Underwater acoustic communication signal open set identification method and system based on multi-task learning network

Abstract

The invention discloses an underwater acoustic communication signal open set identification method and system based on a multi-task learning network. The method comprises the following steps: step 1, inputting an underwater acoustic communication signal to be detected into an impulse noise signal preprocessing module; 2, inputting the preprocessed signal into a graph domain mapping multi-task learning intelligent detection and identification module; step 3, carrying out narrowband processing on the underwater acoustic communication signal to be detected; 4, inputting the to-be-detected underwater acoustic communication signal subjected to narrowband processing into a ViT modulation identification auxiliary module; 5, correcting the preliminary prediction probability vector to obtain a prediction probability vector of the underwater acoustic communication signal to be detected; and step 6, correcting the prediction probability vector of the underwater acoustic communication signal to be detected according to a probability decision open set identification module, and completing open set identification of the underwater acoustic communication signal. According to the method, the recognition precision of the system is improved, the robustness and flexibility are enhanced, and the adaptability and open set recognition capability in a dynamically changing environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic communication signal analysis and processing, and in particular to an open set recognition method and system for underwater acoustic communication signals based on a multi-task learning network. Background Art

[0002] The transmission of underwater acoustic communication signals is significantly affected by the marine environment. Factors such as environmental noise, multipath effects, and changes in underwater equipment can all cause signal distortion and changes. Due to the wide frequency range and complex transmission environment of underwater acoustic signals, traditional modulation recognition methods generally rely on fixed feature extraction algorithms, such as time-frequency analysis and wavelet transform. These methods perform well under fixed noise conditions, but their performance degrades significantly in dynamically changing underwater acoustic environments, especially when encountering strong noise interference or complex underwater terrain. This problem of poor environmental adaptability makes it difficult for traditional methods to cope with the diversity and unpredictability of signals in actual underwater acoustic communications.

[0003] In addition, with the continuous development of underwater acoustic communication systems, more diverse and complex modulation methods have emerged. However, existing modulation recognition methods usually assume that all modulation methods are of known categories. In practical applications, underwater acoustic communication systems often encounter unknown modulation methods, and this "open set recognition" problem has not yet been fully solved. Open set recognition requires that the system not only be able to recognize signals of known categories, but also make appropriate processing when faced with unseen signals to prevent misclassification or recognition failure. However, existing technologies lack effective solutions for how to deal with unknown category signals, especially in a changing underwater acoustic environment. Traditional methods usually rely on fixed thresholds or rules for unknown category detection, which makes it difficult to adapt to complex and changing actual environments.

[0004] In recent years, multi-task learning methods based on deep learning have made significant progress in many fields, but their application in underwater acoustic communications still faces some challenges. First, sharing information between tasks does not always lead to performance improvements, especially when there is strong conflict or interference between tasks, the model may be negatively affected. Second, how to reasonably design task combinations and weight distribution to ensure complementarity and synergy between tasks remains an important research issue. In addition, the complexity and diversity of underwater acoustic communication signals require multi-task learning to handle multiple related tasks simultaneously, while existing methods often lack effective mechanisms to balance the impact of different tasks, resulting in the recognition performance under the multi-task learning framework failing to reach the ideal state.

[0005] In summary, traditional underwater acoustic communication signal detection and recognition methods suffer from poor adaptability and insufficient open-set recognition capabilities in dynamically changing environments. Existing multi-task learning methods still need improvement in task collaborative optimization and weight allocation. These issues limit the performance and reliability of underwater acoustic communication systems in practical applications, and new technical solutions are urgently needed. Summary of the Invention

[0006] In order to solve the problems of poor adaptability, insufficient open set recognition capability and unreasonable task collaborative optimization and weight distribution in the existing technology in a dynamically changing environment, the present invention provides an open set recognition method and system for underwater acoustic communication signals based on a multi-task learning network. The present invention pre-processes the underwater acoustic communication signal to be tested through an impulse noise signal pre-processing module to provide a clearer and more reliable signal input for subsequent signal detection and open set recognition tasks. Afterwards, a preliminary prediction of the signal modulation category is made through a graph domain mapping multi-task learning intelligent detection and recognition module, and the underwater acoustic communication signal to be tested is narrow-band processed. The narrow-band processed underwater acoustic communication signal to be tested is input into the ViT modulation recognition auxiliary module to obtain a corrected prediction probability. Finally, the preliminary prediction probability vector is corrected according to the corrected prediction probability, and the corrected prediction probability vector is input into the probability decision open set recognition module to complete the open set recognition of the underwater acoustic communication signal. The present invention places the impulse noise signal preprocessing module, the image domain mapping multi-task learning intelligent detection and recognition module, and the ViT modulation recognition auxiliary module within a unified framework, facilitating the collaborative processing of complex tasks and the rational allocation of weights. The present invention not only improves the recognition accuracy of the system, but also enhances its robustness and flexibility, and improves its adaptability and open-set recognition capabilities in dynamically changing environments.

[0007] In order to achieve the above object, the technical solution of the present invention is:

[0008] The first aspect of the present invention proposes an open set recognition method for underwater acoustic communication signals based on a multi-task learning network, comprising:

[0009] Step 1: Input the underwater acoustic communication signal to be tested into the impulse noise signal preprocessing module to obtain the preprocessed signal, which is convenient for providing a clearer and more reliable signal input for subsequent signal detection and open set recognition tasks;

[0010] Step 2: Input the preprocessed signal into the image domain mapping multi-task learning intelligent detection and recognition module to obtain the preliminary prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be tested, which facilitates the subsequent correction of the prediction probability;

[0011] Step 3: performing narrowband processing on the underwater acoustic communication signal to be measured according to the relative time-frequency position of the underwater acoustic communication signal to be measured, thereby obtaining the narrowband processed underwater acoustic communication signal to be measured, so as to facilitate the subsequent correction prediction probability;

[0012] Step 4: Input the narrowbanded underwater acoustic communication signal to the ViT modulation identification auxiliary module to obtain the corrected prediction probability, which is convenient for correcting the initial prediction probability vector;

[0013] Step 5: Modify the preliminary prediction probability vector according to the modified prediction probability to obtain the prediction probability vector of the underwater acoustic communication signal to be measured;

[0014] Step 6: According to the probability decision open set recognition module, the predicted probability vector of the underwater acoustic communication signal to be tested is corrected to obtain the open set recognition result of the modulation mode of the underwater acoustic communication signal to be tested, thereby completing the open set recognition of the underwater acoustic communication signal.

[0015] Furthermore, the impulse noise signal preprocessing module is expressed by the following formula:

[0016]

[0017] τ r =(1+2τ0)τ Q

[0018] Among them, y′(n) is the preprocessed signal, y(n) is the underwater acoustic communication signal to be measured, τ0 is the calculation parameter, y(n)| is the modulus of y(n), τ Q is the second quartile value of |y(n)|, τ r is the middle value.

[0019] Furthermore, the image domain mapping multi-task learning intelligent detection and recognition module includes an image domain mapping submodule and a multi-task learning network;

[0020] The image domain mapping submodule is used to perform short-time Fourier transform on the pre-processed signal, so as to make the processed features richer, the signal performance more intuitive, and the anti-interference ability stronger;

[0021] The multi-task learning network processes the output of the image domain mapping submodule to obtain a preliminary prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be measured; wherein the multi-task learning network includes a YOLOv5s target detection network;

[0022] The preliminary prediction probability vector is expressed as follows:

[0023] P={P 2FSK , P 4FSK , P 8FSK , P PSK , P OFDM}

[0024] Among them, P is the initial prediction probability vector, P 2FSKis the binary frequency shift keying probability, P 4FSK is the quaternary frequency shift keying probability, P 8FSK is the octal frequency shift keying probability, P PSK is the phase shift keying probability, P OFDM is the OFDM probability.

[0025] Furthermore, the convolution structure in the multi-task learning network is replaced by a wavelet convolution structure;

[0026] The wavelet convolution structure includes a binarization unit, a first convolution unit, a two-dimensional Haar wavelet transform unit, a second convolution unit, an inverse wavelet transform unit and a summation unit;

[0027] The binarization unit is used to perform a binarization operation on the input of the wavelet convolution structure;

[0028] The first convolution unit is used to perform a convolution operation on the output of the binarization unit;

[0029] The two-dimensional Haar wavelet transform unit is used to perform a two-dimensional Haar wavelet transform on the output of the binarization unit;

[0030] The second convolution unit is used to perform a convolution operation on the output of the two-dimensional Haar wavelet transform unit;

[0031] The inverse wavelet transform unit is used to perform an inverse wavelet transform operation on the output of the second convolution unit;

[0032] The summing unit is used to add the output of the inverse wavelet transform unit and the output of the first convolution unit.

[0033] Furthermore, the narrowband processing of the underwater acoustic communication signal to be measured is performed according to the relative time-frequency position of the underwater acoustic communication signal to be measured, which is expressed by the following formula:

[0034]

[0035] Among them, s narrow (n) is the underwater acoustic communication signal to be tested after narrowband processing, w is the relative duration of the signal, x center is the relative position of the midpoint in time, y center is the relative position of the frequency midpoint, h is the relative bandwidth of the signal, L={x center ,y center , w, h} is the relative time-frequency position of the underwater acoustic communication signal to be measured, R is the step size between time frames, m is the time dimension of the time-spectrum graph, k is the frequency dimension of the time-spectrum graph, T is the signal duration, F s is the sampling rate, and n is the time dimension of the underwater acoustic communication signal to be measured after narrowband processing.

[0036] Furthermore, the ViT modulation identification auxiliary module includes a square spectrum estimation unit and a ViT network;

[0037] The square spectrum estimation unit is used to perform square spectrum estimation on the narrowband processed underwater acoustic communication signal to obtain a corrected prediction probability;

[0038] The ViT network is used to process the output of the square spectrum estimation unit to obtain a modified prediction probability; wherein the modified probability includes a binary phase modulation probability and an orthogonal phase shift keying probability.

[0039] Furthermore, the step five specifically includes:

[0040] According to the modified prediction probability, the modified binary phase modulation probability and the quadrature phase shift keying probability are calculated;

[0041] The phase shift keying probability in the preliminary prediction probability vector is replaced by the modified binary phase modulation probability and orthogonal phase shift keying probability to obtain the prediction probability vector of the underwater acoustic communication signal to be tested, so as to improve the prediction accuracy.

[0042] Furthermore, the modified binary phase modulation probability and quadrature phase shift keying probability calculated according to the modified predicted probability are expressed as follows:

[0043] P BPSK =λP PSK +(1-λ)P B

[0044] P QPSK =λP PSK +(1-λ)P Q

[0045] Among them, P BPSK is the modified binary phase modulation probability, λ is the weight parameter, P PSK is the phase shift keying probability, P B is the binary phase modulation probability, P QPSK is the modified QPSK probability, P Q is the quadrature phase shift keying probability.

[0046] Furthermore, the probabilistic decision open set identification module includes a distance module, a correction vector module and a prediction module;

[0047] The distance module is used to calculate the Euclidean distance between the predicted probability vector of the underwater acoustic communication signal to be measured and the central prediction vectors of multiple modulation types, which is specifically expressed by the following formula:

[0048] D i =P-CV i

[0049] Among them, D i is the Euclidean distance between the predicted probability vector of the underwater acoustic communication signal to be tested and the central prediction vector of the i-th modulation type, P is the predicted probability vector of the underwater acoustic communication signal to be tested, CV i is the center prediction vector of the i-th class;

[0050] The correction vector module is used to input the output of the distance module into the cumulative distribution functions of multiple modulation types respectively to obtain corresponding probability correction vectors, and calculate the known modulation type probability prediction vector and the unknown modulation type probability of the underwater acoustic communication signal to be tested according to the probability correction vector, which is specifically expressed by the following formula;

[0051] V known =P*W

[0052] V unkown =P·(1-W) T

[0053] Among them, V known is the probability prediction vector of the known modulation type, W is the probability correction vector, V unkown is the probability of unknown modulation type, * is Hadmard multiplication;

[0054] The prediction module is used to normalize the output of the correction vector module and calculate the category probability, and the maximum category probability is used as the modulation category of the underwater acoustic communication signal to be tested, which is specifically expressed by the following formula:

[0055]

[0056] Where P′ is the category probability, softmax is the softmax function, and V is the known modulation type probability prediction vector or the unknown modulation type probability.

[0057] The second aspect of the present invention proposes an open set recognition system for underwater acoustic communication signals based on a multi-task learning network, comprising:

[0058] The signal preprocessing module is used to input the underwater acoustic communication signal to be tested into the impulse noise signal preprocessing module to obtain the preprocessed signal, so as to provide a clearer and more reliable signal input for subsequent signal detection and open set recognition tasks;

[0059] The initial prediction module is used to input the preprocessed signal into the image domain mapping multi-task learning intelligent detection and recognition module to obtain the initial prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be tested, which is convenient for the subsequent correction of the prediction probability;

[0060] A narrowbanding module is used to narrowband the underwater acoustic communication signal to be tested according to the relative time-frequency position of the underwater acoustic communication signal to be tested, thereby obtaining the narrowbanded underwater acoustic communication signal to be tested, so as to facilitate the subsequent correction prediction probability;

[0061] The modified probability module is used to input the narrowband processed underwater acoustic communication signal to the ViT modulation identification auxiliary module to obtain the modified prediction probability, so as to facilitate the correction of the preliminary prediction probability vector;

[0062] A correction module is used to correct the preliminary prediction probability vector according to the corrected prediction probability to obtain a prediction probability vector of the underwater acoustic communication signal to be measured;

[0063] The recognition module is used to correct the predicted probability vector of the underwater acoustic communication signal to be tested according to the probability decision open set recognition module, obtain the open set recognition result of the modulation mode of the underwater acoustic communication signal to be tested, and complete the open set recognition of the underwater acoustic communication signal.

[0064] Beneficial effects of the present invention:

[0065] (1) This invention aims to solve the problem that existing methods in the detection and recognition of broadband underwater acoustic communication signals usually rely on a single network model, which makes it difficult to fully capture the multimodal characteristics of the signal, resulting in limited system performance. A hybrid network framework (including an impulse noise signal preprocessing module, a graph domain mapping multi-task learning intelligent detection and recognition module, a ViT modulation recognition auxiliary module and a probabilistic decision open set recognition module) is proposed. This framework combines the advantages of convolutional neural networks and Vision-Transformer, can simultaneously capture the local time-frequency characteristics and global distribution dependencies of the signal, fully explore the complementarity between different features, and improve the detection and recognition performance of broadband underwater acoustic communication signals in complex underwater acoustic environments. Through the coordinated optimization of the overall framework, the robustness and adaptability of signal processing are significantly improved.

[0066] (2) This invention addresses the problem that existing methods typically treat signal detection and open-set recognition as independent tasks, resulting in insufficient information sharing between tasks and reduced overall system efficiency and accuracy. This invention proposes a graph-domain mapping multi-task learning intelligent detection and recognition module. By mapping the high-dimensional features of underwater acoustic communication signals to a graph structure domain, signal detection and modulation recognition are combined for multi-task learning. This combination of graph-domain feature representation and multi-task learning effectively improves the accuracy of signal detection and modulation recognition.

[0067] (3) To address the problem that the local receptive field of CNNs cannot fully capture the global information of the spectral characteristics of broadband underwater acoustic communication signals, this paper proposes a ViT modulation recognition auxiliary module. By extracting square spectrum features from narrowband underwater acoustic communication signals, ViT can effectively extract global features and cross-time / frequency dependencies in the signal, achieving higher accuracy and robustness in modulation recognition tasks than traditional CNNs.

[0068] (4) In order to solve the problem that traditional underwater acoustic communication signal recognition methods can only recognize signals of known categories and cannot process signals of new or unseen categories, resulting in limited flexibility and adaptability of the system, the present invention proposes a probabilistic decision open set recognition module. This module corrects the predicted probability vector and converts it into a standard probability distribution in combination with the Softmax function, thereby effectively realizing the recognition of unknown category signals and significantly improving the adaptability and classification accuracy of the system in a dynamically changing environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A flowchart of an open-set recognition method for underwater acoustic communication signals based on a multi-task learning network provided in an embodiment of the present invention.

[0070] Figure 2 A schematic diagram of narrowband processing provided by an embodiment of the present invention.

[0071] Figure 3 A schematic diagram of a ViT modulation identification auxiliary module provided by an embodiment of the present invention.

[0072] Figure 4 A schematic diagram of the training and testing process of an open-set recognition method for underwater acoustic communication signals based on a multi-task learning network provided in an embodiment of the present invention.

[0073] Figure 5 A schematic diagram of the network structure of an example model provided in an embodiment of the present invention.

[0074] Figure 6 A schematic diagram of a wavelet convolution structure provided by an embodiment of the present invention.

[0075] Figure 7 A schematic diagram of a probabilistic decision open set recognition module provided by an embodiment of the present invention.

[0076] Figure 8 A schematic diagram of amplitude-frequency response curves of different channels provided by an embodiment of the present invention.

[0077] Figure 9 A schematic diagram of the open set recognition results of underwater acoustic communication signals provided by an embodiment of the present invention.

[0078] Figure 10This is an architectural diagram of an open-set recognition system for underwater acoustic communication signals based on a multi-task learning network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0079] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0080] Example 1

[0081] like Figure 1 As shown, the open set recognition method of underwater acoustic communication signals based on a multi-task learning network includes:

[0082] S101: Inputting the underwater acoustic communication signal to be measured into the impulse noise signal preprocessing module to obtain a preprocessed signal.

[0083] Specifically, based on the existing commonly used underwater acoustic communication signal modulation methods, including binary frequency shift keying (2FSK), quaternary frequency shift keying (4FSK), octal frequency shift keying (8FSK), binary phase modulation (BPSK), quadrature phase shift keying (QPSK) and orthogonal frequency division multiplexing (OFDM), the simulation obtains the transmitted signal s(n), and then obtains the underwater acoustic communication signal to be tested after passing through the underwater acoustic multipath channel h(n) and additive ocean ambient noise v(n). The underwater acoustic communication signal to be tested during training can be modeled as:

[0084]

[0085] in, Denotes the convolution operation. The channel h(n) is simulated based on the Argo ocean database using Bellhop channel numerical simulation software.

[0086] The ocean ambient noise w(n) is modeled as Alpha stable distribution noise, and its characteristic function is:

[0087]

[0088] Here, α (0 < α ≤ 2) is a characteristic exponent that defines the strength of the distribution's impulses. When α = 2, it degenerates into a Gaussian distribution. The location parameter a determines the center of the distribution, and the dispersion coefficient γ defines the degree to which the distribution deviates from its mean. Since the Alpha stable distribution does not have second-order or higher-order statistics when α < 2, the mixed signal-to-noise ratio (MSNR) is used to measure the power relationship between the signal and the noise:

[0089]

[0090] in, is the useful signal variance.

[0091] The impulse noise signal preprocessing module is expressed as follows:

[0092]

[0093] τ r =(1+2τ0)τ Q

[0094] Among them, y′(n) is the preprocessed signal, y(n) is the underwater acoustic communication signal to be measured, τ0 is the calculation parameter, y(n)| is the modulus of y(n), τ Q is the second quartile value of |y(n)|, τ r is the middle value.

[0095] S102: Input the preprocessed signal into the image domain mapping multi-task learning intelligent detection and recognition module to obtain a preliminary prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be measured.

[0096] Specifically, the image domain mapping multi-task learning intelligent detection and recognition module includes an image domain mapping sub-module and a multi-task learning network.

[0097] The image domain mapping submodule is used to perform short-time Fourier transform on the preprocessed signal.

[0098] The multi-task learning network processes the output of the image domain mapping submodule to obtain a preliminary prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be measured; among them, the multi-task learning network includes the YOLOv5s target detection network.

[0099] The initial prediction probability vector is expressed as follows: P = {P 2FSK , P 4FSK , P 8FSK , P PSK , P OFDM}

[0100] Among them, P is the initial prediction probability vector, P 2FSK is the binary frequency shift keying probability, P 4FSK is the quaternary frequency shift keying probability, P 8FSK is the octal frequency shift keying probability, P PSK is the phase shift keying probability, P OFDM is the OFDM probability.

[0101] The relative time-frequency position of the underwater acoustic communication signal to be measured is expressed by the following formula: L = {x center,y center ,w,h}

[0102] Among them, w is the relative duration of the signal, x center is the relative position of the midpoint in time, y center is the relative position of the frequency midpoint, h is the relative bandwidth of the signal, L={x center ,y center , w, h} are the relative time-frequency positions of the underwater acoustic communication signal to be measured.

[0103] S103: performing narrowband processing on the underwater acoustic communication signal to be measured according to the relative time-frequency position of the underwater acoustic communication signal to be measured, to obtain the underwater acoustic communication signal to be measured after narrowband processing.

[0104] Specifically, if Figure 2 As shown, the narrowband processing is expressed by the following formula:

[0105]

[0106] Among them, s narrow (n) is the underwater acoustic communication signal to be tested after narrowband processing, R is the step size between time frames, m is the time dimension of the time-spectrum graph, k is the frequency dimension of the time-spectrum graph, T is the signal duration, F s is the sampling rate, and n is the time dimension of the underwater acoustic communication signal to be measured after narrowband processing.

[0107] S104: Inputting the narrowband processed underwater acoustic communication signal to be tested into the ViT modulation identification auxiliary module to obtain a corrected prediction probability.

[0108] Specifically, due to the similarity in time-frequency distribution between BPSK and QPSK, the image-domain mapping multi-task learning intelligent detection and recognition module cannot recognize them. Considering that the square spectrum of the BPSK signal has a clear impulse at twice its carrier frequency, while the QPSK signal does not have this feature, and that the convolutional neural network has limitations in recognizing large blank images due to its limited receptive field, the present invention designs a ViT (Vision-Transformer) modulation recognition auxiliary module to distinguish BPSK and QPSK signals and assist in modulation recognition.

[0109] The ViT modulation identification auxiliary module includes a square spectrum estimation unit and a ViT network. The square spectrum estimation unit is used to perform square spectrum estimation on the underwater acoustic communication signal to be tested after narrowband processing. The ViT network is used to process the output of the square spectrum estimation unit to obtain a modified prediction probability. Among them, the modified probability includes the binary phase modulation probability and the orthogonal phase shift keying probability. The module structure is as follows Figure 3 shown.

[0110] S105: Correcting the preliminary prediction probability vector according to the corrected prediction probability to obtain a prediction probability vector of the underwater acoustic communication signal to be measured.

[0111] Specifically, when the signal to be tested is judged as phase shift keying (PSK) by the image domain mapping multi-task learning intelligent detection and identification module, the corrected binary phase modulation probability and orthogonal phase shift keying probability are calculated based on the corrected prediction probability. The phase shift keying probability in the preliminary prediction probability vector is replaced with the corrected binary phase modulation probability and orthogonal phase shift keying probability to obtain the predicted probability vector of the underwater acoustic communication signal to be tested. The above process is expressed as follows:

[0112] P BPSK =λP PSK +(1-λ)P B

[0113] P QPSK =λP PSK +(1-λ)P Q

[0114] P={P 2FSK , P 4FSK , P 8FSK , P PSK , P OFDM}→P′={P 2FSK , P 4FSK , P 8FSK , P BPSK , P QPSK , P OFDM}

[0115] Among them, P BPSK is the modified binary phase modulation probability, λ is the weight parameter, P PSK is the phase shift keying probability, P B is the binary phase modulation probability, P QPSK is the modified QPSK probability, P Q is the quadrature phase shift keying probability, and P′ is the predicted probability vector of the underwater acoustic communication signal to be measured.

[0116] S106: Correct the predicted probability vector of the underwater acoustic communication signal to be measured according to the probability decision open set recognition module to obtain the open set recognition result of the modulation mode of the underwater acoustic communication signal to be measured, and complete the open set recognition of the underwater acoustic communication signal.

[0117] Preferably, the present invention has two stages: training and testing. Figure 4As shown. During the training phase, the high-amplitude impulse noise in the received broadband underwater acoustic communication signal is first effectively suppressed through the impulse noise preprocessing module. Secondly, based on the idea of high-dimensional data mapping feature representation method, the preprocessed signal is subjected to the image domain mapping multi-task learning intelligent detection and recognition module to achieve broadband underwater acoustic communication signal detection and preliminary modulation recognition under the background of low signal-to-noise ratio and high dynamic changes, and complete the narrowbandization of the underwater acoustic communication signal. After that, the square spectrum feature of the narrowbanded signal is extracted, and the ViT modulation recognition auxiliary module (Vision-Transformer modulation recognition auxiliary module) constructed using the extracted square spectrum feature map is trained to obtain the final modulation recognition prediction probability vector. Finally, the probability decision open set recognition module is trained using the correctly predicted probability vector.

[0118] During the testing phase, the predicted probability vectors of the underwater acoustic communication signal to be tested can be used to obtain signal detection results and closed-set recognition results after being output by the trained multi-task learning network and ViT network. After being corrected by the trained probability decision open-set recognition module, it is converted into open-set classification prediction probability to realize open-set recognition of broadband underwater acoustic communication signals.

[0119] The present invention discloses a method for open-set recognition of broadband underwater acoustic communication signals based on a hybrid multi-task learning network. The method first introduces an impulse noise preprocessing module to preprocess the underwater acoustic communication signal to be tested. Secondly, a graph domain mapping multi-task learning intelligent detection and recognition module is constructed through a target detection network model to realize broadband underwater acoustic communication signal detection and preliminary modulation recognition under a low signal-to-noise ratio and high dynamic change background, and obtain a preliminary prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be tested. After that, the underwater acoustic communication signal is narrowbanded, and the square spectrum feature of the narrowbanded signal is extracted. The extracted square spectrum feature map is identified by the constructed ViT modulation recognition auxiliary module to obtain a corrected prediction probability, and the preliminary prediction probability vector is corrected. Finally, the prediction probability vector of the underwater acoustic communication signal to be tested is corrected by the probability decision open-set recognition module and converted into a classification prediction probability, thereby realizing open-set recognition of broadband underwater acoustic communication signals. The present invention adopts a hybrid multi-task learning network architecture to achieve collaborative processing of complex tasks from pulse noise suppression, signal detection, modulation recognition to open set recognition within a unified framework, which not only improves the recognition accuracy of the system, but also enhances its robustness and flexibility. Especially in complex underwater acoustic environments and open set recognition problems, it shows greater advantages and innovations than traditional methods, and has important application value and broad prospects.

[0120] Example 2

[0121] Based on the above embodiments, the present invention proposes a structure of an image domain mapping multi-task learning intelligent detection and recognition module, which specifically includes:

[0122] Based on the idea of high-dimensional mapping feature characterization method of data, the present invention takes advantage of the graph domain to represent signal characteristics and spatial information, adopts a signal characterization method based on graph domain mapping, maps low-dimensional time series data to high-dimensional graph domain space for characterization, makes the time series signal produce high-dimensional correlation and improves the data feature dimension, and combines with the artificial intelligence multi-task learning framework to complete the underwater acoustic signal detection and modulation recognition under the background of high and low signal-to-noise ratio and high dynamic change.

[0123] The features that can be reflected by different domain mapping methods are often quite different. The time-frequency distribution diagram depicts the energy distribution of the signal at different times and frequencies. It has the advantages of both time domain and frequency domain representation, contains richer features, and compared with time domain data, the signal performance is more intuitive and has stronger anti-interference ability.

[0124] Therefore, the present invention pre-processes the time-frequency graph of the underwater acoustic communication signal to be tested, and uses it as the object for subsequent identification processing. The expression of the time-frequency graph obtained by the short-time Fourier transform (STFT) of the underwater acoustic communication signal to be tested is as follows:

[0125]

[0126] Wherein, w(n) is the selected window function, which is set as Hamming window, N is the number of sampling points of signal y′(n), and ω=2πmk / N represents the discrete frequency.

[0127] From the signal time-frequency graph, the differences between different signals are reflected in the energy distribution structure in the two-dimensional time-frequency plane. Manual methods are difficult to fully and effectively characterize such differences. Therefore, the present invention constructs a multi-task learning network in the image domain to jointly learn signal position detection and debugging recognition tasks. It automatically extracts the deep features contained in the signal time-frequency image, locates the relative position of the signal time-frequency, and completes classification and recognition simultaneously. It is less affected by multipath fading and noise, and is more robust in complex ocean environments.

[0128] Specifically, the present invention adopts an image target detection and recognition network to construct a graph domain multi-task learning network, and the specific network type can be flexibly selected. Taking into account that the YOLO series target detection network has the multi-task learning ability of target detection and classification, the present invention adopts the YOLOv5s target detection network as the example basic model of the multi-task learning intelligent detection and recognition module. It should be noted that: networks with multi-task learning capabilities of image target detection and classification can constitute the example basic model of the graph domain multi-task learning network referred to in the present invention. The example model is mainly composed of four major parts, namely the input end (Input), the backbone network (Backbone), the neck (Neck) network and the output end (Output). The input end preprocesses the image data, such as scaling, cropping, normalization, etc., so that it meets the input requirements of the model. The backbone network extracts feature maps of different scales from the image, performs preprocessing operations on the input image, and extracts features in the image through convolutional layers and pooling layers. The neck network upsamples and downsamples feature maps of different levels to achieve cross-layer feature fusion. The output end performs multi-scale and multi-category predictions on each pixel on each feature map, and combines the feature maps of different scales in the image through convolutional layers and upsampling to classify and regress the target. The network structure of the example model is as follows: Figure 5 shown.

[0129] Considering that the low symbol rate 2FSK and BPSK signals occupy a very small image domain space after image domain mapping, resulting in a decrease in detection recognition rate, the present invention improves the original example model to enhance its small target detection capability. Although the convolution structure in the example model can extract high-precision image features, it is still limited by the size of the convolution kernel receptive field. Therefore, the present invention introduces a wavelet convolution structure to replace the original convolution structure. The wavelet convolution structure of the single-layer decomposition is as follows: Figure 6 As shown in the figure, this structure is designed to expand the receptive field of convolution and effectively capture the low-frequency information in the input image domain features, which has a good effect on multi-scale problems and small target problems.

[0130] The wavelet convolution structure includes a binarization unit, a first convolution unit, a two-dimensional Haar wavelet transform unit, a second convolution unit, an inverse wavelet transform unit, and a summation unit. The binarization unit is used to perform a binarization operation on the input of the wavelet convolution structure. The first convolution unit is used to perform a convolution operation on the output of the binarization unit. The two-dimensional Haar wavelet transform unit is used to perform a two-dimensional Haar wavelet transform on the output of the binarization unit. The second convolution unit is used to perform a convolution operation on the output of the two-dimensional Haar wavelet transform unit. The inverse wavelet transform unit is used to perform an inverse wavelet transform operation on the output of the second convolution unit. The summation unit is used to add the output of the inverse wavelet transform unit and the output of the first convolution unit.

[0131] Specifically, first, a binarization operation is performed on the input of the wavelet convolution structure (the preprocessed signal) to obtain a matrix input image X of size M×N.

[0132] The two-dimensional Haar wavelet transform is used to decompose the matrix input image X of size M×N:

[0133]

[0134] Among them, x MN The data in the Mth row and Nth column of the matrix input image.

[0135] For each row of the image matrix X i =(x i1 ,x i2 ,…,x iN ) performs Haar wavelet transform to obtain the low-frequency part y containing smooth and low-frequency information avg,i,j and the high-frequency part y containing edge and detail information diff,i,j :

[0136]

[0137] After the transformation, each row will be divided into two parts: one is the low frequency part and the other is the high frequency part. Then, a one-dimensional Haar transform is performed on each column after the transformation. Assume that after processing each row, the image obtained is Y, with a size of M×N. For each column Y of the image j =(y 1j ,y 2j ,…,y Mj ) performs one-dimensional Haar transform:

[0138]

[0139] Among them, z avg,j,i is the low-frequency part after transformation, z diff,j,i is the high frequency part after transformation.

[0140] Ultimately, through transformations in both horizontal and vertical dimensions, the image is decomposed into four subbands: Low-frequency components (LL): Capture low-frequency information of the image, such as overall shape or outline. Horizontal high-frequency components (LH): Capture horizontal edge information in the image. Vertical high-frequency components (HL): Capture vertical edge information in the image. Diagonal high-frequency components (HH): Capture diagonal details in the image.

[0141]

[0142] A 3×3 depthwise convolution kernel is then used on each frequency subband to perform convolution operations on each decomposed subband. Because the wavelet transform reduces the spatial resolution of each subband, a smaller convolution kernel can cover a larger area of the original image, increasing the receptive field. The low-frequency subband (LL) primarily contains large-scale information of the image, so applying convolution on it helps capture global features. The high-frequency subbands (LH, HL, HH) contain local edges and details, which can be captured by convolution operations.

[0143]

[0144] Among them, the input image I∈R M / 2×N / 2 , including four sub-bands LL, LH, HL and HH, convolution kernel K∈R 3×3 , output feature map O∈R (M / 2-2)×(N / 2-2) , including the four sub-bands LL′, LH′, HL′ and HH′ of the convolution output.

[0145] After convolution, the inverse wavelet transform (IWT) is used to re-synthesize the convolution results of each sub-band into a complete output, fusing the features of different frequency levels of the image together to obtain the wavelet reconstructed image:

[0146] First, reconstruct each column in the vertical direction and transform the low-frequency part z avg,j,i With the high frequency part z diff,j,i The specific process is expressed as follows:

[0147] y avg,2i-1,j =z avg,j,i +z diff,j,i

[0148] y avg,2i,j =z avg,j,i -z diff,j,i

[0149] Then, the detail information of each row is reconstructed. The low-frequency part y containing smooth, low-frequency information is converted into avg,i,j and the high-frequency part y containing edge and detail information diff,i,j Merge to get the wavelet reconstructed image

[0150]

[0151] in, Reconstruct the image using wavelet.

[0152] Finally, the reconstructed image is convolved with the original image to obtain the final output image:

[0153]

[0154] Among them, K is a 3×3 depth convolution kernel.

[0155] Example 3

[0156] Based on the above embodiments, the present invention proposes a structure of a probabilistic decision open set recognition module, which specifically includes:

[0157] Specifically, the development of underwater acoustic communication equipment is often accompanied by the continuous emergence of various new signal modulation methods. The signal patterns in the underwater environment are more diverse and complex, which brings certain troubles to the supervision of underwater spectrum resources. To this end, the present invention constructs a probabilistic decision open set recognition module to realize the judgment of the type of unknown underwater acoustic communication signal. The probabilistic decision open set recognition module is as follows: Figure 7 shown.

[0158] In the training phase, according to the output probability prediction vector P, the matrix CV consisting of the prediction vectors of the known class centers is calculated as [CV0,...,CV1,...,CV N-1 ], calculated according to the following formula:

[0159]

[0160] Among them, N i represents the number of samples in the i-th category, V i,j Represents the prediction vector P of the j-th sample of the i-th class.

[0161] After that, calculate the distance vector d of the i-th class sample i :

[0162] d i =V i,j -CV i j=1,2,...,N i

[0163] Using d i Fitting the cumulative distribution function CDF of the Weibull distribution i :

[0164]

[0165] Among them, λ i is the scale parameter to be fitted, k i is the shape parameter to be fitted.

[0166] In the testing phase, the Euclidean distance between the probability prediction vector P of the signal to be tested output by the network and the various central prediction vectors is calculated:

[0167] D i =P-CV i

[0168] Among them, D i is the Euclidean distance between the predicted probability vector of the underwater acoustic communication signal to be tested and the central prediction vector of the i-th modulation type, P is the predicted probability vector of the underwater acoustic communication signal to be tested, CV i is the center prediction vector of the i-th class.

[0169] D i Substitute various cumulative distribution functions CDF respectively i (Cumulative distribution function corresponding to various modulation modes), and obtain the N-dimensional probability correction vector W=[w0,...,w i ,...,w N-1 ], and use this vector to calculate the probability that the signal to be tested belongs to the known class and the unknown class respectively, which can be expressed as follows:

[0170] V known =P*W

[0171] V unkown =P·(1-W) T

[0172] Among them, V known is the probability prediction vector of the known modulation type, W is the probability correction vector, V unkown is the probability of unknown modulation type, and * is Hadmard multiplication.

[0173] And normalize the prediction vector V=[V known ,V unkown ]Get the category probability P={P 2FSK , P 4FSK , P 8FSK , P BPSK , P QPSK , P OFDM , P unknown}, the index of the maximum value represents the category label of the final recognition result.

[0174]

[0175] Where P′ is the category probability, softmax is the softmax function, and V is the known modulation type probability prediction vector or the unknown modulation type probability.

[0176] Example 4

[0177] Based on the above embodiments, the present invention proposes an evaluation experiment of an open set recognition method for underwater acoustic communication signals based on a multi-task learning network, which specifically includes:

[0178] In the experiment, the sampling frequency of the modulation signal is set to 100 kHz, the carrier frequency range is [1 kHz, 30 kHz] Hz, the modulation order of the MPSK signal is 2 or 4, and root-raised cosine pulse shaping is used. The OFDM signal subcarriers are randomly modulated using BPSK or QPSK. The other modulation parameters are shown in Table 1. In Table 1, " / " indicates that the parameter is not involved, "[]" indicates that the data is randomly selected within the closed set range, and "{}" indicates that the data is randomly selected from the listed items.

[0179] Table 1 Signal modulation parameters

[0180]

[0181] To simulate the actual underwater acoustic channel environment, Bellhop channel simulation software was used to simulate four underwater acoustic channels under different transmission conditions based on the Argo ocean database. The channel parameters are shown in Table 2:

[0182] Table 2 Simulation channel parameters

[0183]

[0184]

[0185] Their channel system functions are:

[0186] H1(z)=0.07+0.2z -140 +z -653 +0.23z -817 +0.05z -876

[0187] H2(z)=1+0.58z -369 +0.56z -3581

[0188] H3(z)=0.49+0.18z -169 +z -311

[0189] H4(z)=0.7+0.6134z -541 +0.971z -945 +z -1261

[0190] The signal sampling rate is 100kHz, so the maximum propagation delays of these channels are 8.76ms, 35.81ms, 3.11ms, and 12.61ms respectively. Their amplitude-frequency response curves are shown in Figure 2. Figure 8 shown.

[0191] The model was tested on a dataset consisting of 1000 samples of 6 types of known underwater acoustic communication signals and 1 type of unknown underwater acoustic communication signal. The open set recognition results of underwater acoustic communication signals were obtained as follows: Figure 9 As shown. It can be seen that under the conditions of complex underwater acoustic multipath channels and ocean impulse noise, the recognition accuracy of the method of the present invention for various commonly used underwater acoustic communication signals continues to improve with the signal-to-noise ratio. When the mixed signal-to-noise ratio MSNR>5dB, the recognition rate of 6 types of known underwater acoustic communication signals reaches more than 90%, and the recognition rate of unknown underwater acoustic communication signals reaches more than 80%. This verifies the good performance of the method of the present invention.

[0192] Example 5

[0193] Based on the above embodiments, Figure 10 As shown, the present invention proposes an open set recognition system for underwater acoustic communication signals based on a multi-task learning network, comprising:

[0194] The signal preprocessing module is used to input the underwater acoustic communication signal to be measured into the impulse noise signal preprocessing module to obtain the preprocessed signal.

[0195] The initial prediction module is used to input the preprocessed signal into the image domain mapping multi-task learning intelligent detection and recognition module to obtain the initial prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be tested.

[0196] The narrowband module is used to perform narrowband processing on the underwater acoustic communication signal to be tested according to the relative time-frequency position of the underwater acoustic communication signal to be tested, so as to obtain the underwater acoustic communication signal to be tested after narrowband processing.

[0197] The modified probability module is used to input the narrowband processed underwater acoustic communication signal to be tested into the ViT modulation identification auxiliary module to obtain the modified prediction probability.

[0198] The correction module is used to correct the preliminary prediction probability vector according to the corrected prediction probability to obtain the prediction probability vector of the underwater acoustic communication signal to be measured.

[0199] The recognition module is used to correct the predicted probability vector of the underwater acoustic communication signal to be tested according to the probability decision open set recognition module, obtain the open set recognition result of the modulation mode of the underwater acoustic communication signal to be tested, and complete the open set recognition of the underwater acoustic communication signal.

[0200] It should be noted that the open-set recognition system for underwater acoustic communication signals based on a multi-task learning network provided in an embodiment of the present invention is intended to implement the above-mentioned open-set recognition method for underwater acoustic communication signals based on a multi-task learning network. Its specific functions can be referred to the above-mentioned method embodiments and will not be repeated here.

[0201] In summary, the present invention aims at the problem that in the detection and recognition tasks of broadband underwater acoustic communication signals, existing methods usually rely on a single network model, which makes it difficult to fully capture the multimodal characteristics of the signal, resulting in limited system performance. A hybrid network framework (including an impulse noise signal preprocessing module, a graph domain mapping multi-task learning intelligent detection and recognition module, a ViT modulation recognition auxiliary module and a probabilistic decision open set recognition module) is proposed. The framework combines the advantages of convolutional neural networks and Vision-Transformer, can simultaneously capture the local time-frequency characteristics and global distribution dependencies of the signal, fully explore the complementarity between different features, and improve the detection and recognition performance of broadband underwater acoustic communication signals in complex underwater acoustic environments. Through the collaborative optimization of the overall framework, the robustness and adaptability of signal processing are significantly improved. The present invention aims at the problem that existing methods usually process signal detection and open set recognition as independent tasks, resulting in the inability to fully share information between tasks, reducing the overall efficiency and accuracy of the system. The present invention proposes a graph domain mapping multi-task learning intelligent detection and recognition module. By mapping the high-dimensional features of underwater acoustic communication signals to a graph domain, multi-task learning is performed on signal detection and modulation recognition. The combination of this graph domain feature representation and multi-task learning effectively improves the accuracy of signal detection and modulation recognition. To address the problem that the local receptive field of CNN is unable to fully capture the global information of the spectral features of broadband underwater acoustic communication signals, the present invention proposes a ViT modulation recognition auxiliary module. By extracting square spectrum features from narrowband underwater acoustic communication signals, ViT can effectively extract global features and cross-time / frequency dependencies in the signal, achieving higher accuracy and robustness in modulation recognition tasks than traditional CNN. To address the problem that traditional underwater acoustic communication signal recognition methods can usually only recognize signals of known categories and cannot handle signals of new or unseen categories, resulting in limited flexibility and adaptability of the system, the present invention proposes a probabilistic decision open set recognition module. This module corrects the predicted probability vector and converts it into a standard probability distribution in combination with the Softmax function, thereby effectively realizing the recognition of unknown category signals, significantly improving the system's adaptability and classification accuracy in dynamically changing environments.

[0202] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An open set recognition method for underwater acoustic communication signals based on a multi-task learning network, characterized by: include: Step 1: Input the underwater acoustic communication signal to be tested into the impulse noise signal preprocessing module to obtain a preprocessed signal; Step 2: Input the preprocessed signal into the image domain mapping multi-task learning intelligent detection and recognition module to obtain the preliminary prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be tested; Step 3: performing narrowband processing on the underwater acoustic communication signal to be measured according to the relative time-frequency position of the underwater acoustic communication signal to be measured, to obtain the underwater acoustic communication signal to be measured after narrowband processing; Step 4: Input the narrowband processed underwater acoustic communication signal to the ViT modulation identification auxiliary module to obtain the corrected prediction probability; Step 5: Modify the preliminary prediction probability vector according to the modified prediction probability to obtain the prediction probability vector of the underwater acoustic communication signal to be measured; Step 6: According to the probability decision open set recognition module, the predicted probability vector of the underwater acoustic communication signal to be tested is corrected to obtain the open set recognition result of the modulation mode of the underwater acoustic communication signal to be tested, thereby completing the open set recognition of the underwater acoustic communication signal.

2. The method for open set recognition of underwater acoustic communication signals based on a multi-task learning network according to claim 1 is characterized in that: The impulse noise signal preprocessing module is expressed as follows: t r =(1+2τ0)τ Q Among them, y′(n) is the preprocessed signal, y(n) is the underwater acoustic communication signal to be measured, τ0 is the calculation parameter, y(n)| is the modulus of y(n), τ Q is the second quartile value of |y(n)|, τ r is the middle value.

3. The method for open set recognition of underwater acoustic communication signals based on a multi-task learning network according to claim 1 is characterized in that: The image domain mapping multi-task learning intelligent detection and recognition module includes an image domain mapping submodule and a multi-task learning network; The image domain mapping submodule is used to perform short-time Fourier transform on the preprocessed signal; The multi-task learning network processes the output of the image domain mapping submodule to obtain a preliminary prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be measured; wherein the multi-task learning network includes a YOLOv5s target detection network; The preliminary prediction probability vector is expressed as follows: P={P 2FSK ,P 4FSK ,P 8FSK ,P PSK ,P OFDM } Among them, P is the initial prediction probability vector, P 2FSK is the binary frequency shift keying probability, P 4FSK is the quaternary frequency shift keying probability, P 8FSK is the octal frequency shift keying probability, P PSK is the phase shift keying probability, P OFDM is the OFDM probability.

4. The method for open set recognition of underwater acoustic communication signals based on a multi-task learning network according to claim 3 is characterized in that: Replacing the convolution structure in the multi-task learning network with a wavelet convolution structure; The wavelet convolution structure includes a binarization unit, a first convolution unit, a two-dimensional Haar wavelet transform unit, a second convolution unit, an inverse wavelet transform unit and a summation unit; The binarization unit is used to perform a binarization operation on the input of the wavelet convolution structure; The first convolution unit is used to perform a convolution operation on the output of the binarization unit; The two-dimensional Haar wavelet transform unit is used to perform a two-dimensional Haar wavelet transform on the output of the binarization unit; The second convolution unit is used to perform a convolution operation on the output of the two-dimensional Haar wavelet transform unit; The inverse wavelet transform unit is used to perform an inverse wavelet transform operation on the output of the second convolution unit; The summing unit is used to add the output of the inverse wavelet transform unit and the output of the first convolution unit.

5. The method for open set recognition of underwater acoustic communication signals based on a multi-task learning network according to claim 1 is characterized in that: The narrowband processing of the underwater acoustic communication signal to be measured according to the relative time-frequency position of the underwater acoustic communication signal to be measured is expressed by the following formula: Among them, s narrow (n) is the underwater acoustic communication signal to be tested after narrowband processing, w is the relative duration of the signal, x center is the relative position of the midpoint in time, y center is the relative position of the frequency midpoint, h is the relative bandwidth of the signal, L={x center ,y center , w, h} is the relative time-frequency position of the underwater acoustic communication signal to be measured, R is the step size between time frames, m is the time dimension of the time-spectrum graph, k is the frequency dimension of the time-spectrum graph, T is the signal duration, F s is the sampling rate, and n is the time dimension of the underwater acoustic communication signal to be measured after narrowband processing.

6. The method for open set recognition of underwater acoustic communication signals based on a multi-task learning network according to claim 1, characterized in that: The ViT modulation identification auxiliary module includes a square spectrum estimation unit and a ViT network; The square spectrum estimation unit is used to perform square spectrum estimation on the underwater acoustic communication signal to be measured after narrowband processing; The ViT network is used to process the output of the square spectrum estimation unit to obtain a modified prediction probability; wherein the modified probability includes a binary phase modulation probability and an orthogonal phase shift keying probability.

7. The method for open set recognition of underwater acoustic communication signals based on a multi-task learning network according to claim 3 or 6, characterized in that: The step five specifically includes: According to the modified prediction probability, the modified binary phase modulation probability and the quadrature phase shift keying probability are calculated; The phase shift keying probability in the preliminary prediction probability vector is replaced by the modified binary phase modulation probability and quadrature phase shift keying probability to obtain the prediction probability vector of the underwater acoustic communication signal to be tested.

8. The method for open set recognition of underwater acoustic communication signals based on a multi-task learning network according to claim 7 is characterized in that: The modified binary phase modulation probability and quadrature phase shift keying probability obtained by calculating the modified predicted probability are expressed as follows: P BPSK =λP PSK +(1-λ)P B P QPSK =λP PSK +(1-λ)P Q Among them, P BPSK is the modified binary phase modulation probability, λ is the weight parameter, P PSK is the phase shift keying probability, P B is the binary phase modulation probability, P QPSK is the modified QPSK probability, P Q is the quadrature phase shift keying probability.

9. The method for open set recognition of underwater acoustic communication signals based on a multi-task learning network according to claim 1, characterized in that: The probabilistic decision open set recognition module includes a distance module, a correction vector module and a prediction module; The distance module is used to calculate the Euclidean distance between the predicted probability vector of the underwater acoustic communication signal to be measured and the central prediction vectors of multiple modulation types, which is specifically expressed by the following formula: D i JP-CV i Among them, D i is the Euclidean distance between the predicted probability vector of the underwater acoustic communication signal to be tested and the central prediction vector of the i-th modulation type, P is the predicted probability vector of the underwater acoustic communication signal to be tested, CV i is the center prediction vector of the i-th class; The correction vector module is used to input the output of the distance module into the cumulative distribution functions of multiple modulation types respectively to obtain corresponding probability correction vectors, and calculate the known modulation type probability prediction vector and the unknown modulation type probability of the underwater acoustic communication signal to be tested according to the probability correction vector, which is specifically expressed by the following formula; V known =P*W V unkown =P·(1-W) T Among them, V known is the probability prediction vector of the known modulation type, W is the probability correction vector, V unkown is the probability of unknown modulation type, * is Hadmard multiplication; The prediction module is used to normalize the output of the correction vector module and calculate the category probability, and the maximum category probability is used as the modulation category of the underwater acoustic communication signal to be tested, which is specifically expressed by the following formula: Where P′ is the category probability, softmax is the softmax function, and V is the known modulation type probability prediction vector or the unknown modulation type probability.

10. An open set recognition system for underwater acoustic communication signals based on a multi-task learning network, characterized by: include: The signal preprocessing module is used to input the underwater acoustic communication signal to be measured into the impulse noise signal preprocessing module to obtain a preprocessed signal; The initial prediction module is used to input the preprocessed signal into the image domain mapping multi-task learning intelligent detection and recognition module to obtain the initial prediction probability vector and the time-frequency relative position of the underwater acoustic communication signal to be tested; A narrowbanding module is used to perform narrowbanding on the underwater acoustic communication signal to be tested according to the relative time-frequency position of the underwater acoustic communication signal to be tested, so as to obtain the underwater acoustic communication signal to be tested after narrowbanding; A modified probability module is used to input the narrowband processed underwater acoustic communication signal to be tested into the ViT modulation identification auxiliary module to obtain a modified prediction probability; A correction module is used to correct the preliminary prediction probability vector according to the corrected prediction probability to obtain a prediction probability vector of the underwater acoustic communication signal to be measured; The recognition module is used to correct the predicted probability vector of the underwater acoustic communication signal to be tested according to the probability decision open set recognition module, obtain the open set recognition result of the modulation mode of the underwater acoustic communication signal to be tested, and complete the open set recognition of the underwater acoustic communication signal.

Images