Water-impact sound signal detection method based on deep learning image segmentation model
By constructing a deep learning-based image segmentation model and utilizing U-Net mask branches and skip connection structures, accurate separation of water-pounding sound signals was achieved, solving the problem of low detection accuracy in existing technologies and improving detection performance in complex underwater acoustic environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN ENG UNIV
- Filing Date
- 2026-03-20
- Publication Date
- 2026-06-23
Smart Images
Figure CN122265644A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of underwater acoustics, and particularly relates to a method for detecting water impact sound signals based on a deep learning image segmentation model. Background Technology
[0002] Water-entry sound (WES) is a special type of underwater acoustic signal generated by the intense interaction between an object and the water surface during the instant air enters the water and within a short period thereafter. It is a transient sound wave generated the moment an object impacts the water surface and enters the water. Its generation mechanism is complex, typically including an impact phase, a cavity formation phase, and a bubble pulsation phase. Accurate detection of this signal is crucial for underwater target positioning systems. It not only provides the initial time marker needed for target positioning but, combined with hydrophone arrays, can also achieve target orientation estimation. Studies show that water-entry sound signals are characterized by short duration, non-stationarity, and strong impulses in the time domain, and exhibit a broadband distribution and lack obvious line spectrum features in the frequency domain. However, the core challenge in engineering applications lies in signal extraction in complex underwater acoustic environments. In real-world scenarios, target signals are often overwhelmed by marine environmental noise and the self-noise of the observation platform (such as propeller, mechanical vibration, and hydrodynamic noise). These strong interferences result in extremely low signal-to-noise ratios (SNR) and signal-to-interference ratios (SIR), making it very difficult to obtain accurate water-entry sound signals using conventional methods, and easily leading to missed detections and false alarms.
[0003] Existing underwater acoustic transient signal detection techniques mainly encompass traditional methods based on statistical signal processing and early deep learning methods based on time-domain features. In terms of traditional methods, statistical detection methods such as the Power-Law detector, the Page Test algorithm, and higher-order spectral analysis are widely used to capture signal abrupt changes and edge features. For example, the cumulative sum (CUSUM) principle is used to detect falling edges of signals, or higher-order statistics are used to suppress Gaussian background noise. Simultaneously, matrix factorization algorithms such as Robust Principal Component Analysis (RPCA) and Nonnegative Matrix Factorization (NMF) have been introduced, attempting to separate sparse transient signals from low-rank background noise through mathematical optimization. However, these traditional methods often rely on idealized assumptions about noise distribution (such as Gaussian white noise). When faced with complex non-stationary noise and variable interference in the marine environment, their feature generalization ability is limited, and they struggle to characterize the fine structure of the signal.
[0004] In recent years, although deep learning technology has begun to be applied to the field of underwater acoustics to improve detection performance, current mainstream solutions still focus on the direct processing of time-domain waveforms. One paper proposes a weak transient signal detection method based on time-domain deep feature learning, utilizing a one-dimensional deep residual network (ResNet) to directly perform end-to-end feature learning on one-dimensional time-series signals, achieving signal existence discrimination and time-domain location. Similarly, another paper discloses a time-series transient signal detection device based on deep learning, which processes time-series data by constructing a hybrid network model containing convolutional and recurrent units. While these methods avoid the tediousness of manual feature design, they to some extent ignore the rich two-dimensional texture and morphological features of underwater acoustic signals in the time-frequency domain. Especially for signals with unique time-frequency structures (such as impact components and bubble pulsation components), such as pounding sounds, it is difficult to capture their global semantic information on the time-frequency map using only one-dimensional time-domain features, making it difficult to achieve pixel-level fine segmentation and recognition. Summary of the Invention
[0005] The purpose of this invention is to address the problem of low detection accuracy in existing methods for detecting water-striking sound signals. A method for detecting water-striking sound signals based on a deep learning image segmentation model is provided, including:
[0006] I. Construct an underwater acoustic signal image segmentation and detection model; obtain a trained underwater acoustic signal image segmentation and detection model;
[0007] 2. Acquire the underwater acoustic data to be detected; preprocess the underwater acoustic data to be detected to obtain the four-channel video feature matrix of the underwater acoustic data to be detected;
[0008] 3. Input the four-channel video feature matrix of the underwater acoustic data to be detected into the trained underwater acoustic signal image segmentation and detection model for detection and obtain the detection results.
[0009] The first step involves constructing an underwater acoustic signal image segmentation and detection model; obtaining a trained underwater acoustic signal image segmentation and detection model; the specific process is as follows:
[0010] Step 1: Construct a training set of time-frequency feature samples of underwater acoustic signals;
[0011] Step 2: Construct an underwater acoustic signal image segmentation and detection model;
[0012] Step 3: Train the underwater acoustic signal image segmentation and detection model based on the training set of time-frequency feature samples of the underwater acoustic signal.
[0013] Obtain a trained underwater acoustic signal image segmentation and detection model;
[0014] The second step involves preprocessing the underwater acoustic data to be detected to obtain a four-channel video feature matrix of the underwater acoustic data. The specific process is as follows:
[0015] A1: Perform short-time Fourier transform on the one-dimensional time series signal to be detected to obtain the two-dimensional time-frequency complex matrix to be detected;
[0016] A2: Extract the real part, imaginary part, amplitude, and phase of all complex numbers in the two-dimensional time-frequency complex matrix to be detected;
[0017] A3: Construct a four-channel video feature matrix of the underwater acoustic data to be detected based on the real part, imaginary part, amplitude, and phase of all complex numbers in the two-dimensional time-frequency complex matrix to be detected;
[0018] The beneficial effects of this invention are as follows:
[0019] This method reconstructs signal detection into a semantic segmentation task on time-frequency images. The deep learning-based semantic segmentation ad hoc search approach adopts a data-driven model, relying not on manually designed fixed features, but automatically learning and extracting multi-level, multi-scale features from the data through deep neural networks. A multi-channel feature matrix containing the real, imaginary, amplitude, and phase parts of the signal is constructed at the input end to fully utilize the physical information of the signal. This ensures that both the impact and pulsation components of the slamming sound can be accurately separated and preserved even under complex noise interference, effectively solving the problem of slamming sound signals being masked and difficult to detect in actual sea conditions. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the overall network structure of the present invention;
[0021] Figure 2 This is a schematic diagram of the encoder and decoder structure of the present invention;
[0022] Figure 3 This is a schematic diagram of the jump connection structure of the present invention;
[0023] Figure 4 This is a schematic diagram of the U-Net mask branch structure of the present invention;
[0024] Figure 5 A schematic diagram of the received signal in the time domain when the signal-to-interference ratio is -5dB and the signal-to-noise ratio is 20dB.
[0025] Figure 6 A time-frequency domain diagram of the received signal when the signal-to-interference ratio is -5dB and the signal-to-noise ratio is 20dB.
[0026] Figure 7 A schematic diagram of the results of this invention with a signal-to-interference ratio of -5dB and a signal-to-noise ratio of 20dB.
[0027] Figure 8 A schematic diagram of the received signal in the time domain when the signal-to-interference ratio is 20dB and the signal-to-noise ratio is -5dB.
[0028] Figure 9A time-frequency domain diagram of the received signal when the signal-to-interference ratio is 20dB and the signal-to-noise ratio is -5dB.
[0029] Figure 10 A schematic diagram showing the results of the method proposed in this invention when the signal-to-interference ratio is 20dB and the signal-to-noise ratio is -5dB.
[0030] Figure 11 This is a schematic diagram showing the IoU performance of image segmentation models with different network architectures for detecting WES signals as a function of SIR.
[0031] Figure 12 This is a schematic diagram showing the IoU performance of image segmentation models with different network architectures for WES signal detection as a function of SNR.
[0032] Figure 13 A time-domain schematic diagram of the image segmentation model of the present invention applied to the measured hydrophone signal;
[0033] Figure 14 A schematic diagram of the time-frequency domain of the image segmentation model of the present invention for the measured hydrophone signal;
[0034] Figure 15 A schematic diagram of the results of the image segmentation model of the present invention on the measured hydrophone signal. Detailed Implementation
[0035] Specific implementation method one: Combining Figures 1-15 This invention is described;
[0036] I. Construct an underwater acoustic signal image segmentation and detection model; obtain a trained underwater acoustic signal image segmentation and detection model;
[0037] 2. Acquire the underwater acoustic data to be detected; preprocess the underwater acoustic data to be detected to obtain the four-channel video feature matrix of the underwater acoustic data to be detected;
[0038] 3. Input the four-channel video feature matrix of the underwater acoustic data to be detected into the trained underwater acoustic signal image segmentation and detection model for detection and obtain the detection results.
[0039] Specific Implementation Method Two: The difference between this implementation method and Specific Implementation Method One is that:
[0040] The first step involves constructing an underwater acoustic signal image segmentation and detection model; obtaining a trained underwater acoustic signal image segmentation and detection model; the specific process is as follows:
[0041] Step 1: Construct a training set of time-frequency feature samples of underwater acoustic signals;
[0042] Step 2: Construct an underwater acoustic signal image segmentation and detection model;
[0043] Step 3: Train the underwater acoustic signal image segmentation and detection model based on the training set of time-frequency feature samples of the underwater acoustic signal.
[0044] Obtain a trained underwater acoustic signal image segmentation and detection model.
[0045] The other steps and parameters are the same as in Specific Implementation Method 1.
[0046] Specific Implementation Method Three: The difference between this implementation method and Specific Implementation Methods One and Two is that:
[0047] In step one, a training set of time-frequency feature samples of the underwater acoustic signal is constructed; the specific process is as follows:
[0048] Step 11: Acquire the one-dimensional time-series signal received by the hydrophone;
[0049] In a one-dimensional time series signal, the time series signal at time t is represented as: ;
[0050] Steps 1 and 2: Perform Short Time Fourier Transform (STFT) processing on the one-dimensional time-series signal received by the hydrophone to obtain a two-dimensional time-frequency complex matrix; wherein, the complex representation at time t and frequency f in the two-dimensional time-frequency complex matrix is as follows: ;
[0051] Step 13: Extract the real part of all complex numbers in the two-dimensional time-frequency complex matrix. virtual part Amplitude and phase Here, R refers to the matrix composed of the real parts of all complex numbers. This real part R is also a matrix, and the imaginary part, magnitude, and phase are also of the same type. The dimension, rows, columns, and positions remain completely unchanged, only the elements change from complex numbers to single real numbers.
[0052] Based on the two-dimensional time-frequency complex matrix real part of all complex numbers virtual part Amplitude and phase Constructing a four-channel feature matrix Four-channel feature matrix For any specific time point t and frequency point f, the elements are: A vector containing four values yields... The dimension is t×f×4. Expressed as a formula:
[0053] ;
[0054] In the formula, express The real part of the complex number corresponding to the time point t and the frequency point f;
[0055] express The imaginary part of the complex number corresponding to the time point t and the frequency point f;
[0056] express The amplitude values of the complex number corresponding to the time point t and the frequency point f;
[0057] express The phase values of the complex number corresponding to time point t and frequency point f;
[0058] Step 14: For the four-channel feature matrix Normalization is performed to obtain the normalized four-channel feature matrix. The formula is as follows:
[0059]
[0060] In the formula, Indicates the maximum value; This represents the four-channel feature matrix processed by the maximum value function. A specific element within;
[0061] Step 15: Based on the physical characteristics of the sound signal entering the water, the signal is divided into three categories: impact sound, pulsating sound, and interference noise.
[0062] Construction and normalization of the four-channel feature matrix Corresponding label mask ,
[0063] Based on the normalized four-channel feature matrix and label mask Construct a training set of time-frequency feature samples for underwater acoustic signals;
[0064] The specific process is as follows: If the time frequency point The signal belongs to the first The class is in the tag mask. time and frequency points The Each channel is marked as 1 if it corresponds to a specific position, and 0 otherwise. The sound indicates an impact. The time indicates a pulsating sound; Time represents interference noise; T represents the number of time rows, F represents the number of frequency columns, and C represents the number of channels; other steps and parameters are the same as in one of the specific implementation methods one or two.
[0065] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that:
[0066] The underwater acoustic signal image segmentation and detection model in step two includes: a main encoder, a U-Net mask branch, a main decoder, and an output layer;
[0067] The main encoder includes: a first coding unit, a second coding unit, a third coding unit, a fourth coding unit, a fifth coding unit, a sixth coding unit, a seventh coding unit, an eighth coding unit, and a ninth coding unit;
[0068] The first, second, third, fourth, fifth, sixth, seventh, eighth, and ninth coding units each include: a dilated convolutional layer and a ReLU activation function layer;
[0069] The U-Net mask branch includes: a first sub-coding unit, a second sub-coding unit, a third sub-coding unit, a fourth sub-coding unit, a fifth sub-coding unit, a sixth sub-coding unit, a seventh sub-coding unit, a first max pooling layer, a second max pooling layer, a third max pooling layer, a fourth max pooling layer, a fifth max pooling layer, a sixth max pooling layer, a seventh max pooling layer, a first upsampling layer, a second upsampling layer, a third upsampling layer, a fourth upsampling layer, a fifth upsampling layer, a sixth upsampling layer, a seventh upsampling layer, a first sub-decoding unit, a second sub-decoding unit, a third sub-decoding unit, a fourth sub-decoding unit, a fifth sub-decoding unit, a sixth sub-decoding unit, and a seventh sub-decoding unit;
[0070] The first sub-coding unit, the second sub-coding unit, the third sub-coding unit, the fourth sub-coding unit, the fifth sub-coding unit, the sixth sub-coding unit, and the seventh sub-coding unit each include: a dilated convolutional layer and a ReLU activation function layer;
[0071] The first sub-decoding unit, the second sub-decoding unit, the third sub-decoding unit, the fourth sub-decoding unit, the fifth sub-decoding unit, the sixth sub-decoding unit, and the seventh sub-decoding unit each include: a dilated convolutional layer and a ReLU activation function layer;
[0072] The main decoder includes: a first decoding unit, a second decoding unit, a third decoding unit, a fourth decoding unit, a fifth decoding unit, a sixth decoding unit, a seventh decoding unit, an eighth decoding unit, and a ninth decoding unit;
[0073] The first decoding unit, the second decoding unit, the third decoding unit, the fourth decoding unit, the fifth decoding unit, the sixth decoding unit, the seventh decoding unit, the eighth decoding unit, and the ninth decoding unit each include: a dilated convolutional layer and a ReLU activation function layer;
[0074] The output layer comprises, in sequence: a Tanh activation function layer, a Flatten feature tensor flattening layer, and a fully connected layer;
[0075] The other steps and parameters are the same as those in one of the specific implementation methods one to three.
[0076] Specific Implementation Method Five: The difference between this implementation method and Specific Implementation Methods One to Four is that:
[0077] In step three, the underwater acoustic signal image segmentation and detection model is trained based on the training set of time-frequency feature samples of the underwater acoustic signal to obtain the trained underwater acoustic signal image segmentation and detection model. The specific process is as follows:
[0078] S1: The four-channel feature matrix obtained by normalizing the time-frequency feature sample training set of the underwater acoustic signal. The input to the first coding unit is processed to obtain the intermediate feature map B1;
[0079] The intermediate feature map B1 is input into the second decoding unit for processing to obtain the intermediate feature map B2.
[0080] The intermediate feature map B2 is input into the third decoding unit for processing to obtain the intermediate feature map B3.
[0081] The intermediate feature map B3 is input into the fourth decoding unit for processing to obtain the intermediate feature map B4.
[0082] The intermediate feature map B4 is input into the fifth decoding unit for processing to obtain the intermediate feature map B5.
[0083] After the intermediate feature map B3 is skipped and connected to the intermediate feature map B5, it is input into the sixth decoding unit for processing to obtain the intermediate feature map B6.
[0084] After the intermediate feature map B2 is skipped and connected to the intermediate feature map B6, it is input into the seventh decoding unit for processing to obtain the intermediate feature map B7.
[0085] After the intermediate feature map B1 is skipped and connected to the intermediate feature map B7, it is input into the eighth decoding unit for processing to obtain the intermediate feature map B8.
[0086] The intermediate feature map B8 is input into the ninth decoding unit for processing to obtain the intermediate feature map B9.
[0087] S2: Input intermediate feature maps B3, B4, B5, B6, B7, and B8 into the U-Net mask branch for processing to obtain the spatial attention mask;
[0088] S3: Weight the intermediate feature map B9 according to the spatial attention mask to obtain the intermediate feature map D1;
[0089] S4: Input the intermediate feature map D1 into the first decoding unit for processing to obtain the intermediate feature map D2;
[0090] After the intermediate feature map B8 is skipped and connected to the intermediate feature map D2, it is input into the second decoding unit for processing to obtain the intermediate feature map D3.
[0091] After the intermediate feature map B7 is skipped and connected to the intermediate feature map D3, it is input into the third decoding unit for processing to obtain the intermediate feature map D4.
[0092] After the intermediate feature map B6 is skipped and connected to the intermediate feature map D4, it is input into the fourth decoding unit for processing to obtain the intermediate feature map D5.
[0093] After the intermediate feature map B5 is skipped and connected to the intermediate feature map D5, it is input into the fifth decoding unit for processing to obtain the intermediate feature map D6.
[0094] After skipping and connecting intermediate feature map B4 and intermediate feature map D4 to intermediate feature map D6, the intermediate feature map D7 is input into the sixth decoding unit for processing.
[0095] After skipping and connecting intermediate feature map B3 and intermediate feature map D3 to intermediate feature map D7, the intermediate feature map D8 is input into the seventh decoding unit for processing.
[0096] After skipping and connecting intermediate feature map B2 and intermediate feature map D2 to intermediate feature map D8, the intermediate feature map D9 is input into the eighth decoding unit for processing.
[0097] The intermediate feature map D9 is input into the ninth decoding unit for processing to obtain the intermediate feature map E1.
[0098] S5: Input the intermediate feature map E1 into the output layer for processing to obtain the output result. ;
[0099] S6: Based on the output results Label mask in the training set of underwater acoustic signal time-frequency feature samples The calculation of the loss function;
[0100] S7: Repeat S1 to S6, using the adaptive moment estimator (Adam) optimizer to iteratively update the parameters of the underwater acoustic signal image segmentation and detection model according to the loss function. Stop training when the loss function converges to obtain the trained underwater acoustic signal image segmentation and detection model; other steps and parameters are the same as in one of the specific implementation methods one to four.
[0101] Specific Implementation Method Six: The difference between this implementation method and Specific Implementation Methods One through Five is that:
[0102] In step S2, intermediate feature maps B2, B3, B4, B5, B6, B7, and B8 are input into the U-Net mask branch for processing to obtain the spatial attention mask; the specific process is as follows:
[0103] S2.1: Input the intermediate feature map B2 into the first sub-encoding unit for processing to obtain the intermediate feature map C1; input the intermediate feature map C1 into the first max pooling layer for processing to obtain the intermediate feature map C2;
[0104] S2.2: After the intermediate feature map B3 is skipped and connected to the intermediate feature map C2, it is input into the second sub-coding unit for processing to obtain the intermediate feature map C3;
[0105] The intermediate feature map C3 is input into the second max pooling layer for processing to obtain the intermediate feature map C4.
[0106] S2.3: After skipping and connecting the intermediate feature map B4 to the intermediate feature map C4, it is input into the third sub-encoding unit for processing to obtain the intermediate feature map C5;
[0107] The intermediate feature map C5 is input into the third max pooling layer for processing to obtain the intermediate feature map C6.
[0108] S2.4: After skipping and connecting the intermediate feature map B5 to the intermediate feature map C6, it is input into the fourth sub-coding unit for processing to obtain the intermediate feature map C7;
[0109] The intermediate feature map C7 is input into the fourth max pooling layer for processing to obtain the intermediate feature map C8.
[0110] S2.5: After skipping and connecting the intermediate feature map B6 to the intermediate feature map C8, input it into the fifth sub-encoding unit for processing to obtain the intermediate feature map C9;
[0111] The intermediate feature map C9 is input into the fifth max pooling layer for processing to obtain the intermediate feature map C10.
[0112] S2.6: After skipping and connecting the intermediate feature map B7 to the intermediate feature map C10, it is input into the sixth sub-encoding unit for processing to obtain the intermediate feature map C11;
[0113] The intermediate feature map C11 is input into the sixth max pooling layer for processing to obtain the intermediate feature map C12.
[0114] S2.7: After skipping and connecting the intermediate feature map B8 to the intermediate feature map C12, input it into the seventh sub-coding unit for processing to obtain the intermediate feature map C13;
[0115] The intermediate feature map C13 is input into the seventh max pooling layer for processing to obtain the intermediate feature map C14.
[0116] S2.8: Input the intermediate feature map C14 into the first upsampling layer for processing to obtain the intermediate feature map C15;
[0117] The intermediate feature map C15 is input into the first sub-decoding unit for processing to obtain the intermediate feature map C16.
[0118] S2.9: The intermediate feature map C16 is input into the second upsampling layer for processing to obtain the intermediate feature map C17;
[0119] After the intermediate feature map C11 is skipped and connected to the intermediate feature map C17, it is input into the second sub-decoding unit for processing to obtain the intermediate feature map C18.
[0120] S2.10: The intermediate feature map C18 is input into the third upsampling layer for processing to obtain the intermediate feature map C19;
[0121] After the intermediate feature map C9 is skipped and connected to the intermediate feature map C19, it is input into the third sub-decoding unit for processing to obtain the intermediate feature map C20.
[0122] S2.11: The intermediate feature map C20 is input into the fourth upsampling layer for processing to obtain the intermediate feature map C21;
[0123] After the intermediate feature map C7 is skipped and connected to the intermediate feature map C21, it is input into the fourth sub-decoding unit for processing to obtain the intermediate feature map C22.
[0124] S2.12: The intermediate feature map C22 is input into the fifth upsampling layer for processing to obtain the intermediate feature map C23;
[0125] After the intermediate feature map C5 is skipped and connected to the intermediate feature map C23, it is input into the fifth sub-decoding unit for processing to obtain the intermediate feature map C24.
[0126] S2.13: The intermediate feature map C24 is input into the sixth upsampling layer for processing to obtain the intermediate feature map C24;
[0127] After connecting intermediate feature map C3 to intermediate feature map C24, it is input into the sixth sub-decoding unit for processing to obtain intermediate feature map C25.
[0128] S2.14: The intermediate feature map C25 is input into the seventh upsampling layer for processing to obtain the intermediate feature map C26;
[0129] After the intermediate feature map C1 is skipped and connected to the intermediate feature map C26, it is input into the seventh sub-decoding unit for processing to obtain the spatial attention mask.
[0130] This branch, as an embedded sub-network, adopts the classic sub-encoder-decoder architecture.
[0131] The input main encoder features are first processed by convolution and ReLU, and then enter the downsampling path. The spatial resolution is gradually reduced and the number of channels is increased through max pooling layers to extract deep semantic information.
[0132] Then, the upsampling path is entered, and the spatial resolution of the feature map is gradually restored using upsampling convolution.
[0133] In this process, this branch uses internal skip connections to concatenate shallow features from the downsampling path to the corresponding layers of the upsampling path, preserving rich spatial edge details. Finally, this branch outputs a high-resolution spatial attention mask with the same size as the input.
[0134] The other steps and parameters are the same as those in any of the specific implementation methods one to five.
[0135] Specific Implementation Method Seven: The difference between this implementation method and Specific Implementation Methods One through Six is that:
[0136] The specific process in S1, S2, and S3 of skipping and connecting one intermediate feature map to another intermediate feature map.
[0137] Expressed as a formula:
[0138]
[0139] In the formula, Feature maps representing skip connections, The feature map representing the skip connections. This represents the feature map after the skip connections. Indicates a skip connection. The expression represents element-wise multiplication, sigmoid() represents the sigmoid activation function, maxpool() represents max pooling, conv represents convolution, and tanh() represents the tanh activation function.
[0140] The other steps and parameters are the same as those in any of the specific implementation methods one to six.
[0141] Specific Implementation Method Eight: The difference between this implementation method and Specific Implementation Methods One through Seven is that:
[0142] The loss function used in S6 is the cross-entropy loss function.
[0143] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.
[0144] Specific Implementation Method Nine: The difference between this implementation method and Specific Implementation Methods One through Eight is that:
[0145] The second type of underwater acoustic data to be detected is a one-dimensional time-series signal received by a hydrophone.
[0146] Specific Implementation Method Ten: The difference between this implementation method and Specific Implementation Methods One through Nine is that:
[0147] The second step involves preprocessing the underwater acoustic data to be detected to obtain a four-channel video feature matrix of the underwater acoustic data. The specific process is as follows:
[0148] A1: Perform short-time Fourier transform on the one-dimensional time series signal to be detected to obtain the two-dimensional time-frequency complex matrix to be detected;
[0149] A2: Extract the real part, imaginary part, amplitude, and phase of all complex numbers in the two-dimensional time-frequency complex matrix to be detected;
[0150] A3: Construct a four-channel video feature matrix of the underwater acoustic data to be detected based on the real part, imaginary part, amplitude, and phase of all complex numbers in the two-dimensional time-frequency complex matrix to be detected;
[0151] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.
[0152] The core technology of this invention lies in transforming transient signal detection into semantic segmentation of time-frequency images. According to the physical characteristics of water-sound signals, they manifest as a two-dimensional spectrum with a specific structure in the time-frequency domain. Traditional detection methods rely on one-dimensional time-domain analysis or simple thresholding, which is insufficient for handling strongly non-stationary signals. This method reconstructs signal detection into a semantic segmentation task on a time-frequency image. The deep learning-based semantic segmentation ad hoc search approach adopts a data-driven model, not relying on manually designed fixed features, but automatically learning and extracting multi-level, multi-scale features from the data through deep neural networks. A multi-channel feature matrix containing the real, imaginary, amplitude, and phase of the signal is constructed at the input end to fully utilize the physical information of the signal. During processing, a U-Net masking mechanism is introduced, essentially "illuminating" the effective signal region in a complex acoustic background. Finally, a high-confidence signal distribution is obtained through reconstruction by the decoder. Traditional methods are prone to false alarms and missed detections under low signal-to-interference ratio (SIR) and low signal-to-noise ratio (SNR) conditions. However, the image segmentation mechanism of this method can maintain an intersection-to-union ratio (IoU) of 0.9 and 0.82 even when the SIR and SNR are as low as -11dB, which significantly improves the detection performance.
[0153] The hybrid network architecture mechanism included in this invention mainly combines an encoder-decoder structure with a U-Net network structure. The model uses the main encoder to extract features and generates a time-frequency mask through U-Net branches as sub-networks. This mechanism utilizes both the encoder-decoder structure's ability to abstract global features and the U-Net's ability to recover spatial details through skip connections, ensuring that the impact and pulsation components of the water-splashing sound can be accurately separated and preserved even under complex noise interference.
[0154] The dilated convolution feature extraction mechanism included in this invention mainly involves applying stacked dilated convolutional layers in both the encoder and decoder. Leveraging the property of dilated convolution to significantly expand the receptive field without changing the feature map size, this mechanism ensures that the network can capture complex global features in the time-frequency domain and dynamic changes in background noise. In this way, the model can effectively extract a wide range of contextual information while maintaining computational efficiency, providing sufficient global semantic support for subsequent signal separation.
[0155] The multi-scale feature fusion mechanism included in this invention mainly addresses the problems of low-frequency feature bias, insufficient high-frequency information extraction, and gradient vanishing that often occur in deep convolutional networks. This invention introduces a skip connection structure between corresponding layers of the encoder and decoder. This mechanism establishes a direct information path across layers, enabling dynamic selection and fusion of features between different layers. This not only ensures efficient multi-scale feature integration but also effectively alleviates the gradient vanishing phenomenon, enhancing the model's ability to simultaneously represent local details (such as transient signal edges) and global structure.
[0156] The time-frequency attention mask guidance mechanism included in this invention mainly involves the following: the output generated by the U-Net branch is used as an attention mask and fused with the main decoder path for feature extraction. The pixel value in the mask represents the probability that the time-frequency point belongs to the target signal. This mechanism provides explicit guidance for separating weak signals from complex backgrounds, enabling the main decoder to effectively suppress strong interference such as platform self-noise during signal reconstruction, and significantly improving the robustness of the algorithm in low signal-to-noise ratio environments.
[0157] The robust anti-interference mechanism of this invention mainly involves training with a sample set containing diverse time-frequency features and using the Intersection over Union (IoU) as an evaluation metric. Experiments show that, compared to single-network models such as DeepLab and FCN, this method exhibits the smallest decrease in IoU in interference-dominated (low SIR) and noise-dominated (low SNR) environments, demonstrating stronger generalization ability and stability, and effectively solving the problem of difficult detection of sea-pounding acoustic signals due to masking in actual sea conditions.
[0158] The above description is merely of preferred embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention, and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A method for detecting water-pounding acoustic signals based on a deep learning image segmentation model, characterized in that, include: I. Construct an underwater acoustic signal image segmentation and detection model; obtain a trained underwater acoustic signal image segmentation and detection model; II. Acquiring underwater acoustic data to be tested; Preprocessing is performed on the underwater acoustic data to be detected to obtain a four-channel video feature matrix of the underwater acoustic data to be detected.
3. Input the four-channel video feature matrix of the underwater acoustic data to be detected into the trained underwater acoustic signal image segmentation and detection model for detection and obtain the detection results.
2. The method for detecting water-pounding sound signals based on a deep learning image segmentation model according to claim 1, characterized in that, The first step involves constructing an underwater acoustic signal image segmentation and detection model; obtaining a trained underwater acoustic signal image segmentation and detection model; the specific process is as follows: Step 1: Construct a training set of time-frequency feature samples of underwater acoustic signals; Step 2: Construct an underwater acoustic signal image segmentation and detection model; Step 3: Train the underwater acoustic signal image segmentation and detection model based on the training set of time-frequency feature samples of the underwater acoustic signal; Obtain a trained underwater acoustic signal image segmentation and detection model.
3. The method for detecting water-pounding sound signals based on a deep learning image segmentation model according to claim 2, characterized in that, In step one, a training set of time-frequency feature samples of the underwater acoustic signal is constructed; the specific process is as follows: Step 11: Acquire the one-dimensional time-series signal received by the hydrophone; In a one-dimensional time series signal, the time series signal at time t is represented as: ; Steps 1 and 2: Perform Short Time Fourier Transform (STFT) processing on the one-dimensional time-series signal received by the hydrophone to obtain a two-dimensional time-frequency complex matrix; wherein, the complex representation at time t and frequency f in the two-dimensional time-frequency complex matrix is as follows: ; Step 13: Extract the real part of all complex numbers in the two-dimensional time-frequency complex matrix. virtual part Amplitude and phase ; Based on the two-dimensional time-frequency complex matrix real part of all complex numbers virtual part Amplitude and phase Constructing a four-channel feature matrix ; Step 14: For the four-channel feature matrix Normalization is performed to obtain the normalized four-channel feature matrix. ; Step 15: Based on the physical characteristics of the water entry sound signal, the signal is divided into three categories: impact sound, pulsating sound, and interference noise; Construction and normalization of the four-channel feature matrix Corresponding label mask , Based on the normalized four-channel feature matrix and label mask Construct a training set of time-frequency feature samples of underwater acoustic signals.
4. The method for detecting water-pounding sound signals based on a deep learning image segmentation model according to claim 3, characterized in that, The underwater acoustic signal image segmentation and detection model in step two includes: a main encoder, a U-Net mask branch, a main decoder, and an output layer; The main encoder includes: a first coding unit, a second coding unit, a third coding unit, a fourth coding unit, a fifth coding unit, a sixth coding unit, a seventh coding unit, an eighth coding unit, and a ninth coding unit; The first, second, third, fourth, fifth, sixth, seventh, eighth, and ninth coding units each include: a dilated convolutional layer and a ReLU activation function layer; The U-Net mask branch includes: a first sub-coding unit, a second sub-coding unit, a third sub-coding unit, a fourth sub-coding unit, a fifth sub-coding unit, a sixth sub-coding unit, a seventh sub-coding unit, a first max pooling layer, a second max pooling layer, a third max pooling layer, a fourth max pooling layer, a fifth max pooling layer, a sixth max pooling layer, a seventh max pooling layer, a first upsampling layer, a second upsampling layer, a third upsampling layer, a fourth upsampling layer, a fifth upsampling layer, a sixth upsampling layer, a seventh upsampling layer, a first sub-decoding unit, a second sub-decoding unit, a third sub-decoding unit, a fourth sub-decoding unit, a fifth sub-decoding unit, a sixth sub-decoding unit, and a seventh sub-decoding unit; The first sub-coding unit, the second sub-coding unit, the third sub-coding unit, the fourth sub-coding unit, the fifth sub-coding unit, the sixth sub-coding unit, and the seventh sub-coding unit each include: a dilated convolutional layer and a ReLU activation function layer; The first sub-decoding unit, the second sub-decoding unit, the third sub-decoding unit, the fourth sub-decoding unit, the fifth sub-decoding unit, the sixth sub-decoding unit, and the seventh sub-decoding unit each include: a dilated convolutional layer and a ReLU activation function layer; The main decoder includes: a first decoding unit, a second decoding unit, a third decoding unit, a fourth decoding unit, a fifth decoding unit, a sixth decoding unit, a seventh decoding unit, an eighth decoding unit, and a ninth decoding unit; The first decoding unit, the second decoding unit, the third decoding unit, the fourth decoding unit, the fifth decoding unit, the sixth decoding unit, the seventh decoding unit, the eighth decoding unit, and the ninth decoding unit each include: a dilated convolutional layer and a ReLU activation function layer; The output layer comprises, in sequence: a Tanh activation function layer, a Flatten feature tensor flattening layer, and a fully connected layer.
5. The method for detecting water-pounding acoustic signals based on a deep learning image segmentation model according to claim 4, characterized in that, In step three, the underwater acoustic signal image segmentation and detection model is trained based on the training set of time-frequency feature samples of the underwater acoustic signal to obtain the trained underwater acoustic signal image segmentation and detection model. The specific process is as follows: S1: The four-channel feature matrix obtained by normalizing the time-frequency feature sample training set of the underwater acoustic signal. The input to the first coding unit is processed to obtain the intermediate feature map B1; The intermediate feature map B1 is input into the second decoding unit for processing to obtain the intermediate feature map B2. The intermediate feature map B2 is input into the third decoding unit for processing to obtain the intermediate feature map B3. The intermediate feature map B3 is input into the fourth decoding unit for processing to obtain the intermediate feature map B4. The intermediate feature map B4 is input into the fifth decoding unit for processing to obtain the intermediate feature map B5. After the intermediate feature map B3 is skipped and connected to the intermediate feature map B5, it is input into the sixth decoding unit for processing to obtain the intermediate feature map B6. After the intermediate feature map B2 is skipped and connected to the intermediate feature map B6, it is input into the seventh decoding unit for processing to obtain the intermediate feature map B7. After the intermediate feature map B1 is skipped and connected to the intermediate feature map B7, it is input into the eighth decoding unit for processing to obtain the intermediate feature map B8. The intermediate feature map B8 is input into the ninth decoding unit for processing to obtain the intermediate feature map B9. S2: Input intermediate feature maps B3, B4, B5, B6, B7, and B8 into the U-Net mask branch for processing to obtain the spatial attention mask; S3: Weight the intermediate feature map B9 according to the spatial attention mask to obtain the intermediate feature map D1; S4: Input the intermediate feature map D1 into the first decoding unit for processing to obtain the intermediate feature map D2; After the intermediate feature map B8 is skipped and connected to the intermediate feature map D2, it is input into the second decoding unit for processing to obtain the intermediate feature map D3. After the intermediate feature map B7 is skipped and connected to the intermediate feature map D3, it is input into the third decoding unit for processing to obtain the intermediate feature map D4. After the intermediate feature map B6 is skipped and connected to the intermediate feature map D4, it is input into the fourth decoding unit for processing to obtain the intermediate feature map D5. After the intermediate feature map B5 is skipped and connected to the intermediate feature map D5, it is input into the fifth decoding unit for processing to obtain the intermediate feature map D6. After skipping and connecting intermediate feature map B4 and intermediate feature map D4 to intermediate feature map D6, the intermediate feature map D7 is input into the sixth decoding unit for processing. After skipping and connecting intermediate feature map B3 and intermediate feature map D3 to intermediate feature map D7, the intermediate feature map D8 is input into the seventh decoding unit for processing. After skipping and connecting intermediate feature map B2 and intermediate feature map D2 to intermediate feature map D8, the intermediate feature map D9 is input into the eighth decoding unit for processing. The intermediate feature map D9 is input into the ninth decoding unit for processing to obtain the intermediate feature map E1. S5: Input the intermediate feature map E1 into the output layer for processing to obtain the output result. ; S6: Based on the output results Label mask in the training set of underwater acoustic signal time-frequency feature samples The calculation of the loss function; S7: Repeat S1 to S6, using the adaptive moment estimator (Adam) optimizer to iteratively update the parameters of the underwater acoustic signal image segmentation and detection model according to the loss function. Stop training when the loss function converges to obtain the trained underwater acoustic signal image segmentation and detection model.
6. The method for detecting water-pounding acoustic signals based on a deep learning image segmentation model according to claim 5, characterized in that, In step S2, intermediate feature maps B2, B3, B4, B5, B6, B7, and B8 are input into the U-Net mask branch for processing to obtain the spatial attention mask; the specific process is as follows: S2.1: Input the intermediate feature map B2 into the first sub-encoding unit for processing to obtain the intermediate feature map C1; input the intermediate feature map C1 into the first max pooling layer for processing to obtain the intermediate feature map C2; S2.2: After the intermediate feature map B3 is skipped and connected to the intermediate feature map C2, it is input into the second sub-coding unit for processing to obtain the intermediate feature map C3; The intermediate feature map C3 is input into the second max pooling layer for processing to obtain the intermediate feature map C4; S2.3: After skipping and connecting the intermediate feature map B4 to the intermediate feature map C4, it is input into the third sub-encoding unit for processing to obtain the intermediate feature map C5; The intermediate feature map C5 is input into the third max pooling layer for processing to obtain the intermediate feature map C6; S2.4: After skipping and connecting the intermediate feature map B5 to the intermediate feature map C6, it is input into the fourth sub-coding unit for processing to obtain the intermediate feature map C7; The intermediate feature map C7 is input into the fourth max pooling layer for processing to obtain the intermediate feature map C8; S2.5: After skipping and connecting the intermediate feature map B6 to the intermediate feature map C8, input it into the fifth sub-encoding unit for processing to obtain the intermediate feature map C9; The intermediate feature map C9 is input into the fifth max pooling layer for processing to obtain the intermediate feature map C10; S2.6: After skipping and connecting the intermediate feature map B7 to the intermediate feature map C10, it is input into the sixth sub-encoding unit for processing to obtain the intermediate feature map C11; The intermediate feature map C11 is input into the sixth max pooling layer for processing to obtain the intermediate feature map C12; S2.7: After skipping and connecting the intermediate feature map B8 to the intermediate feature map C12, input it into the seventh sub-coding unit for processing to obtain the intermediate feature map C13; The intermediate feature map C13 is input into the seventh max pooling layer for processing to obtain the intermediate feature map C14; S2.8: Input the intermediate feature map C14 into the first upsampling layer for processing to obtain the intermediate feature map C15; The intermediate feature map C15 is input into the first sub-decoding unit for processing to obtain the intermediate feature map C16. S2.9: The intermediate feature map C16 is input into the second upsampling layer for processing to obtain the intermediate feature map C17; After the intermediate feature map C11 is skipped and connected to the intermediate feature map C17, it is input into the second sub-decoding unit for processing to obtain the intermediate feature map C18. S2.10: The intermediate feature map C18 is input into the third upsampling layer for processing to obtain the intermediate feature map C19; After the intermediate feature map C9 is skipped and connected to the intermediate feature map C19, it is input into the third sub-decoding unit for processing to obtain the intermediate feature map C20. S2.11: The intermediate feature map C20 is input into the fourth upsampling layer for processing to obtain the intermediate feature map C21; After the intermediate feature map C7 is skipped and connected to the intermediate feature map C21, it is input into the fourth sub-decoding unit for processing to obtain the intermediate feature map C22. S2.12: The intermediate feature map C22 is input into the fifth upsampling layer for processing to obtain the intermediate feature map C23; After the intermediate feature map C5 is skipped and connected to the intermediate feature map C23, it is input into the fifth sub-decoding unit for processing to obtain the intermediate feature map C24. S2.13: The intermediate feature map C24 is input into the sixth upsampling layer for processing to obtain the intermediate feature map C24; After connecting intermediate feature map C3 to intermediate feature map C24, it is input into the sixth sub-decoding unit for processing to obtain intermediate feature map C25. S2.14: The intermediate feature map C25 is input into the seventh upsampling layer for processing to obtain the intermediate feature map C26; After the intermediate feature map C1 is skipped and connected to the intermediate feature map C26, it is input into the seventh sub-decoding unit for processing to obtain the spatial attention mask.
7. The method for detecting water-pounding acoustic signals based on a deep learning image segmentation model according to claim 6, characterized in that, The specific process in S1, S2, and S3 of skipping and connecting one intermediate feature map to another intermediate feature map. Expressed as a formula: In the formula, Feature maps representing skip connections, The feature map representing the skip connections. This represents the feature map after the skip connections. Indicates a skip connection. The expression represents element-wise multiplication, sigmoid() represents the sigmoid activation function, maxpool() represents max pooling, conv represents convolution, and tanh() represents the tanh activation function.
8. The method for detecting water-pounding acoustic signals based on a deep learning image segmentation model according to claim 7, characterized in that, The loss function used in S6 is the cross-entropy loss function.
9. The method for detecting water-pounding acoustic signals based on a deep learning image segmentation model according to claim 8, characterized in that, The second type of underwater acoustic data to be detected is a one-dimensional time-series signal received by a hydrophone.
10. The method for detecting water-pounding acoustic signals based on a deep learning image segmentation model according to claim 9, characterized in that, The second step involves preprocessing the underwater acoustic data to be detected to obtain a four-channel video feature matrix of the underwater acoustic data. The specific process is as follows: A1: Perform short-time Fourier transform on the one-dimensional time series signal to be detected to obtain the two-dimensional time-frequency complex matrix to be detected; A2: Extract the real part, imaginary part, amplitude, and phase of all complex numbers in the two-dimensional time-frequency complex matrix to be detected; A3: Construct a four-channel video feature matrix of the underwater acoustic data to be detected based on the real part, imaginary part, amplitude, and phase of all complex numbers in the two-dimensional time-frequency complex matrix to be detected.