Single-frequency pulse signal detection method in complex underwater acoustic environment

Abstract

The application discloses a single-frequency pulse signal detection method in a complex underwater acoustic environment and belongs to the field of target detection in a complex underwater acoustic environment. The application solves the problem that the existing detection method ignores the unique structural characteristics of target signals and background noise, resulting in low detection accuracy of target pulse signals in a complex marine environment noise background. The application adopts a data-driven method, does not require artificial experience parameters, and realizes linear skeleton feature extraction and structural enhancement through a linear skeleton structure feature extraction module and a graph attention skeleton feature connection enhancement module, effectively extracts the unique feature structure of target signals and background noise, solves the signal detection problem under the conditions of low input signal-to-interference noise ratio and coexistence of stationary noise components and non-stationary interference components in marine environment noise in a complex marine environment noise background, can stably extract and enhance the linear skeleton feature of the target signal in a complex underwater acoustic environment, and significantly improves the accuracy and robustness of target pulse signal detection. The application is mainly applied in a complex underwater acoustic environment.

Description

Technical Field

[0001] This invention belongs to the field of target detection in complex underwater acoustic environments. Background Technology

[0002] In underwater acoustic positioning scenarios, receiving hydrophones mounted on fixed platforms detect single-frequency pulse signals emitted by high-speed moving targets, providing prior information for subsequent positioning calculations. Due to the extremely high energy of marine environmental noise, target signals are often submerged, significantly reducing the signal-to-interference-plus-noise ratio (SNR) of the hydrophone received signal. Furthermore, complex marine environmental noise contains not only stationary noise components but also non-stationary pulse interference components, further increasing the difficulty of signal processing. Model-driven signal detection techniques rely on accurate analysis and modeling of the target signal and marine environmental noise. However, in complex marine environmental noise environments, the analysis and modeling of the target signal and marine environmental noise often deviate significantly from the actual marine environment, leading to model mismatch and severe performance degradation for model-driven signal processing methods. Classic data-driven deep learning methods often employ convolutional neural networks for end-to-end adaptive feature extraction of the original signal, but neglect the unique structural characteristics of the target signal and background noise, resulting in low accuracy in target pulse signal detection. These problems urgently need to be addressed. Summary of the Invention

[0003] The purpose of this invention is to address the problem that existing detection methods ignore the unique structural characteristics of target signals and background noise, resulting in low accuracy in detecting target pulse signals in complex marine environments with noise backgrounds. This invention provides a method for detecting single-frequency pulse signals in complex underwater acoustic environments.

[0004] A method for detecting single-frequency pulse signals in complex underwater acoustic environments, used to detect whether a moving target emits a position signal, wherein the position signal is a single-frequency pulse signal, the method comprising:

[0005] Construct the training set;

[0006] Construct a pulse signal detection network;

[0007] The pulse signal detection network is trained using the training set to obtain the trained pulse signal detection network;

[0008] After preprocessing the time-domain signal received by the hydrophone, the normalized short-time Fourier transform amplitude spectrum feature is input into the trained pulse signal detection network for extraction, enhancement and detection of multi-level linear skeleton structure features. The network outputs the detection confidence score for the presence of a single-frequency pulse signal. If the detection confidence score for the presence of a single-frequency pulse signal exceeds the threshold, it is determined that the current moving target has emitted a position signal; otherwise, it is determined that the current moving target has not emitted a position signal.

[0009] Preferably, the pulse signal detection network includes a linear skeleton structure feature extraction module, a graph attention skeleton feature connection enhancement module, and a detection module;

[0010] The linear skeleton structure feature extraction module is used to extract multi-level linear skeleton structure features from the received normalized short-time Fourier transform amplitude spectrum features. After the topological connection between the extracted linear skeleton structure features is enhanced by the graph attention skeleton feature connection enhancement module, the linear skeleton structure features with enhanced topological connection are obtained. The detection module performs pulse signal detection and outputs the detection confidence of the presence of a single-frequency pulse signal.

[0011] Preferably, the linear skeleton structure feature extraction module is composed of five sequentially connected feature extraction sub-modules cascaded together, with each feature extraction sub-module used for extracting linear skeleton structure features;

[0012] The normalized short-time Fourier transform amplitude spectrum features received by the linear skeleton structure feature extraction module are used as the input features of the first feature extraction submodule.

[0013] The linear skeleton structure features output by the linear skeleton structure extraction module are used as the output features of the last feature extraction submodule.

[0014] Preferably, the linear skeleton structure feature extraction process for each feature extraction submodule is as follows:

[0015] ;

[0016] ;

[0017] ;

[0018] ;

[0019] in, It is a linear rectified activation function. For normalization layer, It is a 3×3 convolutional layer. For max pooling layer, These are the structural features of the first to third intermediate skeletons, respectively. These are the input and output of the current feature extraction submodule, respectively, and the output features of the last feature extraction submodule in the linear skeleton structure feature extraction module are also used as the output of the linear skeleton structure feature extraction module. .

[0020] Preferably, the processing flow of the graph attention skeleton feature connection enhancement module is as follows:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] in, , These are the input and output of the graph attention skeleton feature connection enhancement module, respectively, and the input of the graph attention skeleton feature connection enhancement module is also the output of the linear skeleton structure feature extraction module. The output of the graph attention skeleton feature connection enhancement module is also used as the input of the detection module. ; These are the first to third intermediate features of the graph attention skeleton feature connection enhancement module. Represents the adjacency matrix. This represents the graph attention layer. Represents the activation function of the exponential linear unit. This indicates the average pooling operation.

[0026] Preferably, the processing flow of the detection module is as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] in, ,and For the input of the detection module, The output of the graph attention skeleton feature connection enhancement module is... These are the first to third intermediate features of the detection module. It is a linear rectified activation function. For normalization layer, It is a fully connected layer. The detection confidence level is the presence of a single-frequency pulse signal.

[0032] Preferably, the loss function of the pulse signal detection network is the cross-entropy loss. ;

[0033] ;

[0034] ;

[0035] in, This indicates whether the single-frequency pulse signal emitted by the moving target actually exists. It indicates that it truly exists. This indicates that it does not exist. This is the mapping function between the input and output of the pulse signal detection network. These are the parameters of the pulse signal detection network. To determine the detection confidence level for the existence of a single-frequency pulse signal, This refers to the normalized short-time Fourier transform amplitude spectrum characteristics received by the pulse signal detection network.

[0036] Preferably, the training set is constructed as follows:

[0037] Different motion parameters and signal parameters are set for the moving target to generate a simulated time-domain signal sequence that is received by the hydrophone;

[0038] The signal parameters include whether the moving target has emitted a position signal or not, and the position signal emitted by the moving target is a single-frequency pulse signal;

[0039] Preprocessing is performed on each time-domain signal in the simulated time-domain signal sequence to obtain the normalized short-time Fourier transform amplitude spectrum characteristics;

[0040] The normalized short-time Fourier transform amplitude spectrum features corresponding to each time-domain signal are used as input data for the pulse signal detection network, and the existence of a single-frequency pulse signal emitted by the moving target is used as a label to form a training sample, thereby realizing the construction of each training sample in the training set.

[0041] Preferably, the preprocessing of the time-domain signal is implemented as follows:

[0042] The short-time Fourier transform (SFT) of the time-domain signal is performed to obtain the SFT amplitude spectrum characteristics, and these SFT amplitude spectrum characteristics are then normalized.

[0043] The beneficial effects of this invention are:

[0044] This invention proposes a method for detecting single-frequency pulse signals in complex underwater acoustic environments. This method sequentially extracts and enhances the linear skeleton features of the target signal, ultimately achieving effective detection of the target signal and improving the accuracy of target pulse signal detection.

[0045] Specifically, considering the linear skeleton structure of the target signal features in the time-frequency domain, a linear skeleton structure feature extraction module is designed to adaptively extract relevant feature structures of the target signal while simultaneously suppressing marine environmental noise-related feature responses. Secondly, taking advantage of the strong connections between points in the target signal skeleton features, a graph attention skeleton feature connection enhancement module is used to further enhance the connections and topological connectivity between points in the target feature structure. Finally, the extracted features are input into a detection module to achieve high-precision and robust detection of target signals in complex underwater acoustic environments.

[0046] This invention employs a data-driven approach, eliminating the need for manual experience parameters. Through a linear skeleton structure feature extraction module and a graph attention skeleton feature connection enhancement module, linear skeleton feature extraction and structure enhancement are achieved respectively. This effectively extracts the unique feature structures of the target signal and background noise, solving the signal detection challenges under conditions of low input signal-to-interference-plus-noise ratio and the coexistence of stationary and non-stationary interference components in complex marine environments. In complex underwater acoustic environments, it can stably extract and enhance the linear skeleton features of the target signal, significantly improving the accuracy and robustness of target pulse signal detection. Attached Figure Description

[0047] Figure 1 This is a flowchart of a single-frequency pulse signal detection method in a complex underwater acoustic environment as described in this invention;

[0048] Figure 2 This is a schematic diagram of the scenario of the present invention;

[0049] Figure 3 This is a block diagram illustrating the principle of the linear skeleton structure feature extraction module of the present invention.

[0050] Figure 4 This is a structural diagram of the pulse signal detection network of the present invention;

[0051] Figure 5 Here is a structural diagram of the feature extraction submodule;

[0052] Figure 6 This is a structural diagram of the graph attention skeleton feature connection enhancement module;

[0053] Figure 7 This is a structural diagram of the detection module;

[0054] Figure 8 The target signal component in the transmitted and received signals is represented by its time-domain waveform.

[0055] Figure 9 This is a time-frequency diagram of the target signal component in the transmitted signal;

[0056] Figure 10 The time-frequency diagram of the target signal component in the received signal;

[0057] Figure 11 The waveform of marine environmental noise in the time domain;

[0058] Figure 12 This is a time-frequency diagram of marine environmental noise.

[0059] Figure 13 The received signal is in the time domain waveform.

[0060] Figure 14 This is a time-frequency diagram of the received signal;

[0061] Figure 15 ROC curves for different signal detection methods under different input signal-to-noise ratios;

[0062] Figure 16 ROC curves for different signal detection methods under different input signal-to-interference ratios. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0065] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0066] Source of idea:

[0067] Physical scene diagram as follows Figure 2 As shown.

[0068] The moving target from its starting position With speed Move to the finish line The transmitting transducer mounted on the moving target emits the target signal during the movement, while the receiving hydrophone mainly collects the target signal propagated through the ocean and the marine environmental noise.

[0069] set up The coordinates of the point are The target's coordinates during its movement are During the target's movement, due to the relative radial motion between the target and the receiving hydrophone, and influenced by the Doppler effect, the relative radial velocity of the target relative to the receiving hydrophone can be expressed as:

[0070] ;

[0071] Due to the target's movement, at the moment of launch The time when the signal is transmitted and received It can be represented as

[0072] ;

[0073] The above equation reflects the Doppler effect caused by the radial motion of the target, which compresses or expands the received signal along the time axis. Consider the time delay of signal propagation to the receiver. Considering the amplitude attenuation during propagation, the received target signal can be expressed as:

[0074] ;

[0075] The receiving hydrophone mainly collects the target signal. In relation to marine environmental noise, consider the presence of stationary noise components. With non-stationary pulse interference components Hydrophone receives signals It can be represented as:

[0076] ;

[0077] The short-time Fourier transform amplitude spectrum characteristics of the hydrophone received signal are obtained by processing the time-domain signal using the short-time Fourier transform. The hydrophone received signal can be expressed as:

[0078] ;

[0079] in, , , , These are the time-frequency domain representations of the hydrophone received signal, the target signal, interference, and noise, respectively. The first characteristic of the amplitude spectrum of the short-time Fourier transform One time frame; The first characteristic of the amplitude spectrum of the short-time Fourier transform Each frequency frame.

[0080] Characteristics of the amplitude spectrum of short-time Fourier transform Normalization is performed to obtain the normalized short-time Fourier transform amplitude spectrum characteristics. ;

[0081] ;

[0082] The normalized short-time Fourier transform amplitude spectrum features can characterize the differences between the target signal and marine environmental noise in the time and frequency domains, providing a reliable feature basis for subsequent target signal detection methods and improving signal recognition accuracy.

[0083] Specific Implementation Method 1: Combination Figure 1 This embodiment describes a single-frequency pulse signal detection method in a complex underwater acoustic environment, used to detect whether a moving target emits a position signal, wherein the position signal is a single-frequency pulse signal. The method includes:

[0084] Construct the training set;

[0085] Construct a pulse signal detection network;

[0086] The pulse signal detection network is trained using the training set to obtain the trained pulse signal detection network;

[0087] After preprocessing the time-domain signal received by the hydrophone, the normalized short-time Fourier transform amplitude spectrum feature is input into the trained pulse signal detection network for extraction, enhancement and detection of multi-level linear skeleton structure features. The network outputs the detection confidence score for the presence of a single-frequency pulse signal. If the detection confidence score for the presence of a single-frequency pulse signal exceeds the threshold, it is determined that the current moving target has emitted a position signal; otherwise, it is determined that the current moving target has not emitted a position signal.

[0088] This implementation reliably detects whether a moving target is emitting a single-frequency pulse position signal. It replaces the manual threshold setting or simple filtering that traditional methods might rely on by training a specialized pulse signal detection network. In actual detection, only preprocessing of the hydrophone signal and extraction of short-time Fourier transform amplitude spectrum features are needed. This information is then input into the network to automatically output the detection confidence level. Finally, by comparing this confidence level with a preset threshold, a clear determination of whether a position signal exists is given, thus achieving efficient and automated processing from the raw signal to the recognition result.

[0089] The implementation method for preprocessing time-domain signals is as follows:

[0090] The short-time Fourier transform (SFT) of the time-domain signal is performed to obtain the SFT amplitude spectrum characteristics, and these SFT amplitude spectrum characteristics are then normalized.

[0091] The specific implementation method for constructing a pulse signal detection network is as follows:

[0092] See Figure 3 and Figure 4 The pulse signal detection network includes a linear skeleton structure feature extraction module, a graph attention skeleton feature connection enhancement module, and a detection module;

[0093] The linear skeleton structure feature extraction module is used to extract multi-level linear skeleton structure features from the received normalized short-time Fourier transform amplitude spectrum features. After the topological connection between the extracted linear skeleton structure features is enhanced by the graph attention skeleton feature connection enhancement module, the linear skeleton structure features with enhanced topological connection are obtained. The detection module performs pulse signal detection and outputs the detection confidence of the presence of a single-frequency pulse signal.

[0094] By using a linear skeleton structure feature extraction module and a graph attention skeleton feature connection enhancement module, linear skeleton feature extraction and structure enhancement are achieved respectively, effectively solving the signal detection problem under conditions of low input signal-to-interference-plus-noise ratio and coexistence of stationary noise components and non-stationary interference components in marine environmental noise.

[0095] Furthermore, considering the characteristic of the target signal exhibiting a linear skeleton structure in the time-frequency domain, this invention designs a linear skeleton structure feature extraction module to effectively extract the target skeleton features. (See [link to relevant documentation]). Figure 4 The linear skeleton structure feature extraction module consists of five sequentially connected feature extraction sub-modules, each of which is used to extract linear skeleton structure features.

[0096] The normalized short-time Fourier transform amplitude spectrum features received by the linear skeleton structure feature extraction module are used as the input features of the first feature extraction submodule.

[0097] The linear skeleton structure features output by the linear skeleton structure extraction module are used as the output features of the last feature extraction submodule.

[0098] See Figure 5 The linear skeleton structure feature extraction process for each feature extraction submodule is as follows:

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] in, It is a linear rectified activation function. For normalization layer, It is a 3×3 convolutional layer. For max pooling layer, These are the structural features of the first to third intermediate skeletons, respectively. These are the input and output of the current feature extraction submodule, respectively, and the output features of the last feature extraction submodule in the linear skeleton structure feature extraction module are also used as the output of the linear skeleton structure feature extraction module. .

[0104] In this preferred embodiment, the feature extraction submodule employs a combination of standard 3×3 convolutional layers and max-pooling layers. Each convolutional layer is responsible for extracting and combining local features, while the subsequent max-pooling layers gradually reduce the spatial resolution of the features while retaining the most salient feature responses. Through this repeated stacking structure, the network can perform deep feature extraction and spatial downsampling, gradually extracting high-level semantic features from the original time-frequency input, laying the foundation for the final detection task.

[0105] Furthermore, the graph attention skeleton feature connection enhancement module, based on the linear topology of the target pulse signal in the feature space, introduces a multi-layer graph attention network to enhance the intrinsic connections between the nodes of the signal. The structure of the graph attention skeleton feature connection enhancement module is as follows: Figure 6 As shown. Specifically, the processing flow of the graph attention skeleton feature connection enhancement module is as follows:

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] in, , These are the input and output of the graph attention skeleton feature connection enhancement module, respectively, and the input of the graph attention skeleton feature connection enhancement module is also the output of the linear skeleton structure feature extraction module. The output of the graph attention skeleton feature connection enhancement module is also used as the input of the detection module. ; These are the first to third intermediate features of the graph attention skeleton feature connection enhancement module. Represents the adjacency matrix. This represents the graph attention layer. Represents the activation function of the exponential linear unit. This indicates the average pooling operation.

[0111] In this preferred embodiment, the graph attention skeleton feature connection enhancement module comprises three graph attention layers. Each graph attention layer dynamically aggregates the feature information of neighboring nodes by calculating the attention weights between nodes. Since the graph attention mechanism can adaptively assign higher weights to important connections, this allows the module to focus on key nodes of the target signal, effectively suppressing outliers caused by noise or interference. By stacking multiple graph attention layers, the module achieves node feature interaction, strengthens the linear structural integrity of the target pulse signal in the feature space, and provides a more robust graph structure feature representation for subsequent detection tasks.

[0112] Furthermore, the detection module integrates the features extracted by the linear skeleton structure feature extraction module and the graph attention skeleton feature connection enhancement module to achieve the final target signal detection. See also Figure 7 The processing flow of the detection module is as follows:

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] in, ,and For the input of the detection module, The output of the graph attention skeleton feature connection enhancement module is... These are the first to third intermediate features of the detection module. It is a linear rectified activation function. For normalization layer, It is a fully connected layer. The detection confidence level is the presence of a single-frequency pulse signal.

[0118] The detection module further extracts high-dimensional features through a series of linear transformation layers and activation function layers. The final output layer of the module is the detection confidence level of the target's presence. This structure makes full use of the output of the front-end feature extraction network, achieving accurate detection of target pulse signals even in the context of low input signal-to-interference-plus-noise ratio and non-stationary marine environmental noise.

[0119] In practical applications, the training set is constructed as follows:

[0120] Different motion parameters and signal parameters are set for the moving target to generate a simulated time-domain signal sequence that is received by the hydrophone;

[0121] The signal parameters include whether the moving target has emitted a position signal or not, and the position signal emitted by the moving target is a single-frequency pulse signal;

[0122] Preprocessing is performed on each time-domain signal in the simulated time-domain signal sequence to obtain the normalized short-time Fourier transform amplitude spectrum characteristics;

[0123] The normalized short-time Fourier transform amplitude spectrum features corresponding to each time-domain signal are used as input data for the pulse signal detection network, and the existence of a single-frequency pulse signal emitted by the moving target is used as a label to form a training sample, thereby realizing the construction of each training sample in the training set.

[0124] The loss function of the pulse signal detection network is cross-entropy loss. ;

[0125] ;

[0126] ;

[0127] in, This indicates whether the single-frequency pulse signal emitted by the moving target actually exists. It indicates that it truly exists. This indicates that it does not exist. This is the mapping function between the input and output of the pulse signal detection network. These are the parameters of the pulse signal detection network. To determine the detection confidence level for the existence of a single-frequency pulse signal, This refers to the normalized short-time Fourier transform amplitude spectrum characteristics received by the pulse signal detection network.

[0128] Verification experiment:

[0129] The target signal is set as a single-frequency pulse signal with a center frequency of 4000 Hz and a pulse duration of 0.01 s. The marine environmental noise is modeled as a superposition of stationary noise components in the form of Gaussian white noise filtered by a bandpass filter and non-stationary pulse interference components in the form of Alpha stable distribution noise. The passband of the bandpass filter is 100-600 Hz, the characteristic exponent of the Alpha stable distribution noise is 1.8, the skew parameter is 0, the scale parameter is 1, and the position parameter is 0. The received signal-to-noise ratio (SNR) and signal-to-interference ratio (SINR) are set to -7.5 dB. Relative motion is assumed between the target and the receiver, with an initial azimuth angle of 30°, a target motion azimuth angle of 40°, and a target velocity of 60 m / s. The target signal components in the transmitted signal, the target signal components in the received signal, and the time-domain waveforms and time-frequency diagrams of the marine environmental noise and the received signal are plotted as follows. Figure 8-14 As shown.

[0130] Depend on Figure 8-14 It is known that the target signal component in the transmitted signal appears between 0.02 and 0.03 seconds, with a center frequency of 4000 Hz. As the target moves away from the receiver, there is a time delay and Doppler frequency shift between the target signal components in the transmitted and received signals. Because the marine environmental noise includes both stationary and non-stationary impulse interference components, and the signal-to-interference-plus-noise ratio (SNR) at the receiver is extremely low, the received signal waveform exhibits obvious spike pulse characteristics in the time domain. In the time-frequency domain, multiple longitudinal bright bands caused by impulse interference appear in the signal's time spectrum. The target signal characteristics are almost completely submerged by the background noise, making it impossible to effectively identify valid target signal features.

[0131] Based on the above analysis, the single-frequency pulse signal detection performance of the method of this invention (DL) and the classical Fourier transform method (FFT) is compared and analyzed. The signal detection performance of the method of this invention and the Fourier transform method is examined when the input signal-to-interference ratio and input signal-to-noise ratio are [-12.5:3.75:-5] dB, and the ROC curves of the two methods are plotted as follows: Figure 15-16 As shown.

[0132] Depend on Figure 15-16 It is evident that, under different input signal-to-interference-plus-noise ratios (SNRs), the performance of the method of this invention is significantly superior to that of the Fourier transform method. As the input SNR decreases, the detection probability of both methods decreases to varying degrees. Due to the short signal duration, the energy accumulation of a single-frequency pulse signal in the frequency domain is limited, causing the Fourier transform method to experience a significant drop in detection performance, almost to the point of failure, when the input SNR is low. In contrast, the method of this invention effectively achieves robust detection of target signals in complex underwater acoustic environments by capturing the linear skeleton structure of target features in the time-frequency domain. Figure 15-16 The detection probability is the probability that the system correctly identifies a target when it actually exists. The false alarm probability is the probability that the system incorrectly identifies a target when it does not exist (only noise / interference exists).

[0133] To further analyze the signal detection performance of the method of the present invention and the classical detection method under different input signal-to-interference-plus-noise ratios, the signal detection performance of the method of the present invention and the Fourier transform method was examined when the input signal-to-interference-plus-noise ratio was [-12.5:2.5:-5] dB. The first detection probability table corresponding to the false alarm probability of 0.1 is shown in Table 1 below.

[0134] Table 1 First Detection Probability Table

[0135]

[0136] As shown in Table 1, for single-frequency pulse signals, the signal detection performance of the method of this invention is significantly better than that of the Fourier transform method under all tested signal-to-interference ratios (SIR). When the SIR decreases from -5 dB to -12.5 dB, the Fourier transform method's detection rate drops from 40.96% to 12.80%. This is because when the SIR decreases, the energy of the interference component in marine environmental noise is approximately evenly distributed in the frequency domain, resulting in complete overlap between the interference and the signal in the frequency domain. The Fourier transform method cannot effectively distinguish between the two based on spectral amplitude, and its detection probability at -5 dB is only 40.96%. The method of this invention achieves stable detection by adaptively extracting and enhancing the characteristic structure of the target signal in the time-frequency domain, fully utilizing the characteristic of single-frequency pulse signals exhibiting a continuous linear skeleton feature structure in the time-frequency domain.

[0137] The signal detection performance of the method of this invention and the Fourier transform method was examined when the input signal-to-noise ratio was [-12.5:2.5:-5] dB. The second detection probability table corresponding to the false alarm probability of 0.1 is shown in Table 2 below.

[0138] Table 2 Second Detection Probability Table

[0139]

[0140] Table 2 shows the detection probabilities of the proposed method and the Fourier transform method under different signal-to-noise ratios (SNRs). It is evident that the proposed method significantly outperforms the Fourier transform method at all tested SNRs. Because the noise components in marine environmental noise exhibit non-uniformity in the frequency domain (i.e., low-frequency noise energy is significantly higher than high-frequency noise energy), this can severely mask the target signal, leading to a sharp decrease in detection probability. The detection probability shows that when the SNR drops from -5 dB to -12.5 dB, the Fourier transform method's probability drops sharply from 63.16% to 10.53%, while the proposed method's probability only decreases from 99.60% to 82.46%. This demonstrates that the proposed method fully utilizes the characteristic structure of the target signal in the time-frequency domain to achieve stable and effective detection of single-frequency target pulse signals. In summary, the proposed method demonstrates better adaptability in target signal detection tasks.

[0141] To examine the advantages of the model of this invention in signal detection tasks, the method of this invention is compared with three classic deep learning models. The parameters of the three classic deep learning models are shown in Table 3. The third to fourth detection probabilities corresponding to the input signal-to-interference ratio (SINR) and input signal-to-noise ratio (SNR) of [-12.5:2.5:-5] dB are examined respectively, and the false alarm probability of 0.1 is shown in Tables 3-5.

[0142] Table 3 Deep Learning Model Parameters

[0143]

[0144] Table 4. Third Detection Probability Table

[0145]

[0146] Table 5. Fourth Detection Probability Table

[0147]

[0148] When the input signal-to-interference-plus-noise ratio (SINR) and signal-to-noise ratio (SNR) are -12.5 dB, the detection probabilities of the proposed method reach 82.63% and 82.46%, respectively, which are at least 20.90% and 5.63% higher than the comparison model. Furthermore, the performance advantage increases further as the input SINR decreases. These results demonstrate that the proposed model possesses stronger feature extraction and anti-interference capabilities, especially maintaining high detection stability under low input SINR conditions, fully validating the effectiveness of the proposed method in complex signal detection tasks.

[0149] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for detecting single-frequency pulse signals in complex underwater acoustic environments, characterized in that, This method is used to detect whether a moving target emits a position signal, wherein the position signal is a single-frequency pulse signal. The method includes: Construct the training set; Construct a pulse signal detection network; The pulse signal detection network is trained using the training set to obtain the trained pulse signal detection network; After preprocessing the time-domain signal received by the hydrophone, the normalized short-time Fourier transform amplitude spectrum feature is input into the trained pulse signal detection network for extraction, enhancement and detection of multi-level linear skeleton structure features. The network outputs the detection confidence score for the presence of a single-frequency pulse signal. If the detection confidence score for the presence of a single-frequency pulse signal exceeds the threshold, it is determined that the current moving target has emitted a position signal; otherwise, it is determined that the current moving target has not emitted a position signal.

2. The method for detecting single-frequency pulse signals in a complex underwater acoustic environment according to claim 1, characterized in that, The pulse signal detection network includes a linear skeleton structure feature extraction module, a graph attention skeleton feature connection enhancement module, and a detection module; The linear skeleton structure feature extraction module is used to extract multi-level linear skeleton structure features from the received normalized short-time Fourier transform amplitude spectrum features. After the topological connection between the extracted linear skeleton structure features is enhanced by the graph attention skeleton feature connection enhancement module, the linear skeleton structure features with enhanced topological connection are obtained. The detection module performs pulse signal detection and outputs the detection confidence of the presence of a single-frequency pulse signal.

3. The method for detecting single-frequency pulse signals in a complex underwater acoustic environment according to claim 2, characterized in that, The linear skeleton structure feature extraction module consists of five sequentially connected feature extraction sub-modules, each of which is used to extract linear skeleton structure features. The normalized short-time Fourier transform amplitude spectrum features received by the linear skeleton structure feature extraction module are used as the input features of the first feature extraction submodule. The linear skeleton structure features output by the linear skeleton structure extraction module are used as the output features of the last feature extraction submodule.

4. The method for detecting single-frequency pulse signals in a complex underwater acoustic environment according to claim 3, characterized in that, The linear skeleton structure feature extraction process for each feature extraction submodule is as follows: ； ； ； ； in, It is a linear rectified activation function. For normalization layer, It is a 3×3 convolutional layer. For max pooling layer, These are the structural features of the first to third intermediate skeletons, respectively. These are the input and output of the current feature extraction submodule, respectively, and the output features of the last feature extraction submodule in the linear skeleton structure feature extraction module are also used as the output of the linear skeleton structure feature extraction module. .

5. The method for detecting single-frequency pulse signals in a complex underwater acoustic environment according to claim 2, characterized in that, The processing flow of the graph attention skeleton feature connection enhancement module is as follows: ； ； ； ； in, , These are the input and output of the graph attention skeleton feature connection enhancement module, respectively, and the input of the graph attention skeleton feature connection enhancement module is also the output of the linear skeleton structure feature extraction module. The output of the graph attention skeleton feature connection enhancement module is also used as the input of the detection module. ; These are the first to third intermediate features of the graph attention skeleton feature connection enhancement module. Represents the adjacency matrix. This represents the graph attention layer. Represents the activation function of the exponential linear unit. This indicates the average pooling operation.

6. The method for detecting single-frequency pulse signals in a complex underwater acoustic environment according to claim 2, characterized in that, The processing flow of the detection module is as follows: ； ； ； ； in, ,and For the input of the detection module, The output of the graph attention skeleton feature connection enhancement module is... These are the first to third intermediate features of the detection module. It is a linear rectified activation function. For normalization layer, It is a fully connected layer. The detection confidence level is the presence of a single-frequency pulse signal.

7. The method for detecting single-frequency pulse signals in a complex underwater acoustic environment according to claim 1, characterized in that, The loss function of the pulse signal detection network is cross-entropy loss. ; ； ； in, This indicates whether the single-frequency pulse signal emitted by the moving target actually exists. It indicates that it truly exists. This indicates that it does not exist. This is the mapping function between the input and output of the pulse signal detection network. These are the parameters of the pulse signal detection network. To determine the detection confidence level for the existence of a single-frequency pulse signal, This refers to the normalized short-time Fourier transform amplitude spectrum characteristics received by the pulse signal detection network.

8. The method for detecting single-frequency pulse signals in a complex underwater acoustic environment according to claim 1, characterized in that, The training set is constructed as follows: Different motion parameters and signal parameters are set for the moving target to generate a simulated time-domain signal sequence that is received by the hydrophone; The signal parameters include whether the moving target has emitted a position signal or not, and the position signal emitted by the moving target is a single-frequency pulse signal; Preprocessing is performed on each time-domain signal in the simulated time-domain signal sequence to obtain the normalized short-time Fourier transform amplitude spectrum characteristics; The normalized short-time Fourier transform amplitude spectrum features corresponding to each time-domain signal are used as input data for the pulse signal detection network, and the existence of a single-frequency pulse signal emitted by the moving target is used as a label to form a training sample, thereby realizing the construction of each training sample in the training set.

9. The method for detecting single-frequency pulse signals in a complex underwater acoustic environment according to claim 1, characterized in that, The implementation method for preprocessing time-domain signals is as follows: The short-time Fourier transform (SFT) of the time-domain signal is performed to obtain the SFT amplitude spectrum characteristics, and these SFT amplitude spectrum characteristics are then normalized.

Images