An open set hydrophone communication modulation recognition method based on a multi-path residual network
By using feature filtering and multi-threshold discrimination methods based on multi-path residual networks, the problem of insufficient recognition of underwater acoustic communication modulation recognition technology in open set scenarios is solved, achieving high-precision modulation recognition and detection of unknown modulation methods, thus improving the adaptability and accuracy of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INST OF ACOUSTICS CHINESE ACAD OF SCI
- Filing Date
- 2025-11-25
- Publication Date
- 2026-08-04
AI Technical Summary
Existing underwater acoustic communication modulation recognition technologies lack model generalization and recognition accuracy, and are mostly limited by the closed set assumption, making them difficult to adapt to the complex underwater acoustic environment of open set scenarios.
A method based on multi-path residual networks is adopted. In the feature selection stage, t-SNE dimensionality reduction and Calinski-Harabasz exponent are used to select modulation features with high discriminative power. In the identification stage, a multi-threshold discrimination framework is introduced to achieve the identification of unknown modulation modes.
It significantly improves the model's generalization ability and recognition accuracy, enabling it to accurately identify known and unknown modulation schemes in adversarial underwater acoustic environments and adapt to non-cooperative communication in complex underwater acoustic channels.
Smart Images
Figure CN121770948B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of underwater acoustic signal processing, and in particular to an open set underwater acoustic communication modulation identification method based on a multi-path residual network. Background Technology
[0002] Underwater acoustic communication modulation identification technology refers to the technique of identifying the modulation scheme of received underwater acoustic communication signals through feature analysis and pattern determination. The accuracy of the modulation identification results directly determines the performance of subsequent processing modules such as demodulation and decoding, making it an indispensable key technology in underwater acoustic communication systems with significant strategic value in both military and civilian fields. As underwater communication networks develop towards heterogeneity and intelligence, underwater acoustic communication modulation technology has become a necessary condition for achieving multi-standard compatibility, dynamic resource optimization, and intelligent network management. Therefore, developing highly robust and adaptable underwater acoustic modulation identification technology is of great significance for enhancing military underwater combat capabilities and promoting the development of the marine economy.
[0003] Machine learning methods have offered new insights into underwater acoustic modulation recognition due to their powerful ability to learn high-dimensional features. Deep neural networks, by learning nonlinear features such as time-frequency distribution and higher-order cumulants, have significantly improved the robustness of recognition under low signal-to-noise ratio conditions. However, existing research still faces two key challenges: first, insufficient model generalization and recognition accuracy; second, most mainstream research is limited to the closed-set assumption, while typical applications such as underwater acoustic countermeasures and underwater reconnaissance are essentially open-set problems, requiring models to detect unknown modulation methods. Therefore, there is an urgent need to explore an underwater acoustic communication modulation recognition method with higher recognition accuracy and adaptability to open-set scenarios to address the challenges posed by non-cooperative communication in complex underwater acoustic environments. Summary of the Invention
[0004] To address the aforementioned problems, this invention provides an open collection method for underwater acoustic communication modulation identification based on a multi-path residual network.
[0005] According to a first aspect, the present invention provides a method for filtering modulation identification features in underwater acoustic communication, the method comprising the following steps:
[0006] Acquire an underwater acoustic communication modulation dataset, which includes modulated underwater acoustic communication signals and their modulation methods;
[0007] The modulated underwater acoustic communication signal is preprocessed to obtain various modulation features;
[0008] A sample set is constructed based on the dataset. The sample set includes multiple types of samples, each type of sample corresponding to a modulation feature. Each sample uses a certain modulation feature of a certain underwater acoustic communication signal as its sample feature and the modulation method of that underwater acoustic signal as its label.
[0009] The feature extraction network in the pre-trained multi-path residual network model is used to extract features from various types of samples in the sample set to obtain high-dimensional feature vectors for each type of sample.
[0010] The high-dimensional feature vector is reduced in dimensionality using the t-SNE technique;
[0011] The Calinski-Harabasz exponent of the t-SNE dimensionality reduction results for each class of samples is calculated. Based on the calculation results, one or more modulation features are selected for modulation identification in underwater acoustic communication.
[0012] In some embodiments, the step of calculating the Calinski-Harabasz index of the t-SNE dimensionality reduction result for each class of samples and selecting one or more modulation features based on the calculation result specifically includes: selecting one or more modulation features for the corresponding class of samples whose Calinski-Harabasz index is higher than a certain threshold.
[0013] In some embodiments, the preprocessing of the modulated underwater acoustic communication signal to obtain multiple modulation features specifically includes: using short-time Fourier transform to preprocess the modulated underwater acoustic communication signal to obtain multiple modulation features of the underwater acoustic communication signal.
[0014] In some embodiments, the types of modulation features specifically include: time-domain waveform features, time-spectrum features, quadratic spectrum features, and quartic spectrum features.
[0015] According to a second aspect, the present invention provides a method for identifying modulation in underwater acoustic communication, the method comprising:
[0016] Acquire the underwater acoustic communication signal to be identified;
[0017] The underwater acoustic communication signal is preprocessed to obtain one or more modulation features of the underwater acoustic communication signal; the type of the modulation feature is determined by the feature screening method described in the first aspect.
[0018] The modulation features are input into a pre-trained multi-path residual network model to obtain the activation vector of the underwater acoustic communication signal to be identified.
[0019] The distance between the activation vector of the underwater acoustic communication signal to be identified and the average activation vector of each modulation scheme is calculated, and the distance is compared with the distance threshold of the corresponding modulation scheme. If the distance between the activation vector of the underwater acoustic communication signal and the average activation vector of each modulation scheme is greater than the corresponding distance threshold, the modulation scheme of the underwater acoustic communication signal is classified as unknown. The average activation vector of each modulation scheme is calculated by the model based on the activation vector of the training samples of each modulation scheme.
[0020] If the distance between the activation vector of the underwater acoustic communication signal to be identified and the average activation vector of a certain modulation scheme is less than the distance threshold of that modulation scheme, then the modulation scheme of the underwater acoustic communication signal is classified as a known category; for underwater acoustic communication signals classified as known categories, the optimized multi-path residual network model is used to determine its modulation scheme.
[0021] In some embodiments, if the modulation features include multiple types, the modulation features are input into a pre-trained multi-path residual network model to obtain the activation vector of the underwater acoustic communication signal. Specifically, this includes: using the feature extraction network in the multi-path residual network model to extract features from different types of modulation features respectively; the model fuses the feature extraction results of different types of modulation features; and further extracts the fused features to obtain the activation vector of the underwater acoustic communication signal.
[0022] The underwater acoustic communication modulation identification method provided by this invention uses t-SNE dimensionality reduction analysis to screen multivariate modulation identification features, and uses modulation features with good class discrimination for modulation identification, significantly improving the model's generalization ability and identification accuracy. Furthermore, the underwater acoustic communication modulation identification method provided by this invention has the ability to identify unknown modulation schemes. In adversarial underwater acoustic environments, the underwater acoustic communication modulation identification method provided by this invention combines high accuracy in closed-set identification with adaptability to open-set scenarios, providing an effective solution for non-cooperative communication analysis in complex underwater acoustic channels. Attached Figure Description
[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 A schematic diagram of the feature screening stage of the underwater acoustic communication modulation recognition method provided by the present invention;
[0025] Figure 2 This is a schematic diagram of the open set identification stage of the underwater acoustic communication modulation identification method in an embodiment of the present invention;
[0026] Figure 3 The t-SNE diagrams and CH indices for each modulation feature in this embodiment of the invention are shown below.
[0027] Figure 4 This is a performance comparison chart between the underwater acoustic communication modulation identification method provided by the present invention and traditional methods. Detailed Implementation
[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings.
[0029] This invention proposes an open-set underwater acoustic communication modulation identification method based on a multi-path residual network. The method includes a feature selection stage and an identification stage. In the feature selection stage, the invention systematically selects modulation features from multiple domains, including time-frequency and statistical domains, to obtain modulation features with high discriminative power. Based on these features, a multi-path residual network model with good identification performance is constructed. The identification stage performs modulation identification based on the modulation features obtained in the feature selection stage. This stage introduces an open-set identification framework based on multi-threshold discrimination, enabling the model to identify unknown modulation modes and adapt to open-set scenarios. This method integrates a feature selection mechanism and adapts to open-set scenarios, combining high accuracy of closed-set identification with adaptability to open-set scenarios in adversarial underwater acoustic environments, providing an effective solution for non-cooperative communication analysis in complex underwater acoustic channels.
[0030] S1, Feature Selection Stage
[0031] Figure 1 The overall framework of the feature selection stage is illustrated. In this stage, the underwater acoustic communication modulated signal samples (received signals) are first preprocessed to extract various modulation feature samples (such as time-domain waveform features, time-spectrum features, quadratic spectrum features, and quartic spectrum features). Next, a trained feature extraction network is used to extract features from the modulation feature samples. Then, t-SNE dimensionality reduction analysis and CH exponent calculation are performed on the feature extraction results for each type of modulation feature sample. Modulation features are selected based on the CH exponent calculation results and used as input to the multi-path residual network model, resulting in an optimized multi-path residual network model. The feature selection stage specifically includes the following steps:
[0032] S11. Obtain underwater acoustic communication modulation dataset, which includes modulated underwater acoustic communication signals and their modulation methods.
[0033] The underwater acoustic communication system generates a modulated signal at the transmitting end, with nine modulation methods including 2ASK, 4ASK, 2FSK, 4FSK, 8FSK, 2PSK, 4PSK, 8QAM, and OFDM. The modulated underwater acoustic communication signal is received at the receiving end of the system. Receive signal It can be represented as:
[0034] (1)
[0035] in, Represents signal convolution. , This represents additive white Gaussian noise. This represents the impulse response of the underwater acoustic channel.
[0036] The modulated underwater acoustic communication signal (i.e., the received signal) The data and their corresponding modulation methods are stored in the underwater acoustic communication modulation dataset.
[0037] S12. Preprocess the centralized modulated underwater acoustic communication signal to obtain various modulation characteristics of the underwater acoustic communication signal.
[0038] For example, the types of modulation features include: time-domain waveform features, time-spectrum features, quadratic spectrum features, and quartic spectrum features.
[0039] In this step, different types of modulation features are obtained by performing different preprocessing on the modulated underwater acoustic communication signal.
[0040] For example, the modulated underwater acoustic communication signal can be analyzed using a short-time Fourier transform. Preprocessing is performed to obtain the underwater acoustic communication signal. Time-spectral characteristics As shown in formula (2);
[0041] (2)
[0042] in, Let w[n] represent the original signal, w[n] represent the window function, m represent the frame index, k represent the frequency index, and N represent the number of points used in the Fourier transform calculation. Similarly, the normalized time-domain waveform characteristics, quadratic spectrum characteristics, and quartic spectrum characteristics of the underwater acoustic signal can be obtained through preprocessing, in preparation for constructing the underwater acoustic communication modulation training sample set and the underwater acoustic communication modulation verification sample set.
[0043] S13. Construct a training sample set and a verification sample set for underwater acoustic communication modulation based on the dataset.
[0044] The underwater acoustic communication modulation training sample set and the underwater acoustic communication modulation verification sample set use a certain modulation feature of a certain underwater acoustic communication signal as the sample feature and the modulation mode of the underwater acoustic signal as the label; each sample set includes multiple types of samples, and each type of sample corresponds to a modulation feature.
[0045] S14. Train the multi-path residual network model using the training sample set;
[0046] The purpose of this step is to improve the model's feature extraction capabilities through training, in preparation for subsequent feature selection.
[0047] In one embodiment, the structure of the multi-path residual network model is as follows: Figure 1As shown, it includes a feature extraction network and a fully connected layer. The feature extraction network is divided into two types: one-dimensional feature extraction networks and two-dimensional feature extraction networks, which are used to extract features from different types of modulation features, respectively. The structures of the one-dimensional and two-dimensional feature extraction networks are shown below. Figure 1 (b) and Figure 1 As shown in (c), the one-dimensional feature extraction network includes one one-dimensional convolutional layer, one one-dimensional batch normalization layer, followed by an activation function and four one-dimensional residual layers; the two-dimensional feature extraction network includes one convolutional layer, one two-dimensional batch normalization layer, followed by an activation function, one pooling layer and four two-dimensional residual layers.
[0048] The model training process includes: inputting different types of modulation feature samples into a multi-path residual network model; one-dimensional features (such as time-domain waveform features, quadratic spectrum features, and quartic spectrum features) into a one-dimensional feature extraction network, which extracts features from each layer of the one-dimensional feature extraction network and transforms them into corresponding high-dimensional feature vectors; two-dimensional features (such as time-spectrum features) into a two-dimensional feature extraction network, which extracts features from each layer of the two-dimensional feature extraction network and obtains corresponding high-dimensional feature vectors; the model inputs these high-dimensional feature vectors into subsequent fully connected layers for further feature extraction to obtain activation vectors of the samples, and derives the modulation mode prediction result based on the activation vectors; the loss function is calculated by comparing the prediction result with the true labels of the training samples; subsequently, the gradient of the loss with respect to all model parameters is calculated using the backpropagation algorithm, and the model parameters are updated based on the gradient, thereby gradually optimizing the network's feature extraction and classification capabilities until the model loss converges and the performance stabilizes.
[0049] The above describes the preparatory work for sample and model training in the feature selection stage. The core process of feature selection is described below. The purpose of feature selection is to select one or more modulation features with high discriminative power from multiple modulation features as model input, thereby solving the problem of redundant features interfering with model generalization and improving the model's recognition accuracy. This invention achieves feature selection by analyzing the high-dimensional features extracted by the model from various samples. In the analysis process, t-SNE technology is first used to reduce the dimensionality and visualize the high-dimensional features extracted by the model. Then, based on the t-SNE results, the CH index of each type of modulation feature sample is calculated to determine the feature clustering and select high-quality features. For example... Figure 1 As shown in (a), the core process of feature selection specifically includes:
[0050] S15. Use the feature extraction network in the trained multi-path residual network model to extract features from various types of samples in the validation sample set to obtain high-dimensional feature vectors for each type of sample.
[0051] The various modulation feature samples in the validation sample set are input into different types of feature extraction networks for feature extraction, resulting in high-dimensional feature vectors for each type of modulation feature sample. After obtaining the high-dimensional feature vectors of all samples, all extracted high-dimensional feature vectors are input into the feature filtering module, and the high-dimensional features are visualized and analyzed using the t-SNE technique of nonlinear dimensionality reduction.
[0052] S16. Use t-SNE technology to reduce the dimensionality of the high-dimensional feature vector to obtain a low-dimensional representation of the high-dimensional feature vector;
[0053] t-SNE (t-Distributed Stochastic Neighbor Embedding) is a non-linear dimensionality reduction technique for visualizing high-dimensional data. It transforms the similarity probabilities between high-dimensional data points into joint probabilities in a low-dimensional (typically 2- or 3-dimensional) space and minimizes the Kullback-Leibler divergence (KL divergence) between these two probability distributions, thus preserving local neighborhood relationships of high-dimensional data in the low-dimensional space. t-SNE ensures that points close in the high-dimensional space remain close in the low-dimensional mapping, forming clear and compact clusters, while simultaneously effectively pushing dissimilar points apart in the low-dimensional space, making cluster boundaries clearer. The core principle of t-SNE is modeling the similarity between data points through probability distributions; its key hyperparameters include perplexity, learning rate, and number of iterations. t-SNE technology is suitable for revealing clustering patterns in high-dimensional data and is widely used in the field of visualizing high-dimensional features output by neural networks. Therefore, this invention uses t-SNE technology to perform dimensionality reduction analysis on the high-dimensional features extracted by the network model, revealing the clustering effects of various features. The process of obtaining the low-dimensional representation of the high-dimensional feature vector using t-SNE technology in this step is described in detail below.
[0054] First, in high-dimensional space, the Gaussian distribution is used to compute the result at a given data point x. i Under the condition of selecting data point x j The conditional probability of being its neighbor As shown in formula (3):
[0055] (3)
[0056] in and Let i and j be the high-dimensional feature vectors. The square of the Euclidean distance between two high-dimensional feature vectors. It is a point The local bandwidth. Among them The perplexity can be adjusted using a binary search method to match the user-defined perplexity. Perplexity refers to how many nearest neighbor data points each data point considers when selecting its neighbors. It is used to control the balance between local and global structures and is one of the key hyperparameters of the t-SNE technique.
[0057] probability Capable of reflecting x in higher-dimensional space i With x j The degree of similarity.
[0058] Next, the corresponding data points are calculated using the Student-t distribution with 1 degree of freedom in the low-dimensional space. and similarity As shown in formula (4):
[0059] (4)
[0060] in, and For high-dimensional data points x i and x j Mapping points in low-dimensional space.
[0061] Finally, the KL divergence of the similarity distribution in the high-dimensional and low-dimensional spaces is minimized. The KL divergence C can be obtained from formula (5):
[0062] (5)
[0063] in, and Let these be the probability distributions in high-dimensional space and low-dimensional space, respectively, which are respectively derived from... and Obtain; joint probability To ensure symmetry, N is the total number of feature vector data points in the high-dimensional space. Using the gradient formula shown in equation (6), the low-dimensional representation is iteratively optimized via gradient descent. To minimize the KL divergence of similarity distributions in high-dimensional and low-dimensional spaces:
[0064] (6)
[0065] This yields a low-dimensional representation of the high-dimensional feature vectors of the validation set samples. Based on this low-dimensional representation, the Calinski-Harabasz index (CH index) for each modulation feature in the validation set is calculated to evaluate the clustering performance of each modulation feature.
[0066] S17. Calculate the CH index of the t-SNE dimensionality reduction result for each type of sample, and select one or more modulation features as modulation features based on the calculation results for underwater acoustic communication modulation identification.
[0067] The Calinski-Harabasz index (CH index) is an internal validation metric used to evaluate clustering analysis results. The CH index assesses clustering quality by calculating the ratio of the inter-cluster scatter matrix (i.e., the dispersion between clusters) to the intra-cluster scatter matrix (i.e., the compactness within clusters). A higher CH index value indicates that signals from different modulation schemes form clearly separated and compact clusters in the feature space. This suggests that the modulation features used are highly effective in distinguishing modulation schemes and can effectively differentiate between them. The CH index comprehensively considers both inter-cluster separation and intra-cluster compactness, thus effectively reflecting the quality of clustering. This invention treats each modulation scheme as a cluster and uses the CH index to evaluate the clustering effect of each modulation feature separately. The modulation feature with the best clustering effect is selected as the final model input based on the CH index. Systematic optimization of multi-dimensional modulation features in the time-frequency domain and statistical domain based on the CH index avoids interference from time-restricted redundant features on the model's generalization ability, thereby improving recognition accuracy.
[0068] The above provides a brief introduction to the CH index. The following section details the specific implementation process of this step.
[0069] First, the CH index needs to be calculated for each type of modulation feature sample. The calculation process of the CH index is introduced below, taking time-domain waveform feature samples as an example.
[0070] The low-dimensional representation dataset X of time-domain waveform feature samples contains n low-dimensional representations. The modulation schemes of the validation sample set include nine types: 2ASK, 4ASK, 2FSK, 4FSK, 8FSK, 2PSK, 4PSK, 8QAM, and OFDM. Therefore, the dataset is divided into K=9 clusters according to the modulation scheme, where each cluster... There is Each cluster contains 10 samples. The intra-cluster scatter matrix is used to measure the dispersion of samples within each cluster. It can be obtained from formula (7):
[0071] (7)
[0072] in, It is a cluster The i-th sample within; It is a cluster The centroid of the cluster, which is the mean of all samples within the cluster, is defined by the following formula: .
[0073] The inter-cluster scatter matrix measures the degree of dispersion among clusters. It can be obtained from formula (8):
[0074] (8)
[0075] Where c is the mean of the entire dataset. .
[0076] Based on the intra-cluster scatter matrix and inter-cluster scatter matrix The CH index of the dataset can be obtained, as shown in formula (9):
[0077] (9)
[0078] After calculating the CH index of the low-dimensional representation of each type of modulation feature sample using the above method, one or more modulation features with prominent CH values are selected as modulation features.
[0079] A higher CH index value indicates better clustering results. Therefore, when selecting modulation features based on the CH index calculation results, one or more modulation features with a CH index higher than a certain threshold can be selected as modulation features, or one or more modulation features with the highest CH index among all modulation features can be selected.
[0080] The above describes the feature selection stage of the underwater acoustic communication modulation identification method provided by this invention. This stage uses t-SNE dimensionality reduction analysis to select multivariate modulation identification features, using modulation features with good discriminative power for modulation identification, which significantly improves the model's generalization ability and identification accuracy. The identification stage is described below. The underwater acoustic communication modulation identification method provided by this invention employs an open set identification method based on multi-threshold discrimination in the identification stage. This allows the model to identify unknown modulation modes. The open set identification method needs to be executed based on the modulation features obtained in the feature selection stage. This is because the modulation features obtained in the feature selection stage can effectively distinguish different modulation modes, making the features of the same modulation mode more concentrated in the feature space. This characteristic is beneficial for open set identification to complete boundary determination in the feature space, which allows the model to more accurately identify unknown modulation modes.
[0081] S2, Identification Phase
[0082] The identification phase is executed based on the optimized multi-path residual network model obtained in the feature selection phase, and its architecture is as follows: Figure 2 As shown, firstly, the optimized multi-path residual network model is trained using the training set. Based on the activation vectors of the training set samples, the average activation vector for each modulation scheme is calculated. Using these as cluster centers for each category, a logits space is constructed. When identifying a test set containing modulated signals of unknown class, the model can determine whether the modulated signal belongs to the unknown class based on the distance between the activation vector of the modulated signal and the average activation vector of each known modulation scheme, thus achieving open set identification.
[0083] The identification phase includes the following steps:
[0084] S21. Acquire the underwater acoustic communication signal to be identified;
[0085] S22. The underwater acoustic communication signal is preprocessed to obtain one or more modulation features of the underwater acoustic communication signal; the type of modulation feature is determined by the feature selection stage.
[0086] S23. Input the modulation features into a pre-trained multi-path residual network model to obtain the activation vector of the underwater acoustic communication signal to be identified. ;
[0087] If the modulation features include multiple types, the step specifically includes: the feature extraction network in the multi-path residual network model extracts features from different types of modulation features respectively, fuses the feature extraction results of different types of modulation features, and further extracts the fused features through a fully connected layer to obtain the activation vector of the underwater acoustic communication signal.
[0088] S24. Calculate the activation vector of the underwater acoustic communication signal to be identified according to formula (10). Average activation vector for each modulation scheme O'Sullivan distance :
[0089] (10)
[0090] Among them, the average activation vector for each modulation scheme The activation vectors are calculated based on the training samples.
[0091] The distance is compared with the distance threshold of the corresponding modulation method: if the distance from the activation vector of the underwater acoustic communication signal to the average activation vector of each modulation method is greater than the corresponding distance threshold, the modulation method of the underwater acoustic communication signal is judged as unknown; if the distance from the activation vector of the underwater acoustic communication signal to be identified to the average activation vector of a certain modulation method is less than the distance threshold corresponding to that modulation method, the modulation method of the underwater acoustic communication signal is judged as known; for underwater acoustic communication signals judged as known, their modulation method is determined based on the activation vector of the signal.
[0092] For underwater acoustic communication signals classified as known categories, the modulation scheme is determined based on the activation vector of the signal. Specifically, this includes: performing softmax processing on the activation vector to obtain the probability of each modulation scheme corresponding to the activation vector, and selecting the modulation scheme with the highest probability as the modulation scheme of the signal.
[0093] In some embodiments, the average activation vector for each modulation scheme Obtained through the following methods:
[0094] A training sample set is constructed, which includes multiple samples. Each sample uses one or more modulation features of a certain underwater acoustic communication signal as sample features and the modulation mode of the underwater acoustic signal as a label. The types of modulation features are determined by the feature selection stage.
[0095] The multi-path residual network model is trained using the training sample set.
[0096] After training is completed, the activation vector of each training sample is obtained through the multi-path residual network model;
[0097] Based on the activation vector of each training sample and its corresponding modulation scheme label, calculate the average activation vector of the training samples for each modulation scheme. , and take this as the cluster center for each category, as shown in formula (11):
[0098] (11)
[0099] Where j is the modulation scheme code. This represents a sample of underwater acoustic communication signal modulation with modulation mode j. Indicates underwater acoustic communication signals The activation vector is the output of the penultimate layer (i.e., the last fully connected layer) of the multipath residual network model.
[0100] The above describes the identification stage of the underwater acoustic communication modulation identification method provided by this invention.
[0101] The following describes the specific application process and simulation test results of the underwater acoustic communication modulation identification method provided by the present invention in practical underwater acoustic communication simulation, with reference to specific embodiments, to illustrate the effectiveness of the present invention.
[0102] S1, Feature Selection Stage:
[0103] The NOF1 shallow sea channel of WATERMARK is selected, and Gaussian white noise is added after convolution with the modulation signal to form the underwater acoustic communication modulation signal, thus forming an underwater acoustic communication modulation dataset. The underwater acoustic communication modulation recognition process is simulated based on this dataset.
[0104] The simulation parameters are as follows: the modulation scheme types are 2ASK, 4ASK, 2FSK, 4FSK, 8FSK, 2PSK, 4PSK, 8QAM, and OFDM; the signal carrier frequency is randomly selected from 750, 1000, and 1250Hz; the symbol rate range is 32, 40, and 50bps; and the roll-off factor is 0.25. The frequency spacing range for 2FSK is 200, 300, and 400Hz; for 4FSK, it is 100, 200, and 300Hz; for 8FSK, it is 100, 125, and 150Hz; and the OFDM bandwidth range is randomly selected from 400 to 800Hz with 100Hz intervals. The simulation signal-to-noise ratio (SNR) uses the common SNR range for underwater acoustic communication. The training set simulation SNR is -10 to 10dB, and the validation set simulation SNR is -10 to 10dB. Here, the SNR is defined as:
[0105] (12)
[0106] The underwater acoustic communication signals in the aforementioned underwater acoustic communication modulation dataset are preprocessed by performing time-domain waveform normalization, time-spectrum feature extraction, and quadratic and quartic spectrum feature extraction to obtain the time-domain waveform features, time-spectrum features, quadratic spectrum features, and quartic spectrum features of the underwater acoustic communication signals. Based on the above modulation features and modulation mode labels, an underwater acoustic communication modulation training sample set and an underwater acoustic communication modulation verification sample set are constructed. Each sample set includes multiple classes of samples, and each class of samples corresponds to a modulation feature. Each sample uses a specific modulation feature of a segment of underwater acoustic communication signal as its sample feature and the modulation mode of that segment of underwater acoustic signal as its label.
[0107] The multi-path residual network model is trained using the training sample set.
[0108] The feature extraction network in the trained multi-path residual network model is used to extract features from various types of samples in the validation sample set to obtain high-dimensional feature vectors for each type of sample; the t-SNE technique is used to reduce the dimensionality of the high-dimensional feature vectors, and the CH index of each type of sample is calculated based on the t-SNE results. One or more modulation features are selected based on the calculation results.
[0109] The results of the t-SNE dimensionality reduction analysis in this embodiment are as follows: Figure 3As shown, the four images correspond to the dimensionality reduction results of four modulation feature samples: time-domain waveform features, time-frequency spectrum features, quadratic spectrum features, and fourth-order spectrum features, respectively. Points of different shapes in the images represent samples of different modulation schemes. It can be seen that in the t-SNE dimensionality reduction results of both time-domain and time-frequency spectrum features, the dimensionality reduction results for different modulation schemes are dispersed, while the dimensionality reduction results for the same modulation scheme are close together, forming relatively clear clusters. This indicates that time-domain and time-frequency spectrum features are effective in distinguishing modulation schemes. The CH index calculation results are as follows... Figure 3 As shown in the upper left corner of each figure: the CH index of the quadratic spectrum is 528, the CH index of the fourth spectrum is 216, the CH index of the time-domain feature is 6028, and the CH index of the time-frequency plot is 5636. It can be seen that the CH indices of the time-domain feature and the time-frequency plot both exceed 5000, which is significantly higher than the other two features. Therefore, the time-domain feature and the time-frequency plot are used as modulation features. The time-domain feature and the time-frequency plot are used as inputs to the multi-path residual network model to obtain an optimized multi-path residual network model for underwater acoustic communication modulation identification.
[0110] S2, Identification Phase:
[0111] First, prepare for the identification process.
[0112] 2ASK, 4ASK, 2FSK, 4FSK, and 2PSK were selected as known modulation schemes, and OFDM was selected as an unknown category. The modulation parameters were consistent with those in the feature selection stage. The simulated signal-to-noise ratio of the training set was -10 to 10 dB, and the simulated signal-to-noise ratio of the test set was -10 to 10 dB.
[0113] Before identification, a multi-path residual network model is trained using a training set. The training set includes multiple training samples, each using the time-domain characteristics and time-frequency plot of a segment of underwater acoustic communication signal as its feature, and the modulation scheme of that underwater acoustic signal as its label. All training samples are from underwater acoustic signals of the aforementioned known modulation scheme categories.
[0114] After training, the activation vector of each training sample is obtained through the multi-path residual network model; based on the activation vector of each training sample and its corresponding sample label, the average activation vector of the training samples for each modulation scheme is calculated, and this average activation vector is used as the average activation vector for each modulation scheme. .
[0115] The above outlines the preparations before the simulation recognition. Next, the model's open-set recognition performance will be tested using a test set. In this test set, the test samples include unknown-class OFDM signals, which are not included in the training set samples.
[0116] Obtain the underwater acoustic communication signal to be identified from the test set; preprocess the underwater acoustic communication signal to obtain the time-domain characteristics and time-frequency diagram of the underwater acoustic communication signal;
[0117] The time-domain features and time-frequency graph are input into the trained multi-path residual network model to obtain the activation vector of the underwater acoustic communication signal to be identified. ;
[0118] Calculate the activation vector of the underwater acoustic communication signal to be identified. Average activation vector for each modulation scheme O'Sullivan distance ;
[0119] The distance is compared with the distance threshold of the corresponding modulation method: if the distance from the activation vector of the underwater acoustic communication signal to the average activation vector of each modulation method is greater than the corresponding distance threshold, the modulation method of the underwater acoustic communication signal is judged as unknown; if the distance from the activation vector of the underwater acoustic communication signal to be identified to the average activation vector of a certain modulation method is less than the distance threshold corresponding to that modulation method, the modulation method of the underwater acoustic communication signal is judged as known; for underwater acoustic communication signals judged as known, their modulation method is determined according to the activation vector of the signal.
[0120] Finally, based on the test simulation results, the underwater acoustic modulation recognition method provided by this invention is compared with existing SoftMax and OpenMax methods. The F1 score is used as the evaluation index for the classification performance of underwater acoustic modulation recognition. The F1 score is a commonly used evaluation index in classification problems; it is the harmonic mean of precision and recall, with a value range of [0, 1]. It is more valuable than precision alone when the data class distribution is unbalanced. The higher the F1 score, the better the model's classification performance.
[0121] Under the same conditions, within the commonly used signal-to-noise ratio range of -10 to 10 dB in underwater acoustic communication, the F1 score comparison curves of the underwater acoustic modulation identification method, SoftMax method, and OpenMax method provided by this invention are as follows: Figure 4 As shown, the underwater acoustic modulation recognition method provided by this invention has a better classification and recognition effect than the existing SoftMax and OpenMax methods.
[0122] As can be seen from the above, the underwater acoustic communication modulation identification method provided by the present invention has the following beneficial effects compared with the prior art:
[0123] 1. The underwater acoustic communication modulation identification method provided by the present invention uses t-SNE dimensionality reduction analysis to optimize the multivariate modulation identification features and uses modulation features with good class discrimination for modulation identification, which significantly improves the model's generalization ability and identification accuracy.
[0124] 2. The underwater acoustic communication modulation identification method provided by this invention completes open set identification based on a multi-path residual network model optimized by feature filtering, and can identify unknown modulation methods with relatively high accuracy. This gives the invention stronger detection capabilities in typical application scenarios such as underwater acoustic countermeasures and underwater reconnaissance.
[0125] The underwater acoustic communication modulation identification method provided by this invention has both high accuracy of closed set identification and adaptability to open set scenarios in adversarial underwater acoustic environments, providing an effective solution for non-cooperative communication analysis in complex underwater acoustic channels.
[0126] In the description of the embodiments of this application, the words "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.
[0127] In the description of the embodiments of this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, B existing alone, and A and B existing simultaneously. Furthermore, unless otherwise stated, the term "multiple" means two or more.
[0128] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.
[0129] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for filtering modulation identification features in underwater acoustic communication, characterized in that, The method includes: Acquire an underwater acoustic communication modulation dataset, which includes modulated underwater acoustic communication signals and their modulation methods; The modulated underwater acoustic communication signal is preprocessed to obtain various modulation features; A sample set is constructed based on the dataset. The sample set includes multiple types of samples, each type of sample corresponding to a modulation feature. Each sample uses a certain modulation feature of a certain underwater acoustic communication signal as its sample feature and the modulation method of that underwater acoustic signal as its label. The feature extraction network in the pre-trained multi-path residual network model is used to extract features from various types of samples in the sample set to obtain high-dimensional feature vectors for each type of sample. The high-dimensional feature vector is reduced in dimensionality using the t-SNE technique; Calculate the Calinski-Harabasz exponent for the t-SNE dimensionality reduction results of each class of samples, and select one or more modulation features based on the calculation results; One or more modulation features obtained through screening are used as input to a multi-path residual network model to obtain an optimized multi-path residual network model, which is then used for modulation identification in underwater acoustic communication.
2. The underwater acoustic communication modulation identification method according to claim 1, characterized in that, The calculation of the Calinski-Harabasz index of the t-SNE dimensionality reduction result for each class of samples, and the selection of one or more modulation features based on the calculation result, specifically includes: selecting one or more modulation features for the corresponding class of samples whose Calinski-Harabasz index is higher than a certain threshold.
3. The underwater acoustic communication modulation identification method according to claim 1, characterized in that, The preprocessing of the modulated underwater acoustic communication signal to obtain multiple modulation features specifically includes: using short-time Fourier transform to preprocess the modulated underwater acoustic communication signal to obtain multiple modulation features of the underwater acoustic communication signal.
4. The underwater acoustic communication modulation identification method according to claim 1, characterized in that, The types of modulation features include: time-domain waveform features, time-spectrum features, quadratic spectrum features, and quartic spectrum features.
5. A method for identifying modulation in underwater acoustic communication, characterized in that, The method includes: Acquire the underwater acoustic communication signal to be identified; The underwater acoustic communication signal is preprocessed to obtain one or more modulation features of the underwater acoustic communication signal; the type of modulation feature is determined by the feature screening method described in any one of claims 1-4. The modulation features are input into a pre-trained multi-path residual network model to obtain the activation vector of the underwater acoustic communication signal; Calculate the distance between the activation vector of the underwater acoustic communication signal and the average activation vector of each modulation scheme, and compare the distance with the corresponding distance threshold; if the distance between the activation vector of the underwater acoustic communication signal and the average activation vector of each modulation scheme is greater than the corresponding distance threshold, then the modulation scheme of the underwater acoustic communication signal is classified as unknown; the average activation vector of each modulation scheme is calculated based on the activation vector of the training samples; If the distance between the activation vector of the underwater acoustic communication signal to be identified and the average activation vector of a certain modulation scheme is less than the corresponding distance threshold, then the modulation scheme of the underwater acoustic communication signal is classified as a known category; for underwater acoustic communication signals classified as known categories, the modulation scheme is determined based on the activation vector of the signal.
6. The underwater acoustic communication modulation identification method according to claim 5, characterized in that, Before acquiring the underwater acoustic communication signal to be identified, the method further includes: A training sample set is constructed, which includes multiple samples. The samples use one or more modulation features of a certain underwater acoustic communication signal as sample features and the modulation mode of the underwater acoustic signal as a label. The type of modulation feature is determined by the feature selection method described in any one of claims 1-4. The multi-path residual network model is trained using the training sample set. After training is completed, the activation vector of each training sample is obtained through the multi-path residual network model; Based on the activation vector of each training sample and its corresponding sample label, the average activation vector of the training samples for each modulation scheme is calculated and used as the average activation vector for each modulation scheme.
7. The underwater acoustic communication modulation identification method according to claim 5, characterized in that, If the modulation features include multiple types, the modulation features are input into a pre-trained multi-path residual network model to obtain the activation vector of the underwater acoustic communication signal. Specifically, the feature extraction network in the multi-path residual network model extracts features from different types of modulation features, fuses the feature extraction results of different types of modulation features, and further extracts the fused features to obtain the activation vector of the underwater acoustic communication signal.