicates that the sample contains only background noise. The probability value p corresponding to the overall existence decision branch is used as the loss function during training, which is defined as:

[0069]

[0070] The model training employs a phased training strategy: when training with a binary time-frequency mask matrix M(t, f) as the supervision signal, only the parameters in the encoder and decoder of the time-frequency segmentation detection network are updated, while the parameters of the signal presence detection branch are not updated; when training with sample-level signal presence labels y exist When training the overall existence decision branch using the supervisory signal, the parameters in the encoder and decoder are fixed, and only the parameters of the overall existence decision branch are updated. Through the above step-by-step training strategy, independent optimization of the time-frequency segmentation task and the overall signal existence judgment task is achieved.

[0071] In this embodiment, the Adam optimization algorithm is used to iteratively update the trainable parameters during model training. The training hyperparameters are set as follows: batch size of 32 and initial learning rate of 1×10⁻⁶. -3 The total number of training epochs was 40. In the training of the signal detection branch, the Adam optimization algorithm was also used for parameter updates. The training hyperparameters for fully connected layer 1 were set as follows: batch size remained at 32, and the initial learning rate was set to 1×10⁻⁶. -3 The total number of training rounds is 5.

[0072] S4: When the initial decision result of step S3 is that there is a suspected communication signal, the signal is frequency shifted and resampled according to the estimated center frequency value of the suspected communication signal to obtain a low sampling rate signal.

[0073] Based on matrix (t, f) is further used to obtain matrix B(t, f), and the detection probability matrix is ​​then calculated according to a preset threshold θ. Binarizing (t, f) yields matrix B(t, f); specifically, given a threshold θ∈(0, 1), the following rules are applied to... Binarize (t,f):

[0074]

[0075] The matrix B(t, f) is traversed, and elements with a value of 1 are extracted. For any element (t, f) with a value of 1 in B(t, f), the surrounding eight elements (t±1, f), (t, f±1), (t±1, f±1), etc., with a value of 1 are merged into the same connected region, until all elements with a value of 1 are included in the corresponding connected region. After dividing into several independent connected regions, the number of elements contained in each connected region is counted, and the connected region with the largest number of elements is selected as the effective time-frequency main region of the target signal. Subsequently, the set of frequency indices corresponding to all elements in the effective time-frequency main region is extracted. The mathematical average of all frequency indices in the set of frequency indices is calculated to obtain the average frequency index f. idx Calculate the center frequency f of the suspected underwater acoustic communication signal. c The specific formula is as follows:

[0076]

[0077] Among them, f s Let x[n] be the sampling rate of the signal;

[0078] For samples identified as containing suspected underwater acoustic communication signals, according to f c The signal is frequency-shifted to adjust the center frequency to f0, resulting in y[n]; y[n] is then sampled at a rate of... Resampling is performed to obtain a low-sampling-rate sequence y'[n].

[0079] In this embodiment, the center frequency f0 is set to 4 kHz, and the target sampling rate is... It is 16 kHz.

[0080] S5: Input the low sampling rate signal into the coding network based on the pre-trained audio model to extract temporal features;

[0081] The encoding network based on the pre-trained audio model includes a convolutional front end, a multi-head self-attention module, and a feedforward network module. Low-rank adaptation modules are inserted into the multi-head self-attention module and the feedforward network module to perform local parameter fine-tuning, so as to reduce the amount of parameter updates while maintaining the representational ability of the pre-trained model.

[0082] In this embodiment, the feature matrix obtained after y'[n] is processed by a coding network based on a pre-trained audio model with LoRA modules inserted is h={h 1, h 2,…, h t,…, h T’}, where T' is the number of feature frames, h t ∈ D For dimension The frame-level feature vectors are obtained; finally, the feature matrix h is input into the classification head.

[0083] In this embodiment, the number of feature frames T′=300, the dimension of the frame-level feature vector D=1280, that is, h is a feature matrix of 1280×300.

[0084] S6: Input the temporal features extracted in step S5 into the classification head, which includes a residual convolutional layer, a channel attention module, and a classification layer. Output the modulation mode determination result of the communication signal and give the final determination result of whether the communication signal exists based on the modulation mode identification result.

[0085] The classification head consists of L cascaded feature extraction units, a global pooling layer, a normalization layer, and a fully connected classification layer. Each feature extraction unit comprises a cascaded one-dimensional residual convolutional layer and a channel attention module. The one-dimensional residual convolutional layer consists of batch normalization, a non-linear activation function, and a one-dimensional convolutional layer. The input features are processed through batch normalization, activation functions, and one-dimensional convolution operations, and then residually added to the original input features. The channel attention module contains a global average pooling layer and cascaded fully connected layers, activated by the sigmoid function to generate channel weight vectors. Finally, these channel weight vectors are used to perform channel-wise multiplication and weighting on the feature matrix output by the one-dimensional residual convolutional layer. After iteration through L feature extraction units, the deep feature F is obtained. (L) F (L) ∈ D×T′ Subsequently, a global pooling layer is used to process the deep features F along the T' dimension. (L) Aggregation is performed to obtain a fixed-length global feature vector v∈ D The vector *v* is then input to a normalization layer to standardize it, resulting in a fixed-length vector *z* with a mean of 0 and a variance of 1. *z* is then input to a fully connected layer, which linearly maps *z* to an N-dimensional class space, outputting an unnormalized classification score vector *s* = [s1, s2]. 2,…, s N ], where N represents the total number of categories, and the Nth component corresponds to the interference category. The score vector is normalized by inputting s into the Softmax function to obtain the posterior probability vector p=[p1, p... 2,…, p N Based on the posterior probability vector The final decision is made based on the maximum value P in the posterior probability vector. max and its corresponding category label C max The final modulation scheme identification result R is determined by the following decision logic, where γ is a preset identification confidence threshold:

[0086]

[0087] When the identification result R is interference, the output identification result is interference; when the identification result R is a modulation of a certain modulation mode, the output is the corresponding modulation mode and the corresponding center frequency, thus completing the detection and modulation mode identification of non-cooperative underwater acoustic communication signals.

[0088] The specific parameter configuration of the classification head in this embodiment is as follows:

[0089] The input features of the one-dimensional residual convolutional layer are sequentially processed through batch normalization, GELU activation, and a one-dimensional convolution operation with a kernel size of 3, a stride of 1, and zero padding of 1, before being added to the original input as residuals. The channel attention module first performs global average pooling along the T′ direction on the feature matrix output from the one-dimensional residual convolutional layer. Then, channel weights are extracted through fully connected layers. The fully connected layer contains a 48×1280 weight matrix and a 48×1 bias vector. The weight matrix is ​​randomly initialized using a normal distribution with a mean of 0, and the bias vector is initialized to 0 and processed by the ReLU activation function. The fully connected layer contains a 1280×48 weight matrix and a 1280×1 bias vector. The weight matrix and bias vector are initialized in the same way as in the fully connected layer, and then activated by the Sigmoid function to obtain a 1280×1 channel weight vector. Finally, this weight vector is used to perform channel-wise multiplication and weighting on the feature matrix output from the one-dimensional residual convolutional layer to obtain the enhanced feature matrix. After iteration through three layers of feature extraction units, the deep feature matrix F is obtained. (3) A 1280-dimensional fixed-length global feature vector v is obtained by aggregating the data over the time dimension using a global pooling layer. This vector is then input to a normalized layer and transformed into a standardized vector z with a mean of 0 and a variance of 1. The vector then enters a fully connected layer. In this embodiment, the total number of categories N=9. The weight matrix of the fully connected layer has a size of 9×1280, and the bias vector has a size of 9×1. The linear mapping outputs a 9-dimensional classification score vector s, which, after softmax activation, yields a posterior probability vector p for the 9 categories. In the final decision logic, a recognition confidence threshold γ=0.5 is set, i.e.:

[0090]

[0091] In this embodiment, the modulation scheme identification module needs to be trained. During the training phase, the one-hot encoded vector y=[y1, y2] of the actual modulation scheme label is used. 2,…, y N [] is the supervisory signal, using cross-entropy loss:

[0092]

[0093] The cross-entropy loss function is used as a unified optimization objective to update the network parameters. Specifically, during backpropagation, the gradient of the cross-entropy loss is not only used to update the network parameters of each layer in the backend classification head, but also propagated back through the computation graph to the low-rank adaptation module inserted in the frontend audio pre-trained encoding network, thereby jointly fine-tuning the low-rank adaptation parameters in the LoRA module.

[0094] In this embodiment, the training hyperparameters are set as follows: batch size is 16, and initial learning rate is 1×10⁻⁶. -4 The maximum number of training epochs is set to 50. During backpropagation, the cross-entropy loss between the 9-dimensional posterior probability vector and the true one-hot encoded label vector y is calculated. The system calculates and updates the parameters of the classifier head and the adaptation matrix of the LoRA module in the pre-trained model. When the recognition accuracy on the validation set does not improve for 5 consecutive epochs of training, the early stopping mechanism is triggered to terminate training and save the optimal model.

[0095] Example 2:

[0096] To verify the effectiveness of the method of the present invention, the following experiments and data analysis were conducted in this embodiment:

[0097] This embodiment conducts experimental verification and data analysis based on steps S1 to S6 of Embodiment 1. In the experiment, a dataset containing signals with various modulation schemes and background noise was constructed. The modulation schemes included 2FSK, 4FSK, 8FSK, BPSK, QPSK, OFDM, LFM, and DSSS. The input signal sampling rate was 48 kHz, and the signal duration was 3 s.

[0098] First, the input signal is analyzed in time and frequency using the short-time Fourier transform in step S2, with a window length of 1024 points and a step size of 128 points, to obtain the initial time-frequency energy spectrum matrix. Then, the initial time-frequency energy spectrum matrix is ​​input into the time-frequency segmentation detection network based on the U-Net structure in step S3 to obtain the initial judgment result of whether there is a suspected communication signal.

[0099] Figure 3 The initial classification confusion matrix is ​​given. This confusion matrix represents the correspondence between the "true class" and the "predicted class," and is derived from... Figure 3 As can be seen, in the test of this embodiment, all 6,000 samples whose true category was noise were correctly identified as noise, and all 6,000 samples whose true category was suspected communication signal were correctly identified as suspected communication signal. That is, the number of false alarm samples was 0, the number of missed detection samples was 0, and the accuracy of binary classification decision reached 100%.

[0100] For samples identified as having suspected communication signals in step S3, further center frequency estimation, frequency shifting, and resampling are performed according to step S4 to obtain a low sampling rate signal with a sampling rate of 16 kHz. Then, the low sampling rate signal is input to the modulation mode identification module in steps S5 and S6, and the corresponding modulation mode decision result is output.

[0101] Figure 4 The confusion matrix of the modulation scheme identification results is shown. Figure 4 The confusion matrix in the model also represents the correspondence between the true and predicted categories for each modulation scheme. The values ​​on the diagonal represent the number of correctly identified samples, while the values ​​off-diagonal represent the number of samples misclassified as other modulation schemes. Figure 4 As can be seen, in the modulation mode identification experiment of this embodiment, a total of 6,000 test samples were identified. Among them, 2FSK, 4FSK, 8FSK and LFM had better identification results, with identification accuracy rates of 95.90%, 93.03%, 98.08% and 100%, respectively. Meanwhile, some PSK type signals had certain similarities in time and frequency characteristics, resulting in a certain degree of mutual confusion.

[0102] In summary, the method of the present invention can effectively determine suspected communication signals and effectively distinguish various modulation methods of underwater acoustic communication signals.

Claims

1. A method for integrated detection and modulation identification of non-cooperative underwater acoustic communication signals, characterized in that, Includes the following steps: S1: The acquired broadband underwater acoustic signal is preprocessed to obtain a time-domain signal sequence; S2: Perform a short-time Fourier transform on the time-domain signal sequence to obtain the initial complex time-frequency spectrum matrix, and then obtain the initial time-frequency energy spectrum matrix; S3: The initial time-frequency energy spectrum matrix is ​​zero-padded and then input into the time-frequency segmentation detection network based on the U-Net structure. The segmentation detection network provides an initial judgment on whether there is a suspected communication signal. S4: When the initial decision result of step S3 is that there is a suspected communication signal, the signal is frequency shifted and resampled according to the estimated center frequency value of the suspected communication signal to obtain a low sampling rate signal. S5: Input the low sampling rate signal into the coding network based on the pre-trained audio model to extract temporal features; S6: Input the time-series features extracted in step S5 into the classification head, output the modulation mode judgment result of the communication signal, and give the final judgment result of whether the communication signal exists based on the modulation mode recognition result.

2. The integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification according to claim 1, characterized in that, Step S1 includes: acquiring the broadband underwater acoustic signal x collected by the underwater acoustic receiving device. raw [n], sampling rate f s The original broadband signal is subjected to low-pass filtering, amplitude normalization, and fixed-time-length clipping to obtain the time-domain signal sequence x[n].

3. The integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification according to claim 2, characterized in that, In step S2, the total length of the time-domain signal sequence x[n] is set to S. The short-time Fourier transform with window length L and step H is used to transform the sequence to obtain the initial complex time-frequency spectrum matrix Z(u,v). ； Where L is an even number, representing the window length; u∈{0,1,…,T-1} is the initial time index. , where is the total number of time indices; v∈{0,1,…,F-1} is the frequency index, F=L / 2+1 is the total number of frequency indices; w[n] is the window function; Further processing of Z(u, v) yields the initial time-frequency energy spectrum matrix P(u, v): 。 4. The integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification according to claim 3, characterized in that, The time-frequency segmentation detection network in step S3 includes a convolutional feature extraction head, an encoder, a decoder, and a signal presence detection branch. The signal presence detection branch is connected to the end of the encoder. The encoder contains K-level downsampling paths, and the decoder contains K-level upsampling paths. The signal presence detection branch is an architecture consisting of a pooling layer and a fully connected layer cascaded together.

5. The integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification according to claim 4, characterized in that, The process of making an initial judgment on the existence of a suspected communication signal through the time-frequency segmentation detection network in step S3 includes: The initial time-frequency energy spectrum matrix P(u, v) is zero-padded in both the time and frequency dimensions to obtain matrix X. Matrix X is then input into the encoder, and the encoder outputs the tensor E. (K) A global pooling operation is performed to obtain a global feature vector g. The global feature vector g is then input into a fully connected layer 1, and finally, a normalized probability value p is output through a Sigmoid activation function. When p is greater than or equal to a preset threshold β, a suspected communication signal is determined to exist, triggering subsequent decoder operations. When p is less than the preset threshold β, a suspected communication signal is determined to exist, and forward propagation is terminated.

6. The integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification according to claim 5, characterized in that, The decoder operation process in step S3 includes: When the decoder operation is triggered, E (K) Enter the K-level upsampling path; E (K) Input a 2D convolutional layer and a Softmax activation function to get a matrix as output. (t,f), where (t,f)∈(0,1) represents the confidence level that an underwater acoustic communication signal exists at the time-frequency unit (t,f); finally, the time and frequency dimensions added by zero padding are removed, thus outputting a matrix consistent with T×F. (t, f).

7. The integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification according to claim 6, characterized in that, The method for obtaining the center frequency estimate in step S4 includes: Based on matrix (t, f) is further used to obtain matrix B(t, f), and the detection probability matrix is ​​then calculated according to a preset threshold θ. Binarizing (t, f) yields matrix B(t, f); The matrix B(t, f) is traversed, and elements with a value of 1 are extracted. For any element (t, f) with a value of 1, its surrounding elements with a value of 1 are merged into the same connected region until all elements with a value of 1 are included in the corresponding connected region. After dividing the matrix into several independent connected regions, the number of elements contained in each connected region is counted, and the connected region with the largest number of elements is selected as the effective time-frequency main region of the target signal. Subsequently, the set of frequency indices corresponding to all elements in the effective time-frequency main region is extracted. The mathematical average of all frequency indices in the set of frequency indices is calculated to obtain the average frequency index f. idx ; Calculate the center frequency f of the suspected underwater acoustic communication signal c The specific formula is as follows: ； Among them, f s Let x[n] be the sampling rate of the signal; For samples identified as containing suspected underwater acoustic communication signals, according to f c The signal is frequency-shifted to adjust the center frequency to f0, resulting in y[n]; y[n] is then sampled at a rate of... Resampling is performed to obtain a low-sampling-rate sequence y'[n].

8. The integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification according to claim 7, characterized in that, In step S5, the encoding network based on the pre-trained audio model includes a convolutional front end, a multi-head self-attention module, and a feedforward network module. A low-rank adaptation module is inserted into the multi-head self-attention module and the feedforward network module to fine-tune local parameters.

9. The integrated method for non-cooperative underwater acoustic communication signal detection and modulation identification according to claim 8, characterized in that, In step S6, the classification head is composed of a feature extraction unit, a global pooling layer, a normalization layer, and a fully connected classification layer cascaded together. The feature extraction unit includes a one-dimensional residual convolutional layer and a channel attention module.

Images