A method and device for passive fusion detection of underwater acoustic communication signals based on multi-node multi-scale feature fusion

Abstract

The application provides a kind of passive fusion detection method and device for underwater acoustic communication signal based on multi-node multi-scale feature fusion.The method comprises deploying multiple hydrophone nodes at multiple spatial positions in a monitored area, receiving underwater acoustic communication signals using the multiple hydrophone nodes, preprocessing the received multiple underwater acoustic communication signals, the preprocessing includes noise reduction and performing short-time Fourier transform (STFT) on the multiple underwater acoustic communication signals after noise reduction to obtain corresponding multiple time-frequency diagrams;Input the multiple time-frequency diagrams into a pre-set multi-node multi-scale feature fusion network to obtain fused features;Input the fused features into a classifier to detect whether there is a communication signal in the fused features.The application can improve the information utilization of the received signal and improve the detection rate under low signal-to-noise ratio.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of passive detection of underwater acoustic signals, and particularly relates to a passive fusion detection method and device for underwater acoustic communication signals based on multi-node multi-scale feature fusion. BACKGROUND

[0002] Passive detection of underwater acoustic signals refers to a process of determining whether a target signal of interest exists from a passively received underwater acoustic signal, and is an important research content in the field of ocean acoustic monitoring. Underwater acoustic signals are mainly divided into types such as ocean ambient noise, marine animal call, ship radiated noise, active sonar signal and underwater acoustic communication (UWAC) signal according to their sound source and waveform characteristics. Among them, the UWAC signal is the carrier of information transmission underwater, and its signal waveform carries the transmitted information, so it is of great significance to improve the efficiency of underwater acoustic communication network and underwater acoustic perception ability to realize passive detection of the UWAC signal.

[0003] Due to the strong spatial variability of the underwater acoustic channel, the signal transmission path is related to the sound speed profile, ocean geography and other factors, and there is a large difference in signal strength and signal quality received by the receiving hydrophone at different depths and different directions. In order to improve the signal perception ability, multiple nodes need to be placed at multiple spatial positions to receive signals, and the signal fusion perception is used to overcome the problem of not receiving, not receiving completely and not judging accurately caused by the spatial difference of sound wave propagation. Signal detection based on multi-node information fusion can process data and information from multiple information sources from multiple angles and levels, thereby improving the accuracy of state and identity estimation, and can solve this problem to a certain extent, thereby improving the accuracy and reliability of the detection of the UWAC signal.

[0004] In the fusion, the traditional method is mainly based on the fusion of the detection values of each sensor node to form a fusion likelihood ratio (LRT) by using the Neyman-Pearson criterion (N-P criterion) and the minimum risk Bayes criterion. The main way of fusion is to realize fusion at the decision level and realize fusion at the feature level. The decision level fusion uses single-bit quantization, each node forms a hard decision, and finally realizes decision level fusion in the fusion center, and the loss of fusion data is the largest. In the research of feature level fusion, there is a multi-bit quantization method, and the fusion performance is improved compared with the decision level fusion. However, the above-mentioned traditional signal fusion detection methods are mostly based on domain knowledge, and the test statistic is constructed, and the corresponding threshold needs to be set according to different environments. The performance is seriously reduced under the condition of low signal-to-noise ratio, and some methods also require certain prior knowledge of signal and noise, which is difficult to work effectively under the condition of third-party reception. In order to extract more effective and robust classification features and improve the robustness to the environment, in recent years, some literatures have proposed some methods based on deep learning, which automatically complete the feature extraction and classification in the traditional method by deep neural network.

[0005] The research on passive fusion detection of signals by using deep learning mainly aims at radio signals, and there is no passive fusion detection of UWAC signals by using deep learning. Moreover, these passive fusion detection methods for radio signals mainly realize multi-node fusion at the decision level to realize the final signal detection; similarly, the modulation recognition network MHFNet which realizes fusion by using features loses a large amount of information in feature extraction and fusion, resulting in low signal detection accuracy. The research on feature extraction and fusion at the feature level can improve the information utilization of received signals and has important value for improving the accuracy of signal detection. SUMMARY

[0006] In view of the problems of the existing passive fusion detection method that too much useful information is lost in the decision level fusion and the detection rate is low under the condition of low signal-to-noise ratio, the present application provides a passive fusion detection method and device for UWAC signals based on multi-node multi-scale feature fusion, which realizes the full use of the information of the received signals of multiple nodes by extracting the multi-scale features of multiple nodes and realizing the fusion of the features, and finally realizes the passive fusion detection of the UWAC signals.

[0007] In the first aspect, the present application provides a passive fusion detection method for underwater acoustic communication signals based on multi-node multi-scale feature fusion, which is used in a monitored area and is arranged with multiple hydrophone nodes. The underwater acoustic communication signals are received by using the multiple hydrophone nodes, and the method comprises the following steps:

[0008] Step 1: preprocessing a plurality of received underwater acoustic communication signals, the preprocessing comprising noise reduction and short-time Fourier transform (STFT) on the plurality of noise-reduced underwater acoustic communication signals to obtain a plurality of corresponding time-frequency diagrams;

[0009] Step 2: inputting the plurality of time-frequency diagrams into a preset multi-node multi-scale feature fusion network to obtain fused features;

[0010] Step 3: inputting the fused features into a classifier to detect whether communication signals exist in the fused features.

[0011] Further, the multi-node multi-scale feature fusion comprises a feature extraction module and a feature fusion module; wherein the feature extraction module comprises a plurality of basic branches corresponding to the plurality of time-frequency diagrams, each basic branch comprising a feature extraction unit and a feature reconstruction unit, the feature extraction unit being configured to extract multi-scale features of a corresponding time-frequency diagram, and the feature reconstruction unit being configured to reconstruct the size dimensions of the multi-scale features to be consistent; the feature fusion module is configured to first fuse the multi-scale features of a single time-frequency diagram to obtain features of the single time-frequency diagram, and then fuse the features of a plurality of time-frequency diagrams to obtain fused features.

[0012] Further, the feature extraction unit adopts an improved ResNet network; the improved ResNet network refers to a network formed after removing the average pooling layer and the fully connected layer in the ResNet network.

[0013] Further, the feature reconstruction unit comprises a convolution layer and a transposed convolution layer connected in sequence.

[0014] Further, the feature fusion module comprises a group grouping splicing (G-concat) layer and a group convolution (G-conv) layer.

[0015] Correspondingly, the multi-scale features of a single time-frequency diagram are fused to obtain features of the single time-frequency diagram, specifically comprising: the G-concat layer is used to fuse the multi-scale features of the single time-frequency diagram to obtain the features x of the single time-frequency diagram according to the following formula: i :

[0016]

[0017] wherein i represents the number of a hydrophone node, and represent two highest layer feature vectors in the multi-scale features of the time-frequency diagram corresponding to the hydrophone node i, m i represents features of the remaining scales in the multi-scale features other than and

[0018] ​The features of the plurality of time-frequency maps are fused to obtain fused features, specifically including: performing two group convolutions in sequence by using a group convolution G-conv layer according to the following formula to obtain the fused features U={u1, u2};

[0019]

[0020] wherein W and V represent intermediate fused features output by the first group convolution, and represent two different convolution kernels corresponding to the hydrophone node i in the first group convolution process; and represent two different convolution kernels in the second group convolution process.

[0021] Further, the classifier adopts a binary classification network based on CNN.

[0022] Further, the method further comprises: jointly training the preset multi-node multi-scale feature fusion network and the classifier in advance by using a transfer data training set; wherein the transfer data training set refers to a data set composed of data having the same distribution as the transmission signal in the test data.

[0023] In a second aspect, the present application provides a device for passive fusion detection of underwater acoustic communication signals based on multi-node multi-scale feature fusion, comprising:

[0024] A plurality of hydrophone nodes are arranged at a plurality of spatial positions in a monitored area and are used to receive underwater acoustic communication signals.

[0025] A preprocessing module is used to preprocess the received plurality of underwater acoustic communication signals, and the preprocessing includes noise reduction and short-time Fourier transform (STFT) on the plurality of underwater acoustic communication signals after noise reduction to obtain a plurality of corresponding time-frequency maps.

[0026] A fusion network module is used to input the plurality of time-frequency maps into a preset multi-node multi-scale feature fusion network to obtain fused features.

[0027] A classification module is used to input the fused features into a classifier to detect whether there is a communication signal in the fused features.

[0028] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the method of the first aspect.

[0029] In a fourth aspect, the present application provides a non-transitory computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the method according to the first aspect.

[0030] Advantages of the present application:

[0031] (1) The present application uses a time-frequency graph as the input of the fusion network. Since the time-frequency graph depicts the energy distribution of the signal at different times and different frequencies, it has the representation advantages in time domain and frequency domain, and contains rich features. Therefore, by extracting and fusing the features of the time-frequency graph at the feature level, the detection of the signal can be realized, the information utilization of the received signal can be improved, and the detection rate under low signal-to-noise ratio can be improved.

[0032] (2) The present application proposes a signal processing flow for multi-node fusion passive detection of UWAC signals under low signal-to-noise ratio conditions, that is, through preprocessing, MMFFN network and finally a CNN-based binary classification network, the detection of UWAC signals is realized. The method can achieve a high detection rate under low signal-to-noise ratio conditions.

[0033] (3) The present application designs a multi-node multi-scale feature fusion deep learning network MMFFN, which can realize multi-scale extraction and fusion of features of multiple received source signals, and fully utilize the information of multi-node received signals. The Feature-Extraction and Feature-Fusion modules are innovatively designed in the network. The Feature-Extraction module innovatively uses the ResNet network as the backbone network for multi-scale feature extraction and forms a BasicNet, so that deeper features can be extracted without the problem of gradient explosion. In order to realize the fusion of multi-node multi-scale features, the Feature-Fusion module proposes to use grouping connection and grouping convolution to group and step by step realize the fusion of multi-node and multi-scale features.

[0034] (4) The experimental results show that when the number of nodes is 3, the present application can achieve a detection rate of 98% when the E s / N0=-2dB of the three nodes. It can effectively detect common UWAC signals under unknown channels, including M-ary Frequency-shift Keying (MFSK), M-ary Phase Shift Keying (MPSK), Orthogonal Frequency Division Multiplexing (OFDM) and Linear Frequency Modulation (LFM) signals. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A flowchart of a passive fusion detection method for underwater acoustic communication signals based on multi-node multi-scale feature fusion is provided for an embodiment of the present application.

[0036] Figure 2 An architecture diagram of a passive fusion detection model for underwater acoustic communication signals based on multi-node multi-scale feature fusion is provided for an embodiment of the present application.

[0037] Figure 3 An improved ResNet network is provided for an embodiment of the present application.

[0038] Figure 4 A basic branch BasicNet is provided for an embodiment of the present application.

[0039] Figure 5 A grouping splicing G-concat layer is provided for an embodiment of the present application.

[0040] Figure 6 A grouping convolution G-conv layer is provided for an embodiment of the present application.

[0041] Figure 7 A classifier is provided for an embodiment of the present application.

[0042] Figure 8 Test results of different node numbers are provided for an embodiment of the present application.

[0043] Figure 9 Impulse responses of different channels are provided for an embodiment of the present application.

[0044] Figure 10 Test results of different channels are provided for an embodiment of the present application.

[0045] Figure 11 A structure diagram of a passive fusion detection device for underwater acoustic communication signals based on multi-node multi-scale feature fusion is provided for an embodiment of the present application.

[0046] Figure 12 A structure block diagram of an electronic device is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0048] The application background of the technical scheme of the present application is that a plurality of hydrophone nodes (hereinafter also referred to as nodes) are arranged at a plurality of spatial positions in a monitored area, and a plurality of hydrophone nodes are used to receive underwater acoustic communication signals. In order to better understand the technical scheme of the present application, the technical problem to be solved by the present application is modeled as follows based on the above application background.

[0049] In the context of multi-node reception, the UWAC signal received by the i-th node can be represented as:

[0050]

[0051] In the formula, r i (t) represents the received signal of the i-th node, x(t) represents the target signal, h i (t) represents the underwater acoustic channel impulse response of the target signal to the i-th node, w i (t) is the ocean ambient noise at the i-th node, N node is the number of nodes.

[0052] The present application assumes that there is no relative displacement between the target sound source and the receiving node, and the underwater acoustic channel can be modeled as a coherent multipath channel, and the model can be represented as:

[0053]

[0054] In the formula, δ(t) is an impulse function, K i is the total number of sound rays reaching the i-th node, i=1, 2, …, N node , A m is the amplitude of the m-th sound ray, τ m is the transmission delay of the m-th sound ray.

[0055] For the UWAC signal, the symbol signal-to-noise ratio is usually used to measure the noise size of the received signal, which is defined as the ratio of the energy of each symbol to the noise power spectral density, denoted as E s / N0; wherein E s represents the signal energy of each symbol, and N0 represents the noise power spectral density.

[0056] Therefore, after sampling the model, it can be represented as:

[0057]

[0058] wherein r i (n), x(n), h i (n) and w i (n) are respectively r i (t), x(t), h i (t) and wi (t) at a sampling rate f s is sampled.

[0059] Based on the above, the passive fusion detection of the UWAC signal of the multi-node is a process of judging whether there is a target communication signal by jointly processing the signals r node (n) received by N i nodes in the same time period, i.e. i = 1, 2, …, N node Therefore, it can be described as a binary hypothesis testing problem:

[0060]

[0061] Wherein, H1 is defined as the existence of the target signal, and H0 is defined as the non-existence of the target signal. In the detection model, the performance of the signal detection algorithm can be evaluated by the false alarm probability (P f ) and the detection probability (P d ):

[0062] P f = P(H1|H0) (5)

[0063] P d = P(H1|H1) (6)

[0064] In order to solve the above technical problems, the present application provides a multi-node multi-scale feature fusion based passive fusion detection method for UWAC signal, as shown in the following steps: Figure 1

[0065] S101: Preprocessing the received multiple underwater acoustic communication signals;

[0066] Specifically, since the UWAC signal is usually distributed above 2 kHz, the signal preprocessing stage first carries out noise reduction preprocessing on the multi-node sampling received signal, including: filtering out the noise below 2 kHz and the radiation noise of the ship by frequency domain filtering.

[0067] Then, the multiple underwater acoustic communication signals after noise reduction are subjected to short-time Fourier transform (STFT) to obtain the corresponding multiple time-frequency diagrams. Specifically, the noise-reduced signals are subjected to segmentation processing, taking 1s as the time window, and the data r(n) in the window is normalized and subjected to STFT, and the expression is as follows:

[0068]

[0069] Wherein, N is the number of sampling points, is the window function, which is the Hanning window here, the window length is 300, and the step length is 30, to obtain the time-frequency diagram T m with fixed output size. ​

[0070] S102: Input multiple time-frequency maps into a preset multi-node multi-scale feature fusion network to obtain the fused features;

[0071] S103: Input the fused features into the classifier to detect whether there are communication signals in the fused features.

[0072] Specifically, the time-frequency graph depicts the energy distribution of a signal at different times and frequencies. It has the advantages of both time and frequency domain representation and contains rich features. Therefore, this embodiment of the invention uses the time-frequency graph as the input of the fusion network. At the feature level, the time-frequency graph is used to extract and fuse features to achieve signal detection, which can improve the utilization of information of the received signal and improve the detection rate under low signal-to-noise ratio.

[0073] In one embodiment, such as Figure 2 As shown in the figure, this embodiment of the invention provides a passive fusion detection model for underwater acoustic communication signals. This network model includes a signal preprocessor, a multi-node multi-scale feature fusion network (MMFFN), and a binary classifier. The detection process of this network model includes: first, the signal preprocessor performs noise reduction preprocessing and STFT time-frequency transformation on the multi-node sampled received signals; then, the resulting multi-node time-frequency map is used as input and fed into the MMFFN network to extract and fuse multi-node multi-scale features; finally, the fused features are fed into the binary classifier for further feature extraction and classification, thereby completing the presence detection of the signal. The workflow of the signal preprocessor is the same as in the above embodiment and will not be repeated here. The MMFFN will be described in detail below.

[0074] Multi-scale feature extraction and fusion is an important concept in the field of computer vision, especially in tasks such as object detection and image recognition. This technology combines features of different scales to better capture various forms of targets, thereby improving the accuracy and stability of detection. The MMFFN proposed in the embodiments of the present application is an end-to-end trainable convolutional neural network that can extract and fuse multi-node multi-scale features. The network as a whole consists of two parts, namely the Feature-Extraction module and the Feature-Fusion module. Among them, Feature-Extraction includes multiple BasicNet branches, and each BasicNet branch corresponds to a node's time-frequency graph; each BasicNet branch mainly completes the extraction of multi-scale features of the node; the Feature-Fusion module completes the connection and fusion of multi-node multi-scale features.

[0075] In the embodiments of the present application, each BasicNet branch includes a feature extraction unit and a feature reconstruction unit, the feature extraction unit is used to extract multi-scale features of the corresponding time-frequency graph, and the feature reconstruction unit is used to reconstruct the size dimension of the multi-scale features to be consistent. As an implementable way, the feature extraction unit uses an improved ResNet network; the improved ResNet network refers to the network formed after removing the average pooling layer and the fully connected layer in the ResNet network. This is because the ResNet introduces residual connection, allowing the input signal to be directly transmitted to the deeper layer of the network, so that the ResNet can train a very deep network structure without performance degradation. However, the present application needs to extract multi-scale features using a deeper network, so the present application selects the original ResNet for modification, and then uses it as the backbone network. The structure of the modified ResNet is as shown in Figure 3 To extract multi-scale features, the embodiments of the present application remove the average pooling layer and the fully connected layer in the original ResNet, and retain the bottom convolution block and the four-layer bottleneck structure. As shown in Figure 4 A single time-frequency graph is input to the backbone network as input, and five scale convolution features R-1 to R-5 are obtained, as shown in Figure 4 The blue square in the middle is the stacking of the features.

[0076] Since the five scale features obtained at this time cannot be directly fused, the present application also designs a feature reconstruction module in the BasicNet branch to realize up-sampling through convolution and transposed convolution, so as to reconstruct the size dimension of the features to be consistent, and the structure is as shown in Figure 4The middle gray and red blocks are shown. When reconstructing, because the top layer features of R-5 are encoded with high-dimensional information, the receptive field is rich, so this feature is used to enable the top end to achieve the classification supervision function, and the embodiment of the application outputs an activation vector of C categories (at this time C = 2) And the bottom layer features of R-1 to R-4 are beneficial to enhance the top classification, and the corresponding output is a single-channel feature map

[0077]

[0078] In the embodiment of the application, the Feature Fusion module includes two parts: a Group concatenation (G-concat) layer and a Group convolution (G-conv) layer. The G-concat layer completes the connection and combination of multi-node multi-scale features, and the G-conv layer realizes supervised feature fusion. The process of the G-concat layer is as shown in Figure 5 The figure shows that num represents the number of nodes. After preprocessing, T m After the BasicNet, the activation vector A i and the single-channel feature map m i are obtained. node When the number of nodes is N i , there are and In order for the activation vector to achieve a supervisory role, the bottom layer feature maps of R-1 to R-4 are copied respectively, and two top layer activation vectors are realized to achieve group connection, so as to obtain the combination of the feature map and the activation vector. For a single time-frequency graph, the above combination process can be expressed by the following formula:

[0079]

[0080] Wherein, A i is composed. After obtaining the combination of the feature map and the activation vector, we realize supervised feature fusion through group convolution, and the input and output are as shown in Figure 6 The figure shows that num represents the number of nodes. Group convolution can increase the diagonal correlation between adjacent layers of filters, thereby helping the model to learn more discriminative features. The specific implementation is as shown in Figure 6

[0081] In the first group convolution, we take as a group to realize convolution, so as to realize the multi-scale fusion of each node. The number of input channels of the first group convolution is as formula (9).

[0082]

[0083] At this time, the length of F is 5*C*N node Where 5 represents 5 layers of features of R-(1-5), C represents a classification category (at this time C=2), and N node represents the number of nodes. The output channel number of the first grouped convolution is C*N node , and the grouping number is N node . Therefore, the input channel number of each group is 5*2, and the output channel number of each group is 2. We define different convolution kernels in the first grouped convolution process as Therefore, the output of the first grouped convolution can be represented as:

[0084]

[0085] In the second grouped convolution, we implement convolution by taking W and V as a group respectively. The input channel number of the second grouped convolution is 2*N node , the output channel number is 2, and the grouping number is 2. Therefore, the input channel number of each group is N node , and the output channel number of each group is 1. The two convolution kernels of the second grouped convolution are Therefore, the output U={u1, u2} of the second grouped convolution can be represented as:

[0086]

[0087] In the embodiment of the application, the Feature-Extraction module innovatively takes the ResNet network as a backbone network for multi-scale feature extraction and forms a BasicNet, so that deeper features can be extracted without the problem of gradient explosion. In the Feature-Fusion module, a grouping and step-by-step fusion method using grouped connection and grouped convolution is proposed to fuse multi-node and multi-scale features. Therefore, the MMFFN provided by the application can realize multi-scale extraction and fusion of features of signals received by multiple receiving sources, and fully utilize multi-node received signal information.

[0088] On the basis of the above embodiments, after obtaining the fused features, a binary classification network C based on CNN is constructed as a classifier in the embodiment of the application to automatically extract detection features for representing the presence or absence of signals in the features and complete binary classification recognition. The network structure is as follows: Figure 7The working process of the classifier includes: firstly, inputting the fused features into a binary classification module; then extracting and classifying high-dimensional features by using two convolution blocks, two maximum pooling layers, two fully connected layers and a Softmax activation function; outputting the probabilities of the received signal containing a communication signal and being pure noise respectively, and the setting of the activation function is the same as that of the decision maker. At this time, the prediction label probability vector output by the Softmax classifier is may be expressed as:

[0089]

[0090] S = f CNN (U) (13)

[0091] wherein, L p represents a prediction label value, S represents an input vector of Softmax, f CNN represents a nonlinear function constituted by the network before the Softmax layer in the classifier C.

[0092] It can be understood that the MMFFN and the classifier described above need to be trained first, and only after being trained do they have the fusion and classification capabilities respectively. Therefore, on the basis of the above embodiments, the embodiments of the present application provide a network training method as follows.

[0093] (1) Constructing a transfer training data set

[0094] The core of the network training process is to find the mapping of the feature space based on the sampled training data. In order to enable the network trained under the training set to still effectively realize this mapping on the test data, it is usually expected that the offline training data set and the online test data set have the same distribution. However, in the UWAC signal detection application, when facing unfamiliar sea areas and different communication scenes, it is difficult to accurately obtain the information of the target channel, so it is difficult to obtain a sufficient number of training data with the same distribution as the test data. Therefore, the embodiments of the present application use the transfer learning idea to construct a data model containing the same distribution as the transmitted signal in the test data, generate a transfer data training set, and complete the training of the network. The transfer data model is:

[0095]

[0096] wherein, represents a received signal, x(n) is a transmitted signal, the signal type is the same as formula (3), is noise.

[0097] Compared with formula (3), formula (14) simplifies the influence of the underwater acoustic multipath channel, although the distribution of the data generated by the two models changes, but both contain the same set of transmitted signals, that is, The same useful signal component is contained in the waveform of r(n), so it has the ability of migration. Considering the modulation parameters of typical UWAC signals and the working parameters of existing laboratory experimental equipment, and on the basis of the model described in formulas (1) to (3), the signal is generated. The detailed signal parameters are shown in Table 1, wherein " / " represents that this parameter does not need to be set, and "{}" represents random selection within the given set. The signal sampling rate is uniformly set to 48 kHz, the length is 8192 sampling points, about 171 ms, at the same time, the center frequency of the LFM signal and the carrier frequency of other signals are set to 15-16 kHz, and the OFDM subcarrier is BPSK or QPSK.

[0098] Table 1 Signal parameters

[0099]

[0100] In the training phase, 30 samples of each hydrophone per signal type are generated under different signal-to-noise ratio conditions. In order to make the number of signals equal to the number of pure noise, the number of corresponding pure noise samples is 30*6=180. In the case of a single node, we set the symbol signal-to-noise ratio range of each node to [-6, 20] dB with a step of 2 dB. At this time, we get a total of 5040 samples. In the case of three nodes, we set the symbol signal-to-noise ratio range of each node to [-6, 14] dB, and the step sizes of the three nodes are 4, 6, and 8 dB respectively. At this time, there are a total of 25920 samples.

[0101] (2) Model training

[0102] In the signal detection model as shown in Figure 2 , the effective detection of signals is realized by jointly training the MMFFN and binary classification network. During training, the cross-entropy loss function is used, which can be expressed as:

[0103]

[0104] Where S is the parameter of the network model, N b is the number of samples in a batch. When the loss function converges, the network training is completed. The learning rate of the model is 0.001, the batch size N b = 32, the loss function is optimized by using the SGD optimizer (momentum = 0.9, weight_decay = 0.0001), and all data is iterated until the overall loss function of the model converges to stop training.

[0105] In a multi-hydrophone receiving system, due to the receiving position, environment, distance, the transmitted signal reaches the receiving end through different channels. In order to be close to the actual application scene, we select two different water areas as water environment based on Argo ocean database, which are recorded as water area 1 and water area 2. Water area 1 is a typical deep-sea sound profile, which contains typical marine acoustic characteristics. Water area 2 is a typical shallow-sea sound profile. In order to fully reflect the characteristics of the underwater acoustic channel as much as possible, the sound source position is selected at the transition layer, the main transition layer and the sound channel axis. Then the Bellhop underwater acoustic channel simulation software is used to generate linear time-invariant underwater acoustic channels under different transmission conditions, and the parameters are shown in Table 2.

[0106] Table 2 Underwater acoustic channel parameters

[0107]

[0108] The transfer functions of each channel are as follows:

[0109] H1(z) = 0.04 + z -353 + 0.508z -570 + 0.283z -644 (16)

[0110] H2(z) = 0.32 + 0.45z -48 + z -61 + 0.9318z -267 (17)

[0111] H3(z) = 0.68 + z -184 + 0.882z -403 (18)

[0112] H4(z) = 0.2039 + 0.3178z -2327 + 0.0759z -2938 + z -3071 (19)

[0113] + 0.5875z -3711 + 0.8845z -4470 + 0.3302z -11697

[0114] In the test phase, according to formula (14), and according to the number of hydrophones, the signal sample is randomly selected to generate a test set. We select different signal-to-noise ratio combinations according to different test purposes, generate 300 signal samples and 300 pure noise samples under each signal-to-noise ratio, that is, a total of 600 test samples under each signal-to-noise ratio.

[0115] In order to verify the effectiveness of the scheme, the following effectiveness verification experimental data is also provided.

[0116] (1) Effectiveness verification of multi-scale feature fusion

[0117] The effectiveness of the method of the present application is illustrated by comparing the detection performance difference between the method of the present application and several deep learning methods under Gaussian noise condition. The results are shown in Table 3. The false alarm rate of the method of the present application is 0.33%, and the detection rate reaches 98.33% under SNR = -15dB. It can be seen that the method of the present application performs better in both detection rate and false alarm rate. In fact, most existing deep learning-based detection methods directly extract features from the original speech signal, while the method of the present application extracts multi-scale features for detection, thereby achieving better results.

[0118] Table 3 Comparison of detection performance of different methods

[0119]

[0120] Literature 1 in Table 3 is "Li Y, Wang B, Shao G, et al. Blind Detection of Underwater Acoustic Communication Signals Based on Deep Learning [J]. IEEE Access, 2020, 8: 204114-204131"; Literature 2 is "Ke D, Huang Z, Wang X, et al. Blind Detection Techniques for Non-Cooperative Communication Signals Based on Deep Learning [J]. IEEE Access, 2019, 7: 89218-89225"; Literature 3 is "Gao J, Yi X, Zhong C, et al. Deep Learning for Spectrum Sensing [J]. IEEE Wireless Communications Letters, 2019, 8(6): 1727-1730".

[0121] (2) Effectiveness verification of multi-node multi-scale feature fusion

[0122] To verify the effectiveness of multi-node multi-scale feature fusion for UWAC signal detection, we increase the number of nodes to achieve signal detection, and compare it with the detection performance of single node. At this time, in order to illustrate the detection gain brought by multi-node feature passive fusion detection, the multi-node test set is set so that the signal-to-noise ratio of each node is equal, and the channel is randomly selected according to the number of hydrophones, and there are 600 test samples under each signal-to-noise ratio. Through such setting, the detection gain brought by the increase of signal-to-noise ratio of a node in multi-node test is avoided, so as to strengthen the detection gain brought by multi-node feature passive fusion detection. The trained single node, three nodes and five nodes network models are respectively tested in different node number test sets, and the detection results are as shown in Figure 8 .

[0123] It can be seen from the figure that the passive fusion detection of multiple nodes is improved compared with single node in detection rate, and the detection signal-to-noise ratio gain can reach about 3dB; and at lower signal-to-noise ratio, the multi-node has higher detection rate. For example, at E s / N0=-8dB, the single node detection rate is 20%, and the multi-node reaches 46%~48%. The detection rate of three nodes and five nodes is similar, at the same time, the false alarm rate of five nodes is 0.13%, and the false alarm rate of three nodes is 1.4%, and the false alarm rate of five nodes is better. This shows that the increase of node number can bring better detection gain.

[0124] (3) Algorithm reliability verification

[0125] To verify the reliability of the algorithm, experiments in different underwater acoustic channels are carried out to verify the robustness of the method to different underwater acoustic multipath channels. In order to ensure the diversity of channel characteristics, the above underwater acoustic channel parameters are used. Under the setting of 48kHz sampling rate, the maximum propagation delay of the four underwater acoustic channels relative to the direct diameter is 18.25ms, 26.27ms, 74.60ms, 6.48ms and 87.77ms. Figure 9 The impulse response of each channel is given in Table 1. The experimental results are shown in Figure 10 It can be seen that the detection performance under different underwater acoustic channel conditions is close, and the difference is small, which proves the robustness of the method to different underwater acoustic multipath channels.

[0126] As shown in Figure 11 , the embodiment of the application provides a kind of underwater acoustic communication signal passive fusion detection device based on multi-node multi-scale feature fusion, comprising: multiple hydrophone nodes, pre-processing module, fusion network module and classification module.

[0127] The multiple hydrophone nodes are arranged at multiple spatial positions of a monitored area to receive underwater acoustic communication signals; a preprocessing module is configured to preprocess the received multiple underwater acoustic communication signals, the preprocessing including noise reduction and short-time Fourier transform (STFT) on the multiple underwater acoustic communication signals after noise reduction to obtain corresponding multiple time-frequency diagrams; a fusion network module is configured to input the multiple time-frequency diagrams into a preset multi-node multi-scale feature fusion network to obtain fused features; and a classification module is configured to input the fused features into a classifier to detect whether the fused features contain communication signals.

[0128] It should be noted that the underwater acoustic communication signal passive fusion detection device based on multi-node multi-scale feature fusion provided by the embodiments of the present application is to realize the above-mentioned methods, and the functions thereof can be referred to the above-mentioned method embodiments, which will not be described here.

[0129] Figure 12 An example of an entity structure diagram of an electronic device is shown in Figure 12 As shown, the electronic device can include a processor 1201, a communications interface 1202, a memory 1203, and a communications bus 1204, wherein the processor 1201, the communications interface 1202, and the memory 1203 complete mutual communication through the communications bus 1204. The processor 1201 can invoke a logical instruction in the memory 1203 to execute the underwater acoustic communication signal passive fusion detection method, which includes arranging multiple hydrophone nodes at multiple spatial positions of a monitored area, receiving underwater acoustic communication signals by using the multiple hydrophone nodes; preprocessing the received multiple underwater acoustic communication signals, the preprocessing including noise reduction and short-time Fourier transform (STFT) on the multiple underwater acoustic communication signals after noise reduction to obtain corresponding multiple time-frequency diagrams; inputting the multiple time-frequency diagrams into a preset multi-node multi-scale feature fusion network to obtain fused features; and inputting the fused features into a classifier to detect whether the fused features contain communication signals.

[0130] In addition, the logic instructions in the memory 1203 described above are implemented in the form of software function units and sold or used as independent products, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the part of the prior art that contributes essentially or the part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0131] The embodiments of the present application also provide a computer program product, which comprises a computer program stored on a non-transitory computer readable storage medium, and the computer program comprises program instructions. When the program instructions are executed by a computer, the computer can execute the method provided by the above-mentioned method embodiments, for example, comprising: deploying a plurality of hydrophone nodes at a plurality of spatial positions of a monitored area, receiving underwater acoustic communication signals by using the plurality of hydrophone nodes; preprocessing the received plurality of underwater acoustic communication signals, the preprocessing comprising noise reduction and performing short-time Fourier transform (STFT) on the plurality of underwater acoustic communication signals after noise reduction to obtain a plurality of corresponding time-frequency diagrams; inputting the plurality of time-frequency diagrams into a preset multi-node multi-scale feature fusion network to obtain fused features; and inputting the fused features into a classifier to detect whether there is a communication signal in the fused features.

[0132] The embodiments of the present application also provide a non-transitory computer readable storage medium, which stores a computer program. When the computer program is executed by a processor, the method provided by the above-mentioned method embodiments is implemented, for example, comprising: deploying a plurality of hydrophone nodes at a plurality of spatial positions of a monitored area, receiving underwater acoustic communication signals by using the plurality of hydrophone nodes; preprocessing the received plurality of underwater acoustic communication signals, the preprocessing comprising noise reduction and performing short-time Fourier transform (STFT) on the plurality of underwater acoustic communication signals after noise reduction to obtain a plurality of corresponding time-frequency diagrams; inputting the plurality of time-frequency diagrams into a preset multi-node multi-scale feature fusion network to obtain fused features; and inputting the fused features into a classifier to detect whether there is a communication signal in the fused features.

[0133] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0134] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A passive fusion detection method for underwater acoustic communication signals based on multi-node, multi-scale feature fusion, comprising deploying multiple hydrophone nodes at multiple spatial locations in the monitored area, and utilizing the multiple hydrophone nodes to receive underwater acoustic communication signals, characterized in that, include: Step 1: Preprocess the received multiple underwater acoustic communication signals. The preprocessing includes noise reduction and performing a short-time Fourier transform (STFT) on the noise-reduced multiple underwater acoustic communication signals to obtain the corresponding multiple time-frequency diagrams. Step 2: Input multiple time-frequency maps into a preset multi-node multi-scale feature fusion network to obtain the fused features; The multi-node, multi-scale feature fusion includes a feature extraction module and a feature fusion module. The feature extraction module includes multiple basic branches corresponding to multiple time-frequency maps. Each basic branch includes a feature extraction unit and a feature reconstruction unit. The feature extraction unit extracts multi-scale features from the corresponding time-frequency map, and the feature reconstruction unit reconstructs the size and dimensions of the multi-scale features to be consistent. The feature fusion module first fuses the multi-scale features of a single time-frequency map to obtain the features of that single time-frequency map, and then fuses the features of multiple time-frequency maps to obtain the fused features. The feature fusion module includes a group-to-group concatenation G-concat layer and a group-to-group convolution G-conv layer; Correspondingly, the multi-scale features of a single time-frequency map are fused to obtain the features of the single time-frequency map. Specifically, this includes fusing the features of the single time-frequency map using a G-concat layer according to the following formula. : Where i represents the number of the hydrophone node, and These represent the two highest-level feature vectors in the multi-scale features of the time-frequency plot corresponding to hydrophone node i. In addition to multi-scale features and Features at other scales besides those; The features of multiple time-frequency maps are fused to obtain the fused features. Specifically, this involves performing two grouped convolutions using a G-conv layer according to the following formula to obtain the fused features. ; in, and This represents the intermediate fused features from the first group convolution output. , ; and This represents the two different convolution kernels corresponding to hydrophone node i during the first group convolution process; and This represents two different convolution kernels in the second group convolution process; Step 3: Input the fused features into the classifier to detect whether there are communication signals in the fused features.

2. The passive fusion detection method for underwater acoustic communication signals based on multi-node, multi-scale feature fusion according to claim 1, characterized in that, The feature extraction unit uses an improved ResNet network; the improved ResNet network refers to the network formed by removing the average pooling layer and the fully connected layer from the ResNet network.

3. The passive fusion detection method for underwater acoustic communication signals based on multi-node, multi-scale feature fusion according to claim 1, characterized in that, The feature reconstruction unit includes a convolutional layer and a transposed convolutional layer connected sequentially.

4. A passive fusion detection method for underwater acoustic communication signals based on multi-node, multi-scale feature fusion according to any one of claims 1 to 3, characterized in that, The classifier uses a CNN-based binary classification network.

5. A passive fusion detection method for underwater acoustic communication signals based on multi-node, multi-scale feature fusion according to any one of claims 1 to 3, characterized in that, Also includes: The pre-set multi-node, multi-scale feature fusion network and classifier are jointly trained using a transfer data training set. The training set of the migration data refers to a dataset consisting of data that has the same distribution as the transmitted signals in the test data.

6. A passive fusion detection device for underwater acoustic communication signals based on multi-node, multi-scale feature fusion, characterized in that, include: Multiple hydrophone nodes are deployed at multiple spatial locations within the monitored area to receive underwater acoustic communication signals; The preprocessing module is used to preprocess multiple received underwater acoustic communication signals. The preprocessing includes noise reduction and performing short-time Fourier transform (STFT) on the noise-reduced underwater acoustic communication signals to obtain multiple corresponding time-frequency diagrams. The fusion network module is used to input multiple time-frequency maps into a preset multi-node multi-scale feature fusion network to obtain fused features. The multi-node, multi-scale feature fusion includes a feature extraction module and a feature fusion module. The feature extraction module includes multiple basic branches corresponding to multiple time-frequency maps. Each basic branch includes a feature extraction unit and a feature reconstruction unit. The feature extraction unit extracts multi-scale features from the corresponding time-frequency map, and the feature reconstruction unit reconstructs the size and dimensions of the multi-scale features to be consistent. The feature fusion module first fuses the multi-scale features of a single time-frequency map to obtain the features of that single time-frequency map, and then fuses the features of multiple time-frequency maps to obtain the fused features. The feature fusion module includes a group-to-group concatenation G-concat layer and a group-to-group convolution G-conv layer; Correspondingly, the multi-scale features of a single time-frequency map are fused to obtain the features of the single time-frequency map. Specifically, this includes fusing the features of the single time-frequency map using a G-concat layer according to the following formula. : Where i represents the number of the hydrophone node, and These represent the two highest-level feature vectors in the multi-scale features of the time-frequency plot corresponding to hydrophone node i. In addition to multi-scale features and Features at other scales besides those; The features of multiple time-frequency maps are fused to obtain the fused features. Specifically, this involves performing two grouped convolutions using a G-conv layer according to the following formula to obtain the fused features. ; in, and This represents the intermediate fused features from the first group convolution output. , ; and This represents the two different convolution kernels corresponding to hydrophone node i during the first group convolution process; and This represents two different convolution kernels in the second group convolution process; The classification module is used to input the fused features into the classifier to detect whether there are communication signals in the fused features.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

Images