Underwater weak signal arrival time prediction method, model training method and model

Abstract

The invention discloses an underwater weak signal arrival time prediction method, a model training method and a model. The method comprises the following steps: acquiring a received sound signal data sequence, and an original emission signal sequence corresponding to the sequence; performing feature extraction on the sound signal data sequence to obtain original signal features; performing cross-correlation calculation on the sound signal data sequence and a transmitting signal sequence to obtain cross-correlation parameters; performing feature extraction on the feature parameter to obtain a first cross-correlation feature; performing feature mask extraction on the first cross-correlation feature to obtain a cross-correlation feature mask; multiplying the cross-correlation feature mask by the first cross-correlation feature, adjusting the weight of the first cross-correlation feature, and outputting a second cross-correlation feature; carrying out feature extraction on the second cross-correlation feature and the original signal feature, and outputting a key feature; and predicting the arrival time of the underwater weak signal based on the key features. According to the method, the problem that the arrival time of the sound signal cannot be accurately detected by a traditional matched filtering method under the condition of low signal-to-noise ratio can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of acoustic signal processing, and in particular to an underwater weak signal arrival time prediction method, a model training method and a model. Background Art

[0002] Acoustic tomography is a technology that uses the propagation characteristics of sound waves to perform spatial imaging. The existing river acoustic tomography flow monitoring technology is a new river measurement technology with the advantages of non-invasiveness and real-time. Due to the ability to quickly and conveniently obtain temperature and flow parameters of a large observation area and the advantages of non-invasive observation, acoustic tomography has attracted many scholars to study. Using two-point acoustic tomography technology to conduct long-term observations of hydrological parameters in shallow seas, estuaries or waterways, the hydroacoustic transducers at two stations send and receive signals, and accurately calculate the travel time and sound speed of the acoustic signal at the two stations, the average temperature and flow information of the monitoring profile can be accurately inverted, and then combined with the dynamic water level and terrain, the river flux in the observed water area can be accurately calculated. Although the phase of the acoustic tomography transmission signal is modulated by the m sequence to improve the signal-to-noise ratio, due to the influence of natural conditions, such as the afternoon effect and the presence of a large amount of underwater noise, the signal-to-noise ratio will drop sharply, causing the receiving end to rely on the traditional matched filtering method. The method cannot accurately calculate the arrival time of the acoustic signal, affecting the final flow rate and flow calculation. Therefore, there is an urgent need for a technical solution that can monitor the arrival time of the acoustic signal under low signal-to-noise ratio conditions. With the development of neural networks, deep learning network models have brought new ideas to solve this problem. As a data-driven method, deep learning network models have strong adaptability and good generalization ability. Summary of the invention

[0003] In view of this, the purpose of the present invention is to propose a method for predicting the arrival time of underwater weak signals, a model training method and a model, which can solve the problem that the traditional matched filtering method cannot accurately detect the arrival time of sound signals under low signal-to-noise ratio conditions (signal-to-noise ratio lower than -10dB).

[0004] According to one aspect of the present invention, a method for predicting arrival time of underwater weak signals is provided, the method comprising:

[0005] Acquire a received acoustic signal data sequence and an original transmitted signal sequence corresponding to the sequence;

[0006] Performing feature extraction on the acoustic signal data sequence to obtain original signal features;

[0007] For the acoustic signal data sequence, a cross-correlation calculation is performed with the transmission signal sequence to obtain a cross-correlation parameter; for the characteristic parameter, a feature extraction is performed to obtain a first cross-correlation feature;

[0008] Performing feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiplying the mutual correlation feature mask by the first mutual correlation feature, adjusting the weight of the first mutual correlation feature, and outputting a second mutual correlation feature;

[0009] Based on the second cross-correlation feature and the original signal feature, feature extraction is performed to output key features;

[0010] Based on this key feature, the arrival time of underwater weak signals is predicted.

[0011] In the above technical solution, first, the received acoustic signal data sequence and its corresponding original transmitted signal sequence are obtained. Secondly, the acoustic signal data sequence is feature extracted to obtain the original signal feature. The feature extraction process aims to extract useful information from the signal and provide the necessary basis for subsequent analysis. Next, the acoustic signal data sequence and the original transmitted signal sequence are cross-correlated to calculate the cross-correlation parameter. The cross-correlation analysis can quantify the similarity between two signals. By calculating the cross-correlation parameter, the relative time delay of the signal can be determined. The cross-correlation parameter is further feature extracted to obtain the first cross-correlation feature. This step aims to further refine the key information in the cross-correlation parameter. Subsequently, the first cross-correlation feature is feature mask extracted to obtain the cross-correlation feature mask. The feature mask can be regarded as the screening or weighting of the feature. By applying the mask, the important part of the feature can be highlighted. The cross-correlation feature mask is multiplied by the first cross-correlation feature to adjust the weight of the first cross-correlation feature, thereby outputting the second cross-correlation feature. This step weights the feature by masking so that the feature is more in line with the actual application requirements. Next, feature extraction is performed on the second cross-correlation feature and the original signal feature to output the key feature. The key feature is the most valuable information obtained after processing, which can more accurately reflect the nature of the signal. Finally, the key feature contains important information about the signal, which can be used to more accurately predict the arrival time of the signal. Through multiple rounds of feature extraction and processing, the useful information in the signal can be fully mined, thereby improving the accuracy of the prediction. In a complex underwater environment, the signal may be subject to various interferences. This method can effectively suppress interference and maintain the stability of the prediction through means such as feature masks. This method is suitable for underwater communications, underwater detection and other fields, and can significantly improve the efficiency and accuracy of signal processing, providing strong technical support for related work.

[0012] In some embodiments, the acoustic signal data sequence is one-dimensional acoustic data.

[0013] In the above technical solution, currently, the input requirements of most neural network models are two-dimensional data, such as images. This method uses a one-dimensional sound signal sequence as input, and can directly analyze and process the received one-dimensional sound signal sequence without performing additional conversion operations, such as short-time Fourier transform, etc. These methods of converting one-dimensional data into two-dimensional data may cause the loss of certain features in the data. Therefore, this method can retain the features of the original signal to the maximum extent, thereby improving the accuracy and effectiveness of signal processing.

[0014] In some embodiments, a received acoustic signal data sequence and an original transmitted signal sequence corresponding to the sequence are obtained;

[0015] Performing feature extraction on the acoustic signal data sequence to obtain original signal features;

[0016] For the acoustic signal data sequence, a cross-correlation calculation is performed with the transmission signal sequence to obtain a cross-correlation parameter; for the characteristic parameter, a feature extraction is performed to obtain a first cross-correlation feature;

[0017] Set N rounds of feature extraction, and perform the following steps in each round of feature extraction:

[0018] Performing feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiplying the mutual correlation feature mask by the first mutual correlation feature, adjusting the weight of the first mutual correlation feature, and outputting a second mutual correlation feature;

[0019] The original signal features and the second cross-correlation features of this round are used as data input for the next round of original signal feature extraction, and the next round of feature extraction is performed; the first cross-correlation features of this round are used as data input for the next round of first cross-correlation feature extraction, and the next round of feature extraction is performed;

[0020] The above steps are repeated until the set number of feature extraction rounds is completed, and the final second cross-correlation feature and the final original signal feature are output;

[0021] For the final second cross-correlation feature and the final original signal feature, feature extraction is performed to output key features;

[0022] Based on this key feature, the arrival time of underwater weak signals is predicted.

[0023] In the above technical solution, the original signal features contain the basic information and initial features of the signal. By using them as the input for the next round of feature extraction, it can be ensured that these basic features continue to play a role in the entire iterative process, providing a stable foundation for the final feature extraction. The newly output feature (the second cross-correlation feature) is the optimized feature after feature mask adjustment in the current round, which contains the latest understanding and refinement of the signal features in the current round. By using it as the input for the next round of feature extraction, these optimized feature information can be further integrated into the subsequent feature extraction process to achieve the supplementation and optimization of feature information, making the final features more comprehensive and accurate. Through the accumulation of feature information, the expressive power of the features can be enhanced, so that the final features can more accurately describe the characteristics and nature of the signal. This enhanced expressive power helps to improve the performance of subsequent feature-based signal analysis and processing tasks.

[0024] According to another aspect of the present invention, a neural network model training method for underwater weak signal arrival time prediction is provided, wherein the neural network is a dual-channel network model, including a cross-correlation feature extraction channel, an original signal feature extraction channel, an interaction module, an attention mechanism module, and a fully connected layer; the method comprises:

[0025] Obtaining a sample acoustic signal data sequence, an original transmitted signal sequence corresponding to the sequence, and a true label corresponding to the sequence;

[0026] For the acoustic signal data sequence, input the original signal feature extraction channel to perform feature extraction to obtain the original signal features;

[0027] For the acoustic signal data sequence, a cross-correlation calculation is performed with the transmission signal sequence to obtain a cross-correlation parameter; for the feature parameter, the cross-correlation feature extraction channel is input to perform feature extraction to obtain a first cross-correlation feature;

[0028] The first mutual correlation feature is input into an interactive module to extract a feature mask to obtain a mutual correlation feature mask; the mutual correlation feature mask is multiplied by the first mutual correlation feature to adjust the weight of the first mutual correlation feature and output a second mutual correlation feature;

[0029] The second cross-correlation feature and the original signal feature are input into the attention mechanism module for feature extraction and key features are output;

[0030] Based on this key feature, the fully connected layer is input to predict the arrival time of underwater weak signals;

[0031] Based on the arrival time and the true label, calculate the loss value; and,

[0032] The neural network model is trained based on the loss value.

[0033] In the above technical scheme, in order to better use the above method, the present application proposes a neural network model training method for underwater weak signal arrival time prediction. The dual-channel network model includes two feature extraction channels, namely, a cross-correlation feature extraction channel and an original signal feature extraction channel. This dual-channel design can extract signal features from different angles. The cross-correlation feature extraction channel focuses on the time correlation feature of the signal, while the original signal feature extraction channel extracts the basic features of the signal. The interaction module is used to extract and adjust the weight of the first cross-correlation feature by feature mask, and output the second cross-correlation feature by multiplying the feature mask with the first cross-correlation feature. This process can highlight the important part of the feature and suppress unimportant information. The attention mechanism module further extracts the second cross-correlation feature and the original signal feature, and outputs the key features. The attention mechanism can make the model pay more attention to the important information in the feature, improve the feature expression ability and the prediction performance of the model. Based on the key features, the fully connected layer predicts the signal arrival time. The fully connected layer can comprehensively analyze the features and output the final prediction result. The dual-channel network model extracts signal features from different angles and can capture the characteristics and laws of the signal more comprehensively. Combining the interaction module and the attention mechanism module to optimize and focus on features enables the model to understand and analyze signals more accurately.

[0034] In some embodiments, obtaining a sample acoustic signal data sequence, an original transmitted signal sequence corresponding to the sequence, and a true label corresponding to the sequence includes:

[0035] collecting at least two types of background noise and at least two types of target sound signals;

[0036] For each target sound signal, selecting an adding time from the target sound signal, synthesizing the target sound signal and any background noise at the adding time, and using the synthesized result as a sample sound signal data sequence;

[0037] Based on the target generated signal as the original transmission signal sequence;

[0038] Based on the adding time as the true label of the sample acoustic signal data sequence;

[0039] Repeat the above synthesis steps until the data set is constructed.

[0040] In the above technical solution, background noise refers to various interfering sounds existing in the underwater environment, such as water flow sound, marine life sound, etc. Collecting different types of background noise can simulate the diversity of the underwater environment and provide rich background conditions for sample data. Different types of background noise have different frequency characteristics, intensity distribution, etc., which can help the model learn the ability to identify target signals in different noise environments. The target sound signal is a signal that needs to be detected and predicted for arrival time, such as the sound signal of an underwater target. Different types of target standard signals may have different characteristics, such as different frequency components, signal shapes, etc., which can improve the diversity and representativeness of the data set and enable the model to adapt to signals of different target types. For each target sound signal, an addition time is selected from the target sound signal. The addition time is the time point when the target sound signal is synthesized with the background noise. Selecting different addition times can simulate the situation where the target signal appears at different times and increase the time diversity of the data set. At the selected addition time, the target sound signal is synthesized with any background noise to obtain a sample sound signal data sequence. The synthesis process needs to consider factors such as signal superposition and noise interference to ensure that the synthesized sample sound signal data sequence can truly reflect the superposition of the target signal and the background noise in the underwater environment. The synthesized sample acoustic signal data sequence contains the characteristic information of the target signal and background noise, providing basic data for subsequent feature extraction and model training. Based on the target acoustic signal as the original emission signal sequence. The original emission signal sequence is the signal form of the target acoustic signal when it is not interfered by noise, and it contains the original characteristic information of the target signal. In the process of model training, the original emission signal sequence can be used as a reference signal to help the model better understand and learn the characteristics of the target signal. The true label makes the sample acoustic signal data sequence based on the added time as the true label. The true label is the actual arrival time of the target acoustic signal in the synthesized sample acoustic signal data sequence, which is an important basis for model training and prediction. By using the added time as the true label, it can be ensured that the model can accurately learn the arrival time characteristics of the target signal during the training process, thereby improving the prediction performance of the model. This method constructs an underwater acoustic signal dataset by synthesizing the target acoustic signal and background noise, and provides the original emission signal sequence and true label for each sample, which can provide high-quality and diverse data support for the training of the underwater signal processing model, and help improve the prediction performance and application effect of the model.

[0041] In some embodiments, the data set includes at least three classic low signal-to-noise ratio data sets of -10dB, -15dB, and -20dB.

[0042] In the above technical solution, when constructing a dataset for underwater weak signal arrival time prediction, at least three classic low signal-to-noise ratio datasets are included. These datasets can provide abundant signal samples in low signal-to-noise ratio environments, helping the model to better learn and adapt to the signal characteristics under low signal-to-noise ratio conditions. Under the signal-to-noise ratio condition of -10dB, the intensity of the signal and the noise are close, and the interference of noise on the signal is more obvious. This dataset can simulate the situation of the signal in a medium noise environment, and is used to train the signal detection and recognition ability of the model in a medium noise environment. Under the signal-to-noise ratio condition of -15dB, the intensity of the noise further increases, and the characteristics of the signal are more difficult to identify. This dataset can simulate the situation of the signal in a higher noise environment, and is used to train the robustness of the model in a high noise environment. Under the signal-to-noise ratio condition of -20dB, the intensity of the noise is much greater than the signal, and the signal is almost submerged by the noise. This dataset can simulate the situation of the signal in an extreme noise environment, and is used to train the extreme performance of the model in an extreme noise environment. These three datasets cover noise environments from low to high, and can comprehensively train the signal processing capabilities of the model under different signal-to-noise ratio conditions.

[0043] According to another aspect of the present invention, a neural network model for predicting the arrival time of underwater weak signals is provided. Based on the above method, the neural network is a dual-channel network model, comprising:

[0044] A cross-correlation feature extraction channel is configured to perform cross-correlation calculation on the acoustic signal data sequence and the transmission signal sequence to obtain a cross-correlation parameter; perform feature extraction on the feature parameter to obtain a first cross-correlation feature;

[0045] An original signal feature extraction channel, which is configured to perform feature extraction on the acoustic signal data sequence to obtain original signal features;

[0046] An interaction module is configured to perform feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiply the first mutual correlation feature by the mutual correlation feature mask, adjust the weight of the first mutual correlation feature, and output a second mutual correlation feature;

[0047] An attention mechanism module is configured to extract features from the second cross-correlation feature and the original signal feature and output key features;

[0048] The fully connected layer is configured to predict the arrival time of underwater weak signals based on the key features.

[0049] In the above technical scheme, in order to better use the above method, the present application proposes a neural network model for predicting the arrival time of underwater weak signals. Each part corresponds to each step of the above method. The specific principles of each step have been explained above and will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 It is a flow chart of an embodiment of a method for predicting arrival time of underwater weak signals of the present invention;

[0052] Figure 2 It is a flow chart of another embodiment of the underwater weak signal arrival time prediction method of the present invention;

[0053] Figure 3 It is a flow chart of an embodiment of a neural network model training method for underwater weak signal arrival time prediction according to the present invention;

[0054] Figure 4 It is a schematic diagram of transmitted acoustic signals of another embodiment of the underwater weak signal arrival time prediction method of the present invention;

[0055] Figure 5 It is a composite diagram of three signals of another embodiment of the underwater weak signal arrival time prediction method of the present invention;

[0056] Figure 6 It is a neural network model diagram of another embodiment of the underwater weak signal arrival time prediction method of the present invention;

[0057] Figure 7 It is a diagram of the interaction module structure of another embodiment of the underwater weak signal arrival time prediction method of the present invention;

[0058] Figure 8 Schematic diagram of an attention mechanism module of another embodiment of the underwater weak signal arrival time prediction method of the present invention;

[0059] Fig. 9 This is a diagram showing the prediction result of the acoustic signal starting position of another embodiment of the underwater weak signal arrival time prediction method of the present invention;

[0060] Fig.10 This is a schematic table comparing the detection results of the neural network model provided by another embodiment of the underwater weak signal arrival time prediction method of the present invention and some other methods on the real signal + real noise data set.

[0061] Fig.11This is a schematic comparison chart of the detection results of adding and subtracting some modules of the neural network model on the real signal + Gaussian white noise data set under the condition of a signal-to-noise ratio of -15dB provided by another embodiment of the underwater weak signal arrival time prediction method of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples. It is particularly noted that the following examples are only used to illustrate the present invention, but are not intended to limit the scope of the present invention. Similarly, the following examples are only partial embodiments of the present invention rather than all embodiments, and all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0063] The present invention provides an underwater weak signal arrival time prediction method, a model training method and a model, which can solve the problem that the traditional matched filtering method cannot accurately detect the arrival time of the sound signal under low signal-to-noise ratio conditions (signal-to-noise ratio is lower than -10dB).

[0064] Embodiment 1

[0065] See also Figure 1 , a method for predicting arrival time of underwater weak signals, the method comprising:

[0066] A1. Acquire a received acoustic signal data sequence and an original transmitted signal sequence corresponding to the sequence;

[0067] A2. Extract features from the acoustic signal data sequence to obtain original signal features;

[0068] A3. Perform cross-correlation calculation on the acoustic signal data sequence and the transmission signal sequence to obtain cross-correlation parameters; perform feature extraction on the feature parameters to obtain a first cross-correlation feature;

[0069] Cross-correlation is a mathematical operation used to compare the similarity of two signals (or data sequences). In signal processing, cross-correlation can help us analyze the position and overlap of one signal in another signal. The calculation process is as follows:

[0070] 1. Prepare the signal: Assume that we have two discrete signals, denoted as x[n] and y[n], where n is the discrete index of time.

[0071] 2. Select the shift amount (delay): Introduce a shift amount k, which represents the number of steps a signal needs to be translated. Cross-correlation calculates the similarity of signals at different delays by changing the value of k.

[0072] 3. Calculate the dot product sum: For each shift amount k:

[0073] Rxy [k] = Σ n x[n]·y[n+k]

[0074] Where y[n+k] represents the signal y shifted right by k units. If y[n+k] exceeds the signal boundary, it is usually regarded as zero (zero padding).

[0075] 4. Adjust the shift range: Usually, the shift amount k will vary from negative to positive. The specific range depends on the length of the signal.

[0076] 5. Analysis results: Calculation results R xy [k] is a new sequence that describes the similarity of the original signal with different shift amounts. The magnitude and sign of the value provide clues to the strength of the correlation and the direction of the phase shift.

[0077] Through cross-correlation, the phase difference between signals can be identified, the repetitive pattern of signals can be detected, or a specific signal pattern can be found in noise. It is a very commonly used tool in the fields of signal processing and time series analysis. In this embodiment, the index of the point greater than a certain threshold (0.7 times the maximum value in the sequence in this patent) in the cross-correlation sequence after cross-correlation calculation is mainly obtained.

[0078] A4. Perform feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiply the mutual correlation feature mask by the first mutual correlation feature, adjust the weight of the first mutual correlation feature, and output a second mutual correlation feature;

[0079] A5. Extract the second cross-correlation feature and the original signal feature, and output the key feature;

[0080] A6. Based on this key feature, predict the arrival time of underwater weak signals.

[0081] In this embodiment, first, the received acoustic signal data sequence and its corresponding original transmitted signal sequence are obtained. Secondly, feature extraction is performed on the acoustic signal data sequence to obtain the original signal feature. The feature extraction process is intended to extract useful information from the signal and provide the necessary basis for subsequent analysis. Next, the acoustic signal data sequence and the original transmitted signal sequence are cross-correlated to calculate the cross-correlation parameter. Cross-correlation analysis can quantify the similarity between two signals. By calculating the cross-correlation parameter, the relative time delay of the signal can be determined. The cross-correlation parameter is further feature extracted to obtain the first cross-correlation feature. This step is intended to further refine the key information in the cross-correlation parameter. Subsequently, feature mask extraction is performed on the first cross-correlation feature to obtain a cross-correlation feature mask. The feature mask can be regarded as the screening or weighting of the feature. By applying the mask, the important part of the feature can be highlighted. The cross-correlation feature mask is multiplied by the first cross-correlation feature to adjust the weight of the first cross-correlation feature, thereby outputting the second cross-correlation feature. This step weights the feature by masking so that the feature is more in line with the actual application requirements. Next, feature extraction is performed on the second cross-correlation feature and the original signal feature to output the key feature. The key feature is the most valuable information obtained after processing, which can more accurately reflect the nature of the signal. Finally, the key feature contains important information about the signal, which can be used to more accurately predict the arrival time of the signal. Through multiple rounds of feature extraction and processing, the useful information in the signal can be fully mined, thereby improving the accuracy of the prediction. In a complex underwater environment, the signal may be subject to various interferences. This method can effectively suppress interference and maintain the stability of the prediction through means such as feature masks. This method is suitable for underwater communications, underwater detection and other fields, and can significantly improve the efficiency and accuracy of signal processing, providing strong technical support for related work.

[0082] In this embodiment, the acoustic signal data sequence is one-dimensional acoustic data. At present, the input requirements of most neural network models are two-dimensional data, such as images. This method uses a one-dimensional acoustic signal sequence as input, and can directly analyze and process the received one-dimensional acoustic signal sequence without performing additional conversion operations, such as short-time Fourier transform, etc. These methods of converting one-dimensional data into two-dimensional data may cause the loss of certain features in the data. Therefore, this method can retain the features of the original signal to the maximum extent, thereby improving the accuracy and effectiveness of signal processing.

[0083] In this embodiment, as a preferred embodiment, please refer to Figure 2 , the method steps are as follows:

[0084] B1. Acquire an acoustic signal data sequence and an original transmission signal sequence corresponding to the sequence;

[0085] B2. Extracting features from the acoustic signal data sequence to obtain original signal features;

[0086] B3. Perform cross-correlation calculation on the acoustic signal data sequence and the transmission signal sequence to obtain cross-correlation parameters; perform feature extraction on the feature parameters to obtain a first cross-correlation feature;

[0087] B4. Set N rounds of feature extraction, and perform the following steps in each round of feature extraction:

[0088] B41, performing feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiplying the mutual correlation feature mask by the first mutual correlation feature, adjusting the weight of the first mutual correlation feature, and outputting a second mutual correlation feature;

[0089] B42, using the original signal features and the second cross-correlation features of this round as data input for the next round of original signal feature extraction, and performing the next round of feature extraction; using the first cross-correlation features of this round as data input for the next round of first cross-correlation feature extraction, and performing the next round of feature extraction;

[0090] B5, loop the above steps until the set number of feature extraction rounds is completed, and output the final second cross-correlation feature and the final original signal feature;

[0091] B6. Extract the final second cross-correlation feature and the final original signal feature, and output the key feature;

[0092] B7. Based on this key feature, predict the arrival time of underwater weak signals.

[0093] In this embodiment, the original signal features contain the basic information and initial features of the signal. By using them as the input of the next round of feature extraction, it can be ensured that these basic features continue to play a role in the entire iterative process, providing a stable basis for the final feature extraction. The newly output feature (the second cross-correlation feature) is the optimized feature after the feature mask adjustment in the current round, which contains the latest understanding and refinement of the signal features in the current round. Using it as the input of the next round of feature extraction, these optimized feature information can be further integrated into the subsequent feature extraction process to achieve the supplementation and optimization of feature information, so that the final feature is more comprehensive and accurate. Through the accumulation of feature information, the expressive power of the feature can be enhanced, so that the final feature can more accurately describe the characteristics and essence of the signal. This enhanced expressive power helps to improve the performance of subsequent feature-based signal analysis and processing tasks.

[0094] Embodiment 2

[0095] See also Figure 3A neural network model training method for underwater weak signal arrival time prediction, wherein the neural network is a dual-channel network model, including a cross-correlation feature extraction channel, an original signal feature extraction channel, an interaction module, an attention mechanism module, and a fully connected layer; the method includes:

[0096] C1. Obtain a sample acoustic signal data sequence, an original transmitted signal sequence corresponding to the sequence, and a true label corresponding to the sequence;

[0097] C2. For the acoustic signal data sequence, input the original signal feature extraction channel to perform feature extraction to obtain the original signal features;

[0098] C3. For the acoustic signal data sequence, perform cross-correlation calculation with the transmission signal sequence to obtain cross-correlation parameters; for the feature parameters, input the cross-correlation feature extraction channel to perform feature extraction to obtain a first cross-correlation feature;

[0099] C4. Input the first mutual correlation feature into the interactive module, perform feature mask extraction, and obtain a mutual correlation feature mask; multiply the mutual correlation feature mask by the first mutual correlation feature, adjust the weight of the first mutual correlation feature, and output a second mutual correlation feature;

[0100] C5: For the second cross-correlation feature and the original signal feature, input the attention mechanism module to extract the features and output the key features;

[0101] C6, based on this key feature, inputs into the fully connected layer to predict the arrival time of underwater weak signals;

[0102] C7. Calculate the loss value based on the arrival time and the true label; and

[0103] C8. Training the neural network model based on the loss value.

[0104] In this embodiment, in order to better use the method described in one of the embodiments, the present application proposes a neural network model training method for underwater weak signal arrival time prediction. The dual-channel network model includes two feature extraction channels, namely, a cross-correlation feature extraction channel and an original signal feature extraction channel. This dual-channel design can extract signal features from different angles. The cross-correlation feature extraction channel focuses on the time correlation feature of the signal, while the original signal feature extraction channel extracts the basic features of the signal. The interaction module is used to extract and adjust the weight of the first cross-correlation feature by feature mask, and output the second cross-correlation feature by multiplying the feature mask with the first cross-correlation feature. This process can highlight the important part of the feature and suppress unimportant information. The attention mechanism module further extracts the second cross-correlation feature and the original signal feature, and outputs the key feature. The attention mechanism can make the model pay more attention to the important information in the feature, improve the feature expression ability and the prediction performance of the model. Based on the key features, the fully connected layer predicts the signal arrival time. The fully connected layer can comprehensively analyze the features and output the final prediction result. The dual-channel network model extracts signal features from different angles and can capture the characteristics and laws of the signal more comprehensively. Combining the interaction module and the attention mechanism module to optimize and focus on features enables the model to understand and analyze signals more accurately.

[0105] In this embodiment, obtaining a sample sound signal data sequence, an original transmission signal sequence corresponding to the sequence, and a true label corresponding to the sequence includes: collecting at least two types of background noise and at least two types of target sound signals; for each target sound signal, selecting an adding time from the target sound signal, synthesizing the target sound signal and any background noise at the adding time, and using the synthesis result as the sample sound signal data sequence; using the target sound signal as the original transmission signal sequence; using the adding time as the true label of the sample sound signal data sequence; repeating the above synthesis steps until the data set is constructed.

[0106] In this embodiment, background noise is various interference sounds existing in the underwater environment, such as water flow sound, marine life sound, etc. Collecting different types of background noise can simulate the diversity of the underwater environment and provide rich background conditions for sample data. Different types of background noise have different frequency characteristics, intensity distribution, etc., which can help the model learn the ability to identify target signals in different noise environments. The target sound signal is a signal that needs to be detected and predicted for arrival time, such as the sound signal of an underwater target. Different types of target standard signals may have different characteristics, such as different frequency components, signal morphology, etc., which can improve the diversity and representativeness of the data set and enable the model to adapt to signals of different target types. For each target sound signal, an addition time is selected from the target sound signal. The addition time is the time point when the target sound signal is synthesized with the background noise. Selecting different addition times can simulate the situation where the target signal appears at different times and increase the time diversity of the data set. At the selected addition time, the target sound signal is synthesized with any background noise to obtain a sample sound signal data sequence. The synthesis process needs to consider factors such as signal superposition and noise interference to ensure that the synthesized sample sound signal data sequence can truly reflect the superposition of the target signal and background noise in the underwater environment. The synthesized sample acoustic signal data sequence contains the characteristic information of the target signal and background noise, providing basic data for subsequent feature extraction and model training. Based on the target acoustic signal as the original emission signal sequence. The original emission signal sequence is the signal form of the target acoustic signal when it is not interfered by noise, and it contains the original characteristic information of the target signal. In the process of model training, the original emission signal sequence can be used as a reference signal to help the model better understand and learn the characteristics of the target signal. The true label makes the sample acoustic signal data sequence based on the added time as the true label. The true label is the actual arrival time of the target acoustic signal in the synthesized sample acoustic signal data sequence, which is an important basis for model training and prediction. By using the added time as the true label, it can be ensured that the model can accurately learn the arrival time characteristics of the target signal during the training process, thereby improving the prediction performance of the model. This method constructs an underwater acoustic signal dataset by synthesizing the target acoustic signal and background noise, and provides the original emission signal sequence and true label for each sample, which can provide high-quality and diverse data support for the training of the underwater signal processing model, and help improve the prediction performance and application effect of the model.

[0107] In this embodiment, the data set includes at least three classic low signal-to-noise ratio data sets of -10dB, -15dB, and -20dB. When constructing a data set for underwater weak signal arrival time prediction, at least three classic low signal-to-noise ratio data sets are included. These data sets can provide abundant signal samples in low signal-to-noise ratio environments, helping the model to better learn and adapt to signal characteristics under low signal-to-noise ratio conditions. Under the signal-to-noise ratio condition of -10dB, the intensity of the signal and the noise are close, and the interference of the noise on the signal is more obvious. This data set can simulate the situation of the signal in a medium noise environment, and is used to train the signal detection and recognition capabilities of the model in a medium noise environment. Under the signal-to-noise ratio condition of -15dB, the intensity of the noise is further increased, and the characteristics of the signal are more difficult to identify. This data set can simulate the situation of the signal in a higher noise environment, and is used to train the robustness of the model in a high noise environment. Under the signal-to-noise ratio condition of -20dB, the intensity of the noise is much greater than the signal, and the signal is almost submerged by the noise. This data set can simulate the situation of the signal in an extreme noise environment, and is used to train the extreme performance of the model in an extreme noise environment. These three datasets cover noise environments ranging from low to high, and can comprehensively train the model's signal processing capabilities under different signal-to-noise ratio conditions.

[0108] Embodiment 3

[0109] A neural network model for underwater weak signal arrival time prediction, based on the method described in Embodiment 1 and Embodiment 2, wherein the neural network is a dual-channel network model, comprising:

[0110] A cross-correlation feature extraction channel is configured to perform cross-correlation calculation on the acoustic signal data sequence and the transmission signal sequence to obtain a cross-correlation parameter; perform feature extraction on the feature parameter to obtain a first cross-correlation feature;

[0111] An original signal feature extraction channel, which is configured to perform feature extraction on the acoustic signal data sequence to obtain original signal features;

[0112] An interaction module is configured to perform feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiply the first mutual correlation feature by the mutual correlation feature mask, adjust the weight of the first mutual correlation feature, and output a second mutual correlation feature;

[0113] An attention mechanism module is configured to extract features from the second cross-correlation feature and the original signal feature and output key features;

[0114] The fully connected layer is configured to predict the arrival time of underwater weak signals based on the key features.

[0115] In this embodiment, in order to better use the methods described in Embodiment 1 and Embodiment 2, the present application proposes a neural network model for predicting the arrival time of underwater weak signals, each part of which corresponds to each step of the above method. The specific principles of each step have been explained above and will not be elaborated here.

[0116] Embodiment 4

[0117] In order to better explain the present invention, the following will be combined with Examples 1 to 3 to further explain the present invention, as follows: Figures 4 to 11 This embodiment provides a deep learning model for detecting arrival time of underwater weak signals by a river acoustic tomography system, comprising the following steps:

[0118] Training data preparation: Use two types of background noise (Gaussian white noise and river environment noise collected in the field) and two types of target sound signals (real emission signals and simulated emission signals), select a suitable set of starting positions to generate a total of three experimental data sets (1. Gaussian white noise + real emission signal, 2. real noise + simulated signal, 3. real noise + real emission signal). At the same time, each of them synthesizes three classic low signal-to-noise ratio (-10dB, -15dB, -20dB) data sets. For example, Figure 4 It is an analog transmitted sound signal composed of a 7th-order m-sequence modulated by BPSK. Figure 5 The three pictures from top to bottom are the three signal data synthesized under -20dB signal-to-noise ratio. (It should be noted that Figure 4 and Figure 5 They are all visualizations of one-dimensional signal sequences, in order to help view the data more intuitively, such as a one-dimensional sequence:

[0119] [0,1,2,3,4,5,6,7,8,9,10], after visualization, is a line segment from the coordinate point (0,0) to (1,1) to (10,10) in a coordinate system. However, the input of this model is a one-dimensional signal sequence, not Figure 4 , Figure 5 Schematic diagram) The first figure is a synthetic sound signal of simulated signal + real noise; the second figure is a synthetic sound signal of real signal + real noise; the third figure is a synthetic sound signal of real signal + Gaussian white noise. At the same time, the time when the target sound signal (real signal and simulated signal) is added is recorded as the real label corresponding to the data. The label is the starting position point of the signal. A data and the corresponding label constitute a data sample. A group of numbers with equal interval length are selected as the starting position of the signal, that is, the label value. Each data set is synthesized into a total of 4000 samples with a sequence length of 20000 to form a data set; the data set is randomly shuffled and divided into training set, test set and validation set according to 8:1:1 to complete the preparation of training data;

[0120] Model building: Build a dual-channel network model for a one-dimensional sound signal sequence (such as Figure 6 ), the input is a one-dimensional sound signal data sequence, one of the channels will first perform cross-correlation calculation on the received sound signal after the signal is input; each channel consists of 5 residual blocks, each residual block consists of three one-dimensional convolutional layers, the first convolutional layer kernel size is 1, the remaining two convolutional layers kernel size is 15, and the step size is 1. The purpose of deploying the first convolutional layer is mainly to increase the number of channels to meet the channel consistency of the jump connection part. The remaining two convolutional layers are followed by batch normalization and activation unit blocks (relu and linear activation). The number of output channels of each residual block is 8, 16, 32, 64, and 128 respectively. Deploy interaction modules between two channels (such as Figure 7 ), a pooling layer + one-dimensional convolution block + Sigmoid activation function is used to generate a mask for the feature, so as to extract the important features in the cross-correlation channel and combine them with the original signal channel features, and then pass through the attention mechanism module to make the model pay more attention to the effective features and discard the invalid features; the attention mechanism module consists of channel attention and spatial attention (such as Figure 8 ). It should be noted that the input data of this attention module is one-dimensional, which is different from the common image data (two-dimensional). The initial design of this module only includes spatial attention. Through continuous adjustment and modification in experiments, channel attention and spatial attention are formed (such as Figure 8 The design combined with the one-dimensional convolution block (shown in Figure 2) makes it better adapt to the processing requirements of one-dimensional data. Finally, the fully connected block is connected to output the number of points at the starting point of the predicted sound signal. It should be noted that the convolution layer and pooling layer in the model are both one-dimensional convolution and one-dimensional pooling, so there is no need to perform additional transformation on the input signal (which may lose important features).

[0121] Model training: directly input a one-dimensional sound signal sequence data set into the model without additional processing of the one-dimensional sound signal sequence, such as short-time Fourier transform, etc.; the loss function is designed to be the mean square error between the predicted starting point and the label Mean absolute error As an evaluation index. i is the tag value, is the output result. The smaller the MSE and MAE, the closer the output result is to the label value. The goal of model training is to find the model parameters that minimize MSE and MAE. We use data to train the neural network in batches, and each round of training is divided into 50 times; we use the Adam optimizer and set the learning rate schedule. When the error value of the validation set does not decrease for 5 consecutive times, the learning rate will be reduced; at the same time, we use the validation set to control the number of model training rounds, and stop training when the error value of the validation set does not decrease for many consecutive times; finally, we use the test set to test whether the model results achieve the expected effect. If the error of the result on the test set is similar to that of the training set and the validation set, it means that the model has strong generalization ability and there is no overfitting phenomenon, and the training is completed;

[0122] Model call: directly input the received one-dimensional sound signal sequence to be detected into the model (such as Figure 4 ), the neural network model can directly output the predicted starting position points of the target sound signal. Fig. 9 As shown in the figure, under the condition of -15dB signal-to-noise ratio, the starting position predicted by the model almost coincides with the label value, while the starting position of the signal detected by the traditional matched filtering method is far from the label value position.

[0123] The advantages of the present invention are as follows: Currently, most neural network models require two-dimensional data as input, such as pictures, etc. The neural network model used in the method of the present invention is for one-dimensional sound signal sequences, and the received one-dimensional sound signal sequences can be directly input into the model without performing additional operations such as short-time Fourier transform on the one-dimensional sequence. This type of method of converting one-dimensional data into two-dimensional data will inevitably lose some features in the data, so the method of this model can protect the original signal features to the greatest extent.

[0124] Compared with the existing detection methods of river acoustic tomography flow measurement systems, such as matched filtering and some simple neural model applications, the neural network model used in the method of the present invention uses a dual-channel model and an interaction module and an attention mechanism module deployed between channels, which can better capture the features of the received sound signal. It can complete the starting point detection task of the target sound signal under low signal-to-noise ratio (less than -10dB). Fig.10As shown, the detection results of the neural network model provided by the embodiment of the present invention and some other methods on the real signal + real noise data set are compared. It can be seen that when the signal-to-noise ratio is lower than -10dB, the performance of the traditional detection technology matched filtering method deteriorates sharply. However, the detection results obtained by classic deep learning methods such as CNN have smaller deviations than traditional technologies. The neural network model proposed in the patent of this invention achieved the best results. Compared with the matched filtering method, the average absolute error was reduced by 81.8%, 84.5%, and 82.1% at -10dB, -15dB, and -20dB respectively. Compared with other classic deep learning-based technologies, the model we proposed has also achieved different degrees of gain in detection error. Fig.11 The figure shows the comparison of the detection results of the neural network model with some modules added or removed on the real signal + Gaussian white noise data set under the condition of -15dB signal-to-noise ratio. It can be seen that compared with the matched filtering method, the dual-channel network model alone can reduce the mean absolute error of the detection results by 8.7%. It can be observed that the mean absolute error is reduced to a certain extent when the interaction module is deployed between the original signal channel and the cross-correlation channel and when the attention module is introduced into the model without deploying the interaction module. Finally, the mean absolute error of the model prediction is the lowest when the interaction module and the attention module are deployed in the network at the same time. These observations verify the effectiveness and stability of the proposed dual-channel network architecture, interaction module, and attention module in detecting the arrival time of acoustic tomography signals under low signal-to-noise ratio conditions.

[0125] The above descriptions are only some embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any equivalent device or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for predicting arrival time of underwater weak signals, characterized in that: The method comprises: Acquire a received acoustic signal data sequence and an original transmitted signal sequence corresponding to the sequence; Performing feature extraction on the acoustic signal data sequence to obtain original signal features; For the acoustic signal data sequence, a cross-correlation calculation is performed with the transmission signal sequence to obtain a cross-correlation parameter; for the characteristic parameter, a feature extraction is performed to obtain a first cross-correlation feature; Performing feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiplying the mutual correlation feature mask by the first mutual correlation feature, adjusting the weight of the first mutual correlation feature, and outputting a second mutual correlation feature; Based on the second cross-correlation feature and the original signal feature, feature extraction is performed to output key features; Based on this key feature, the arrival time of underwater weak signals is predicted.

2. The method for predicting arrival time of underwater weak signals according to claim 1, characterized in that: The acoustic signal data sequence is one-dimensional acoustic data.

3. The method for predicting arrival time of underwater weak signals according to claim 1, characterized in that: The method comprises: Acquire a received acoustic signal data sequence and an original transmitted signal sequence corresponding to the sequence; Performing feature extraction on the acoustic signal data sequence to obtain original signal features; For the acoustic signal data sequence, a cross-correlation calculation is performed with the transmission signal sequence to obtain a cross-correlation parameter; for the characteristic parameter, a feature extraction is performed to obtain a first cross-correlation feature; Set N rounds of feature extraction, and perform the following steps in each round of feature extraction: Performing feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiplying the mutual correlation feature mask by the first mutual correlation feature, adjusting the weight of the first mutual correlation feature, and outputting a second mutual correlation feature; The original signal features and the second cross-correlation features of this round are used as data input for the next round of original signal feature extraction, and the next round of feature extraction is performed; the first cross-correlation features of this round are used as data input for the next round of first cross-correlation feature extraction, and the next round of feature extraction is performed; The above steps are repeated until the set number of feature extraction rounds is completed, and the final second cross-correlation feature and the final original signal feature are output; For the final second cross-correlation feature and the final original signal feature, feature extraction is performed to output key features; Based on this key feature, the arrival time of underwater weak signals is predicted.

4. A neural network model training method for underwater weak signal arrival time prediction, characterized in that: The neural network is a dual-channel network model, including a cross-correlation feature extraction channel, an original signal feature extraction channel, an interaction module, an attention mechanism module, and a fully connected layer; the method includes: Obtaining a sample acoustic signal data sequence, an original transmitted signal sequence corresponding to the sequence, and a true label corresponding to the sequence; For the acoustic signal data sequence, input the original signal feature extraction channel to perform feature extraction to obtain the original signal features; For the acoustic signal data sequence, a cross-correlation calculation is performed with the transmission signal sequence to obtain a cross-correlation parameter; for the feature parameter, the cross-correlation feature extraction channel is input to perform feature extraction to obtain a first cross-correlation feature; The first mutual correlation feature is input into an interactive module to extract a feature mask to obtain a mutual correlation feature mask; the mutual correlation feature mask is multiplied by the first mutual correlation feature to adjust the weight of the first mutual correlation feature and output a second mutual correlation feature; The second cross-correlation feature and the original signal feature are input into the attention mechanism module for feature extraction and key features are output; Based on this key feature, the fully connected layer is input to predict the arrival time of underwater weak signals; Based on the arrival time and the true label, calculate the loss value; and, The neural network model is trained based on the loss value.

5. A neural network model training method for underwater weak signal arrival time prediction as claimed in claim 4, characterized in that: The acoustic signal data sequence is one-dimensional acoustic data.

6. A neural network model training method for underwater weak signal arrival time prediction as claimed in claim 4, characterized in that: The cross-correlation feature extraction channel and the original signal feature extraction channel both include N residual blocks, and the method includes: Obtaining a sample acoustic signal data sequence, an original transmitted signal sequence corresponding to the sequence, and a true label corresponding to the sequence; For the acoustic signal data sequence, the first residual block of the original signal feature extraction channel is input to perform feature extraction to obtain the original signal feature; For the acoustic signal data sequence, a cross-correlation calculation is performed with the transmission signal sequence to obtain a cross-correlation parameter; for the feature parameter, a first residual block of the cross-correlation feature extraction channel is input to perform feature extraction to obtain a first cross-correlation feature; Set N rounds of feature extraction, and perform the following steps in each round of feature extraction: Performing feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiplying the mutual correlation feature mask by the first mutual correlation feature, adjusting the weight of the first mutual correlation feature, and outputting a second mutual correlation feature; The original signal features and the second cross-correlation features of this round are used as the data for the next round of original signal feature extraction, and are input into the next residual block of the original signal feature extraction channel to perform the next round of feature extraction; the first cross-correlation features of this round are used as the data for the next round of first cross-correlation feature extraction, and are input into the next residual block of the cross-correlation feature extraction channel to perform the next round of feature extraction; Repeat the above steps until the set number of feature extraction rounds is completed, and output the final second cross-correlation feature and the final original signal feature; The final second cross-correlation feature and the final original signal feature are input into the attention mechanism module for feature extraction and key features are output; Based on this key feature, the arrival time of underwater weak signals is predicted; Based on the arrival time and the true label, calculate the loss value; and, The neural network model is trained based on the loss value.

7. A neural network model training method for underwater weak signal arrival time prediction as claimed in claim 4, characterized in that: Obtaining a sample acoustic signal data sequence, an original emission signal sequence corresponding to the sequence, and a true label corresponding to the sequence, including: collecting at least two types of background noises and at least two types of target sound signals; For each target sound signal, selecting an adding time from the target sound signal, synthesizing the target sound signal and any background noise at the adding time, and using the synthesized result as a sample sound signal data sequence; Based on the target generated signal as the original transmission signal sequence; Based on the adding time as the true label of the sample acoustic signal data sequence; Repeat the above synthesis steps until the data set is constructed.

8. A neural network model training method for underwater weak signal arrival time prediction as claimed in claim 7, characterized in that: The data set includes at least three classic low signal-to-noise ratio data sets of -10dB, -15dB and -20dB.

9. A neural network model for predicting arrival time of underwater weak signals, characterized in that: Based on the method described in any one of claims 1 to 8, the neural network is a dual-channel network model, comprising: A cross-correlation feature extraction channel is configured to perform cross-correlation calculation on the acoustic signal data sequence and the transmission signal sequence to obtain a cross-correlation parameter; perform feature extraction on the feature parameter to obtain a first cross-correlation feature; An original signal feature extraction channel, which is configured to perform feature extraction on the acoustic signal data sequence to obtain original signal features; An interaction module is configured to perform feature mask extraction on the first mutual correlation feature to obtain a mutual correlation feature mask; multiply the first mutual correlation feature by the mutual correlation feature mask, adjust the weight of the first mutual correlation feature, and output a second mutual correlation feature; An attention mechanism module is configured to extract features from the second cross-correlation feature and the original signal feature and output key features; The fully connected layer is configured to predict the arrival time of underwater weak signals based on the key features.

10. A neural network model for predicting arrival time of underwater weak signals as claimed in claim 9, characterized in that: The cross-correlation feature extraction channel and the original signal feature extraction channel both contain N residual blocks for feature extraction; the residual blocks in each channel are linked by jumps, and each residual block is connected to the interaction module.

Images