Broadband multi-underwater acoustic signal multi-node fusion detection and identification method based on BF-YOLOv11

Abstract

The invention discloses a broadband multi-underwater acoustic signal multi-node fusion detection and identification method based on BF-YOLOv11, and finally realizes fusion detection and identification of broadband multi-underwater acoustic signals through a neural network model fused after YOLOv11 and a backbone network. The method comprises the following steps: firstly, preprocessing a signal received by multiple nodes to reduce environmental noise in the received signal, and carrying out time-frequency processing of short-time Fourier transform; the method comprises the steps that firstly, a signal time-frequency graph is input into BF-YOLOv11, features are extracted based on a backbone network of the YOLOv11, feature fusion of C3K2 and C2PSA layers of multiple nodes is achieved through a FusionNet structure, and then detection and recognition of multiple signals are achieved through a detection head and a classification head of the YOLOv11. And common underwater acoustic signals under an unknown channel can be effectively detected and identified.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic signal passive detection and recognition, and particularly relates to a multi-node fusion detection and recognition method for wideband multi-underwater acoustic signals based on BF-YOLOv11. Background Art

[0002] Underwater acoustic signal passive detection and recognition refers to the process of determining whether there is a communication signal of interest in the passively received underwater acoustic signal and realizing modulation recognition. Underwater acoustic signals are carriers for information transmission underwater, and their signal waveforms carry the transmitted information. Therefore, realizing passive detection and recognition of underwater acoustic signals is of great significance for improving the efficiency of underwater acoustic communication networks and underwater acoustic perception capabilities. In the actual underwater environment, in order to obtain more information, it is necessary to master a sufficiently wide frequency band range and use effective wideband signal detection techniques to discover useful signals. During underwater communication, the number of users increases sharply, and different wireless communication requirements occupy different frequency bands. In the civilian field, in order to use limited spectrum resources more efficiently, it is necessary to constantly sense the occupancy of different frequency bands so that other users can access the spectrum in a timely manner. In the military field, in order to ensure smooth and covert communication, signals need to be dispersed on different frequency bands for communication. Therefore, in order to capture and detect abnormal signals in a timely manner, it is necessary to detect multi-band communication signals on a wideband spectrum.

[0003] Due to the strong spatial variability of the underwater acoustic channel, the signal transmission path is related to factors such as the sound speed profile and ocean geography. There are significant differences in the received signal strength and signal quality when receiving hydrophones are at different depths and different azimuths. In order to improve the signal perception ability, it is necessary to deploy multiple nodes at multiple spatial positions to receive signals, and through multi-node signal fusion perception, to overcome the problems of not receiving, not receiving completely, and misjudging caused by the spatial differences in acoustic wave propagation. Signal detection and recognition 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 multi-node fusion detection and recognition of wideband multi-underwater acoustic signals.

[0004] Currently, there is no research on multi-node fusion detection and recognition of wideband multi-underwater acoustic signals. The research mainly focuses on detection and recognition in the case of a single node. The main methods include serial detection and recognition and parallel detection and recognition, that is, the acquisition of sub-frequency bands under a wideband is realized through a band-pass filter, and then the detection and recognition classifier for a single frequency band is used to realize the detection and recognition under a wideband. Such a method has high requirements for the design of the filter, requires an understanding of the characteristics of the transmitted and received signals, and there will be serious performance degradation in the detection of unknown signals. On the other hand, such methods require a high computational cost, the detection and recognition process is slow, and the workload is large.

[0005] Based on these studies, multi-node fusion detection and recognition at the decision level can be achieved through voting and other means. However, when conflicts occur in the decision results of different nodes, it is difficult for decision fusion methods to make effective fusions. During the actual underwater acoustic signal reconnaissance process, situations where conflicts occur in the decision results of different nodes often arise, and decision fusion methods are no longer applicable to underwater scenarios. Summary of the Invention

[0006] Currently, there is no research on wideband multi-underwater acoustic signal multi-node fusion detection and recognition. The research mainly focuses on detection and recognition in the case of a single node. The fusion detection and recognition at the decision level that can be achieved based on these studies will be affected by the decision results of different hydrophones. To achieve wideband multi-underwater acoustic signal multi-node fusion detection and recognition, the present invention proposes a wideband multi-underwater acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11, which converts the wideband detection and recognition problem of multi-underwater acoustic signals into an object detection problem in the field of deep learning. A neural network model is constructed by fusing YOLOv11 and a backbone network later. The received signals of multiple nodes are used as inputs. The features are extracted based on the backbone network of YOLOv11, and then a FusionNet structure based on the BiFPN model is designed to achieve feature fusion of the C3K2 and C2PSA layers of multiple nodes. Subsequently, the detection and classification heads of YOLOv11 are used to achieve wideband fusion detection and modulation recognition of underwater acoustic signals.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A wideband multi-underwater acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11, comprising:

[0009] Preprocess the underwater acoustic signals received by multi-node sampling to obtain the time-frequency distribution diagrams of each signal;

[0010] Use the time-frequency distribution diagrams of each signal as inputs, and complete the detection and recognition of the underwater acoustic signals received by multi-node sampling based on the trained BF-YOLOv11; the BF-YOLOv11 is composed of YOLOv11 and a FusionNet structure, and the FusionNet structure realizes weighted fusion of multi-scale features of multiple nodes based on a bidirectional feature pyramid network.

[0011] Further, the underwater acoustic signals received by multi-node sampling are expressed as:

[0012]

[0013]

[0014] where rij (t) represents the j-th signal received by the i-th node, x j (t) represents the signal transmitted by sound source j, h ij (t) represents the underwater acoustic channel impulse response from sound source j to the i-th node, w i (t) is the ocean ambient noise at the i-th node, N n is the number of nodes, N s is the number of sound sources, δ(t) is the impulse function, K ij is the total number of sound rays from j to the i-th node, A m is the amplitude of the m-th sound ray, τ m is the transmission delay of the m-th sound ray.

[0015] Further, the preprocessing of the underwater acoustic signals sampled and received by multiple nodes includes:

[0016] Denoising the underwater acoustic signals sampled and received by multiple nodes and performing time-frequency processing by short-time Fourier transform.

[0017] Further, the detection and recognition of the underwater acoustic signals sampled and received by multiple nodes based on the trained BF-YOLOv11 includes:

[0018] Extracting features from the time-frequency distribution diagrams of each signal based on the backbone network of YOLOv11, then using the FusionNet structure to achieve feature fusion of the C3K2 and C2PSA layers of multiple nodes, and subsequently using the detection head and classification head of YOLOv11 to achieve the detection and recognition of multiple underwater acoustic signals.

[0019] Further, the underwater acoustic signals sampled and received by multiple nodes include: multi-level frequency shift keying signals, multi-level phase shift keying signals, orthogonal frequency division multiplexing signals, and chirp signals.

[0020] Further, the loss function of the BF-YOLOv11 adopts the CIoU loss function.

[0021] Compared with the prior art, the beneficial effects of the present invention are: [[ID=B]]

[0022] A broadband multi-underwater acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11 provided by the present invention can effectively detect and recognize common underwater acoustic signals in an unknown channel. Its advantages and contributions are:

[0023] 1. A signal processing flow for broadband multi-underwater acoustic signal multi-node fusion detection and recognition is proposed; through preprocessing and BF-YOLOv11, the detection and recognition of underwater acoustic signals are realized, and the proposed method can achieve a high detection and recognition rate under low signal-to-noise ratio conditions.

[0024] 2. A neural network model BF-YOLOv11, which is a fusion of YOLOv11 and the backbone network, is designed. It can extract and fuse the features of signals received by multiple receiving sources, and realize the detection and recognition of underwater acoustic signals.

[0025] 3. In order to strengthen the fusion of multi-node features, a bidirectional feature pyramid network structure is used in the network, and the backbone network post-fusion is achieved through weighted feature fusion and bidirectional cross-scale connection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flowchart of a method for multi-node fusion detection and recognition of broadband multi-underwater acoustic signals based on BF-YOLOv11 provided in an embodiment of the present invention;

[0027] Figure 2 A schematic diagram of the YOLOv11 network structure provided by an embodiment of the present invention;

[0028] Figure 3 A schematic diagram of the fusion process provided by an embodiment of the present invention;

[0029] Figure 4 A schematic diagram of the BiFPN structure provided by an embodiment of the present invention;

[0030] Figure 5 N provided in the embodiment of the present invention s =1 when the test results;

[0031] Figure 6 N provided in the embodiment of the present invention s =2 when the test results. DETAILED DESCRIPTION

[0032] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0033] like Figure 1 As shown, a broadband multi-node fusion detection and recognition method for multi-underwater acoustic signals based on BF-YOLOv11 includes:

[0034] Preprocess the underwater acoustic signals received by multi-node sampling to obtain the time-frequency distribution diagram of each signal;

[0035] The time-frequency distribution diagram of each signal is used as input, and the underwater acoustic signals received by multi-node sampling are detected and recognized based on the trained BF-YOLOv11. The BF-YOLOv11 is composed of YOLOv11 and FusionNet structures. The FusionNet structure realizes weighted fusion of multi-scale features of multiple nodes based on a bidirectional feature pyramid network.

[0036] Furthermore, the underwater acoustic signals received by the multi-node sampling include M-ary Frequency-shift Keying (MFSK), M-ary Phase Shift Keying (MPSK), Orthogonal Frequency Division Multiplexing (OFDM), and Linear Frequency Modulation (LFM) signals.

[0037] The following is a specific introduction.

[0038] 1. System Model

[0039] In the context of multi-node reception, the underwater acoustic signal received by the i-th hydrophone can be expressed as:

[0040]

[0041] In the formula, r ij (t) represents the j-th signal received by the i-th node, x j (t) represents the signal transmitted by the sound source j, h ij (t) represents the underwater acoustic channel impulse response of the signal transmitted by the sound source j arriving at the i-th node, w i (t) is the ocean ambient noise at the i-th node, N n is the number of nodes, N s is the number of sound sources.

[0042] The present invention assumes that there is no relative displacement between the target sound source and the receiving node. Then, the underwater acoustic channel can be modeled as a coherent multipath channel, and its model can be expressed as:

[0043]

[0044] In the formula, δ(t) is the impulse function, K ij is the total number of acoustic rays of the signal transmitted by j arriving at the i-th node, i = 1, 2,..., N n , j = 1, 2,..., N s , A m is the amplitude of the m-th acoustic ray, τ m is the transmission delay of the m-th acoustic ray.

[0045] For underwater acoustic signals, the symbol signal-to-noise ratio is usually used to measure the noise level of the received signal. It is defined as the ratio of the energy of each symbol to the noise power spectral density, denoted as E s / N0, where E srepresents the signal energy of each symbol, and N0 represents the noise power spectral density.

[0046] Therefore, after sampling the model, it can be expressed as:

[0047]

[0048] Among them, r ij (n), x j (n), h ij (n) and w i (n) respectively by r ij (t), x j (t), h ij (t) and w i (t) with sampling rate f s Sampling obtained.

[0049] Passive fusion detection of underwater acoustic signals of multiple nodes is achieved by n The signal r received by the nodes in the same time period ij (n),i=1,2,…,N n ,j=1,2,…,N s The process of joint processing to determine whether the target communication signal exists can be described as a binary hypothesis testing problem:

[0050]

[0051] Among them, H1 is defined as the presence of the target signal, and H0 is defined as the absence of the target signal.

[0052] Define the accuracy (P) as the correct recognition rate among the detected targets i.

[0053]

[0054] TP (True Positive) indicates a correctly detected and recognized signal, while FP (False Positive) indicates an incorrectly detected signal.

[0055] Define recall (R), which is the probability of correctly detecting and identifying all targets.

[0056]

[0057] Among them, FN (False Negative) is the true target signal that has not been detected.

[0058] The F1 score (F1) is defined as the harmonic mean of the precision and recall rates, which comprehensively measures the detection performance of the model.

[0059]

[0060] 2. Key Technologies

[0061] First, through preprocessing, the present invention effectively suppresses the noise in the received multi-node broadband underwater acoustic signal. Second, the preprocessed signal is input into BF-YOLOv11. The features are extracted based on the backbone network of YOLOv11, and then the FusionNet structure is used to achieve feature fusion of the C3K2 and C2PSA layers of multiple nodes. Subsequently, the detection and classification heads of YOLOv11 are used to achieve the detection and recognition of multiple signals.

[0062] 2.1 Preprocessing

[0063] In the signal preprocessing stage, the multi-node sampled received signal is subjected to noise reduction preprocessing and time-frequency processing of short-time Fourier transform (STFT). First, since underwater acoustic signals are usually distributed above 2 kHz, we filter out the noise below 2 kHz and the radiated noise of ships through frequency domain filtering.

[0064] The time-frequency distribution diagram depicts the energy distribution of the signal at different times and frequencies. It has the representation advantages in both the time domain and the frequency domain and contains rich features. Therefore, the present invention uses the time-frequency distribution diagram as the input of the network, extracts and fuses the features of the time-frequency distribution diagram to achieve signal detection. We perform segmented processing on the noise-reduced signal. Taking 1 s as the time window, after normalizing the data r(n) within the window, we perform STFT, and its expression is as follows:

[0065]

[0066] where N is the number of sampling points, is the window function, which is the Hanning window here, to obtain a time-frequency diagram with a fixed output size.

[0067] 2.2 BF-YOLOv11

[0068] The proposed BF-YOLOv11 is a neural network based on the post-fusion of YOLOv11 and the backbone network. The preprocessed received signals of multiple nodes are used as the input. The backbone network of YOLOv11 is used to extract the features of multiple nodes, and then the FusionNet structure based on the BiFPN model is used to achieve feature fusion of the C3K2 and C2PSA layers of multiple nodes. Finally, the detection and classification heads of YOLOv11 are used to achieve wide-band fusion detection and modulation recognition of underwater acoustic signals.

[0069] 2.2.1 YOLOv11

[0070] YOLOv11 has made significant improvements compared to previous versions, which enable YOLOv11 to exhibit more powerful performance and advantages in object detection tasks. YOLOv11 adopts an improved backbone and neck architecture. By introducing new convolutional mechanisms (C3k2 and C2PSA) and depthwise separable convolution (DWConv), the efficiency and quality of feature extraction are significantly improved. The C3k2 mechanism sets the c3k parameter to False in the shallow layer of the network, similar to the C2f structure in YOLOv8, enabling YOLOv11 to extract features more effectively in the shallow network. The C2PSA mechanism is a multi-head attention mechanism embedded inside the C2 mechanism, similar to embedding a PSA (Pyramid Spatial Attention) mechanism in C2, which can better capture spatial context information. At the same time, YOLOv11 T m has been optimized in terms of architecture design and training process to provide faster processing speed. Through fine architecture design, YOLOv11 can provide faster processing speed while maintaining the best balance between accuracy and performance. YOLOv11 also introduces mixed-precision training technology, which speeds up the training process and reduces the video memory occupancy. By reducing the number of parameters, YOLOv11 can run more efficiently and faster on edge devices, thus meeting the requirements of more practical application scenarios. Its structure is as Figure 2 .

[0071] 2.2.2 FusionNet

[0072] The present invention realizes the feature fusion of the C3K2 and C2PSA layers of multiple nodes based on the designed FusionNet structure, as Figure 3 shown. The FusionNet structure designed by the present invention mainly realizes the weighted fusion of multi-scale features of multiple nodes based on the Bidirectional Feature Pyramid Network (BiFPN). BiFPN was first proposed by the Google Brain team in the EfficientDet object detection algorithm. The intention of this network structure is to solve the problem of multi-scale feature fusion in the network. The main structure of BiFPN can be summarized as: effective bidirectional cross-scale connection and weighted feature fusion. When fusing features with different resolutions, the previous feature fusion methods in object detection algorithms were to first adjust them to the same resolution and then sum them up. Most of the previous methods treated features of different scales equally, that is, fusing features with the same weight. However, we found that input features with different resolutions usually contribute unevenly to the output features. To solve this problem, the idea of BiFPN is to add an extra weight to each input and let the network learn the importance of each input feature. The final BiFPN structure integrates cross-scale connection and fast normalization fusion. In EfficientDet, the feature levels utilized by the backbone network are C3 to C7. The complete BiFPN feature fusion method is asFigure 4 as shown

[0073] The BiFPN structure strengthens the fusion of multi-node multi-scale features and enriches the semantic information of features through weighted feature fusion and bidirectional cross-scale connections. The present invention proposes to add an adjusted BiFPN structure on the basis of YOLOv11 to make the object detection algorithm have stronger performance.

[0074] 2.3 Loss function

[0075] The present invention introduces the CIoU (Complete IoU) loss function. When the predicted box and the ground truth box have the same aspect ratio but different width and height values, existing loss functions may have problems, limiting the convergence speed and accuracy. On the basis of DIoU, CIoU further introduces a penalty term for the aspect ratio to optimize the position and shape of the predicted box simultaneously. CIoU = IoU + center point distance loss + aspect ratio loss. Among them, the center point distance loss calculates the distance between the center points of the predicted box and the ground truth box. The aspect ratio loss considers the aspect ratio difference between the predicted box and the ground truth box and is calculated by a function based on the relative aspect ratio. The calculation formula is as follows:

[0076]

[0077] Among them, ρ 2 (b, b gt ) represents the square of the Euclidean distance between the center point b of the predicted box and the center point b gt of the ground truth box. c represents the diagonal length of the minimum closed region of the predicted box and the ground truth box. α is a balance parameter, defined as for dynamically adjusting the weight of the aspect ratio penalty term; v is a measure of aspect ratio similarity, defined as:

[0078]

[0079] Among them, w and h are the width and height of the predicted box, w gt [[ID=3%2]]and h gt are the width and height of the ground truth box. The IoU part measures the overlap degree between the predicted box and the ground truth box. The larger the IoU, the smaller the loss. The center point distance part: by minimizing the distance between the center points of the predicted box and the ground truth box, it ensures that the position of the predicted box is as close as possible to the ground truth box. The aspect ratio part ensures that the shape of the predicted box is consistent with the ground truth box by considering the aspect ratio difference between the predicted box and the ground truth box. By optimizing these three parts simultaneously, the CIoU loss can more comprehensively evaluate the quality of the predicted box and improve the accuracy of the object detection algorithm.

[0080] 3. Experimental results

[0081] Considering the modulation parameters of typical underwater acoustic signals comprehensively, signals are generated based on the model described by the system model. The detailed signal parameters are shown in Table 1, where " / " indicates that this parameter does not need to be set, and "{}" indicates a random selection within the given set. The signal sampling rate is uniformly set to 48 kHz. At the same time, the center frequency of the LFM signal and the carrier frequencies of other signals are both set to 15 - 16 kHz, and the OFDM sub - carriers are BPSK or QPSK.

[0082] Table 1 Signal Parameters

[0083]

[0084] In the training stage, set the number of sound sources N under broadband s ={1, 2, 3}, and the signal - to - noise ratio range of different received signals is [0, 20] dB with a step size of 2 dB. Under different received signal - to - noise ratio conditions and different numbers of sound sources, each hydrophone generates 30 samples.

[0085] In a multi - hydrophone receiving system, due to reasons such as receiving positions, environments, and distances, the transmitted signals reach the receiving end through different channels. In order to be close to the actual application scenario, we select two different water areas as the water area environment based on the Argo ocean database, denoted 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 positions are respectively selected at the thermocline, main thermocline, and sound channel axis. Then, using the Bellhop underwater acoustic channel simulation software, linear time - invariant underwater acoustic channels under different transmission conditions are generated, and their parameters are shown in Table 2.

[0086] Table 2 Underwater Acoustic Channel Parameters

[0087]

[0088] The transfer functions of each channel are obtained as follows:

[0089] H1(z)=0.04 + z -353 +0.508z -570 +0.283z -644 (11)

[0090] H2(z)=0.32 + 0.45z -48 +z -61 +0.9318z -267 (12)

[0091] H3(z)=0.68 + z -184 +0.882z -403 (13)

[0092]

[0093] When the number of sound sources is 1, the test results obtained are as follows Figure 5 shown. It can be seen that the method of the present invention (yolov11-feature) can achieve a gain of 3 - 4 dB and 1 - 2 dB in terms of accuracy, recall rate, and F1 value compared to the single-node (yolov11) and decision-level detection and recognition (yolov11-decision). At the same time, when the signal-to-noise ratio increases to 10 dB, the algorithm of the present invention can reach about 100% in terms of accuracy.

[0094] When the number of sound sources is 2, the test results obtained are as follows Figure 6 shown. It can be seen that regardless of the change in the signal-to-noise ratio of the two sound source signals, the accuracy rate of the method of the present invention is better than that of the single-node and decision fusion methods.

[0095] In summary, the present invention can effectively detect and identify common underwater acoustic signals in an unknown channel.

[0096] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A wideband multi-aquatic acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11, characterized in that, Including: Preprocess the underwater acoustic signals received by multi-node sampling to obtain the time-frequency distribution maps of each signal; Take the time-frequency distribution maps of each signal as input, and complete the detection and recognition of the underwater acoustic signals received by multi-node sampling based on the trained BF-YOLOv11; the BF-YOLOv11 is composed of the YOLOv11 and the FusionNet structure, and the FusionNet structure realizes the weighted fusion of multi-scale features of multi-nodes based on the bidirectional feature pyramid network.

2. The broadband multi-aquatic acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11 according to claim 1, characterized in that, The underwater acoustic signals received by multi-node sampling are expressed as: where r ij (t) represents the j-th signal received by the i-th node, x j (t) represents the signal transmitted by sound source j, h ij (t) represents the underwater acoustic channel impulse response from sound source j to the i-th node, w i (t) is the ocean ambient noise at the i-th node, N n is the number of nodes, N s is the number of sound sources, δ(t) is the impulse function, K ij is the total number of acoustic rays from j to the i-th node, A m is the amplitude of the m-th acoustic ray, τ m is the propagation delay of the m-th acoustic ray.

3. The broadband multi-aquatic acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11 according to claim 1, characterized in that The preprocessing of the underwater acoustic signals received by multi-node sampling includes: Perform noise reduction processing on the underwater acoustic signals received by multi-node sampling and perform time-frequency processing of short-time Fourier transform.

4. The broadband multi-aquatic acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11 according to claim 1, characterized in that, The completion of the detection and recognition of the underwater acoustic signals received by multi-node sampling based on the trained BF-YOLOv11 includes: Extract features from the time-frequency distribution maps of each signal based on the backbone network of YOLOv11, then use the FusionNet structure to realize the feature fusion of the C3K2 and C2PSA layers of multi-nodes, and then use the detection head and classification head of YOLOv11 to realize the detection and recognition of multiple underwater acoustic signals.

5. The wideband multi-aquatic acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11 according to claim 1, characterized in that The underwater acoustic signals received by multi-node sampling include: multi-level frequency shift keying signals, multi-level phase shift keying signals, orthogonal frequency division multiplexing signals, and linear frequency modulation signals.

6. The wideband multi-aquatic acoustic signal multi-node fusion detection and recognition method based on BF-YOLOv11 according to claim 1, characterized in that, The loss function of the BF-YOLOv11 adopts the CIoU loss function.

Images