Underwater acoustic source distance estimation method based on normal wave phase and multi-task neural network

Abstract

The present application relates to a kind of underwater sound source distance estimation based on normal wave phase and multi-task neural network, first in certain geoacoustic parameter range, simulation sound pressure phase is calculated, and simulation data and measured data are normalized;Then the sound pressure phase calculated is used as training data, horizontal distance, seabed bottom quality sound velocity and mode information are used as supervisory signal to train a multi-task model;The model is preliminarily tested using test data to verify the effectiveness of the method;Finally, a small amount of normalized measured complex sound pressure is used to fine-tune the model, and it is tested again on the remaining measured data.The robustness of distance prediction ability is improved;Using transfer learning to alleviate the distribution difference between simulation data and measured data, experiments show that only a small amount of measured data can greatly improve the distance prediction performance of the model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of ship and ocean engineering, and relates to an underwater sound source distance estimation method based on normal wave phase and a multi-task neural network. BACKGROUND

[0002] The document "A multi-task learning convolutional neural network for source localization in deep ocean. The Journal of the Acoustical Society of America, 148 2, 873." discloses an underwater sound source positioning method based on a convolutional neural network and multi-task learning. The method simulates the complex sound pressure received by the array at different distances and depths of the sound source in various underwater acoustic environments by using sound field propagation software to calculate the cross-spectral density matrix, uses the cross-spectral density matrix as the input quantity, and uses the distance and depth as the label to train the deep residual network. The author constructs a multi-branch network structure, so that the deep neural network can learn two labels at the same time. The method performs well in real environment experiments. However, the method requires a large amount of simulation data to alleviate the environment parameter mismatch problem during training, and has a very high demand for computing resources; at the same time, only the task label is used as the multi-task learning supervision signal, and other information in the sound field is ignored; in addition, the difference between the simulated sound field and the real sound field is ignored, and the method cannot guarantee the performance in the real environment. In summary, the method has limited application value in practical engineering. SUMMARY

[0003] Technical problems to be solved

[0004] In order to avoid the shortcomings of the prior art, the application provides an underwater sound source distance estimation method based on normal wave phase and a multi-task neural network, which provides a deep learning underwater sound source positioning method with low computing resource demand, robustness and environmental adaptability during training.

[0005] Technical scheme

[0006] An underwater sound source distance estimation method based on normal wave phase and a multi-task neural network, characterized in that the steps are as follows:

[0007] Step 1: The phase calculation formula of the adiabatic normal wave is as follows:

[0008]

[0009] Where: Φ m (f;r) represents the sound pressure phase, k m is the real part of the horizontal wave number;

[0010] The broadband sound pressure phase calculated by the above formula is:

[0011] Φ M = [Φ M (1),..., Φ M (f),..., Φ M (F)]

[0012] Normalization is performed by the following formula:

[0013]

[0014]

[0015] where Re(·) and Im(-) represent the real part and the imaginary part of a complex vector, respectively;

[0016] For the measured sound field, the broadband complex sound pressure is used for representation:

[0017] p = [p(1),..., p(f),..., p(F)]

[0018] Normalization is performed by the following formula:

[0019]

[0020]

[0021] The labels of the regression task are normalized to have a uniform scale, and the labels are normalized by the following formula:

[0022]

[0023] where R i represents the distance label, S i represents the substrate sound velocity label, M i represents the number of modal combinations, and n represents the number of training samples; finally, the normalized simulated sound field phase, the measured complex sound pressure, and the distance, depth, and substrate sound velocity labels of the corresponding simulated sound field are obtained;

[0024] Step 2: Construct a multi-task neural network model, and the network adopts a stack of residual blocks: the first layer is a 64-channel convolution layer Conv with a size of 3x3; the main connection part of each residual block adopts two convolution layers with a kernel size of 3x3, followed by a rectified linear unit ReLu activation function and a batch normalization layer Batch Normalization; the channel number of the first two residual blocks is 64, and that of the last two is 128; the skip connection part directly adds the input data and the main connection mapping data for the first, second and fourth residual blocks, and the third residual block first maps the data using a 128-channel convolution layer, and then adds the main connection mapping data; the data is encoded and flattened by the above neural network to obtain a 1024-dimensional deep feature, and then three linear Fc layers are used to obtain a 1-dimensional prediction result, respectively corresponding to the distance, the bottom velocity and the mode;

[0025] The multi-task neural network model is trained in the format of the data calculated in step 1, and the loss function of each prediction result is mean square error MSE during training, and the task weight is set to 1;

[0026] The final loss function is:

[0027]

[0028] Wherein: and are the predictions of the distance, the bottom velocity and the mode by the neural network, respectively.

[0029] Step 3: Perform transfer learning on the obtained neural network, and discard the prediction linear layers of the bottom velocity and the mode during transfer learning, and only fine-tune the distance prediction linear layer.

[0030] Step 4: Use the neural network after transfer learning in step 3 to obtain the distance estimation result.

[0031] The transfer learning of the neural network in step 3 first trains the model using the simulated sound pressure phase, and then fine-tunes the model using the measured sound field complex sound pressure, so as to alleviate the distribution difference between the simulation data and the measured data.

[0032] Beneficial effects

[0033] The application provides an underwater sound source distance estimation based on normal wave phase and a multi-task neural network.

[0034] The application uses simulated normal wave phase to pre-train the neural network instead of complex sound pressure, so that the calculation amount of the pre-training stage is reduced without considering the sound source depth and the receiving depth; the distance prediction is assisted by predicting geoacoustic parameters and signal parameters, so that the robustness of the distance prediction capability is improved; the distribution difference between the simulation data and the measured data is relieved by using transfer learning, and experiments prove that only a small amount of measured data is needed to greatly improve the distance prediction performance of the model. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The multi-task neural network structure is used.

[0036] Figure 2 The transfer learning process is used.

[0037] Figure 3 The topography and the sound velocity profile of the sea trial are used.

[0038] Figure 4 The prediction result comparison of the pre-training model is used.

[0039] Figure 5 The prediction result comparison of the transfer model is used. DETAILED DESCRIPTION

[0040] The application will be further described in combination with embodiments and drawings:

[0041] The application is realized by the following technical solutions:

[0042] The underwater sound source distance estimation based on normal wave phase and the multi-task neural network comprises:

[0043] (1) the normal wave phase is used instead of the cross-spectral density matrix to reduce the number of training samples generated during training;

[0044] (2) the task label and the sound field description are used as the supervision signal of the multi-task learning;

[0045] (3) the transfer learning is used to relieve the distribution difference between the simulation data and the measured data;

[0046] The present application mainly comprises the following contents:

[0047] According to the adiabatic normal mode theory, the complex sound pressure of the sound field can be represented by the following formula:

[0048]

[0049]

[0050]

[0051] where p(f;z s ,z r ,r) represents the complex sound pressure at the frequency point f, which is determined by the receiver depth z s , the sound source depth z r and the horizontal distance r. M represents the modal order present in the signal. a m (f;z s ,z r ,r) represents the sound pressure amplitude. Φ m (f;r) represents the sound pressure phase. Q is a constant related to the medium density at the sound source. ψ m is the depth-dependent equation for mode m. k m and β m are the real and imaginary parts of the horizontal wave number, respectively. It can be seen from the above expression that the sound pressure phase does not depend on the depth and sound source depth information, and contains information related to the horizontal distance r. Therefore, the sound pressure phase of the simulation can avoid the interference of the depth information during the neural network training, reduce the calculation amount, and at the same time alleviate the depth mismatch problem. Considering the sound pressure phase of multiple modes is represented by the following formula:

[0052] Φ M (f) = ∑ m∈M exp(jΦ m (f;r))

[0053] where M represents the modal order contained in the signal.

[0054] The multi-task learning paradigm is given by the following formula:

[0055]

[0056] where N represents the number of tasks, α i is the weight of the i-th task, L(·;·) is the defined loss function, ψ(·;θ s ) is a function defined by the task-shared parameter θ s , is a function defined by the task-specific parameter , x is the input data, y iis the supervision signal of the i-th task. For the source phase, it is described by both the horizontal distance r and the mode M simultaneously, and it is also related to the seabed sound speed. Therefore, in this task, we consider building multi-task learning on both the horizontal distance, the mode and the seabed sound speed to make full use of the supervision signal of the sound field. For the distance and the seabed sound speed, we build regression tasks to learn continuous variables. For the mode, considering that the data corresponding to different mode combinations are still related in distance, if we classify the mode, it may exacerbate the distribution of samples at the same distance in the feature space under different modes, leading to learning unreasonable data manifold. Therefore, we still use the classification task for the mode. Specifically, we directly continuously number different mode combinations, and use the number as the regression label.

[0057] Fine-tuning based transfer learning first trains a neural network using a large amount of source domain data, and then fine-tunes the neural network parameters using a small amount of target domain data and a lower learning rate. In this paper, we consider a large amount of simulation data as the source domain data and a small amount of measured data as the target domain data. First, we train the model using the simulated sound pressure phase, and then fine-tune the model using the measured sound field complex sound pressure to alleviate the distribution difference between the simulation data and the measured data. DETAILED EMBODIMENTS

[0059] Step one: the broadband sound pressure phase calculated by the adiabatic normal mode formula given in the above is represented by the following formula:

[0060] Φ M = [Φ M (1),..., Φ M (f),..., Φ M (F)]

[0061] Referring to the energy distribution of the actual signal, only the highest 4th order normal mode and the highest 2nd order normal mode combination are considered.

[0062] Normalization is performed using the following formula:

[0063]

[0064]

[0065] where Re(·) and Im(·) represent the real part and the imaginary part of the complex vector, respectively.

[0066] For the measured sound field, the broadband complex sound pressure is still used to represent it, as shown in the following formula:

[0067] p = [p(1),..., p(f),..., p(F)]

[0068] Normalization is performed using the following formula:

[0069]

[0070]

[0071] In order to make the labels of the regression task uniform in scale, the labels are normalized by the following formula:

[0072]

[0073] where R i represents the distance label, S i represents the bottom sound speed label, M i represents the number of modal combinations, and n represents the number of training samples. Finally, the normalized simulated sound field phase, measured complex sound pressure, and distance, depth, and bottom sound speed labels corresponding to the simulated sound field are obtained.

[0074] Step two: according to the format of the data calculated in step one, the multi-task neural network model is constructed and trained according to the following formula: Figure 1 The network mainly uses the stacking of residual blocks to construct the main part of the network. The first layer is a 64-channel convolution layer (Conv) with a size of 3x3; the main connection part of each residual block then uses two convolution layers with a convolution kernel size of 3x3, followed by a rectified linear unit (ReLu) activation function and a batch normalization layer (BatchNormalization). The channel number of the first two residual blocks is 64, and that of the last two is 128; the skip connection part of the first, second, and fourth residual blocks directly adds the input data and the main connection mapping data, and the third residual block first maps the data using an 128-channel convolution layer, and then adds the main connection mapping data. The data is encoded and flattened by the above neural network to obtain a 1024-dimensional depth feature, and then passed through three linear (Fc) layers to obtain a 1-dimensional prediction result, corresponding to distance, bottom sound speed, and mode respectively. The loss function of each prediction result is the mean square error (MSE) during training. Considering the equal importance of the three tasks, the task weight is set to 1. The final loss function is given by the lower right formula:

[0075]

[0076] where and are the predictions of the neural network for distance, bottom sound speed, and mode, respectively.

[0077] After training the neural network using the above steps, the pre-trained model is tested using test data collected in a weak distance-dependent waveguide environment, Figure 3The terrain and sound velocity profile are shown. 12 groups of linear frequency modulation signals are respectively emitted at 7 different distances, and the frequency is 20-200Hz. Due to the influence of noise, only 100-200Hz is used as the target frequency band. The test standard is calculated by using the absolute mean error MAPE and the reliable prediction probability PCL. The two standards are as follows:

[0078]

[0079]

[0080] Where N is the number of samples, R i , are the actual distance and the predicted distance respectively.

[0081] Step four: The obtained neural network is migrated according to the Figure 2 flow. One sample is randomly sampled from the 12 signals at 7 distances as the training sample of the migration learning, and the remaining 11 samples are used for testing. The entire neural network is used as the task sharing parameter. Since the measured data does not contain the mode and the bottom speed label, the prediction linear layer of the bottom speed and the mode is discarded during the migration learning, and only the distance prediction linear layer is fine-tuned.

[0082] Step five: The fine-tuned network is tested again using the measured data to obtain the final distance estimation result.

[0083] In order to better illustrate the purpose and advantages of the present application, the content of the present application will be further explained in combination with the drawings and examples:

[0084] Figure 4 The test results of the pre-training model. The multi-task model MTL, the matching field method MP-MIR and the single-task model STL are compared. Among them, the single-task model only uses distance as the label for training. It can be seen that MTL is weaker than MP-MIR at long distances, but better than MP-MIR at medium and short distances. According to Table 1, the average prediction performance of MTL is better than that of MP-MIR. The performance of STL is low because it does not fully utilize the information of the sound field.

[0085] Table 1, comparison of pre-training model performance

[0086]

[0087] Figure 5To test the results of the migration model. The multi-task based migration model MTL-TL, the single-task based migration model STL-TL and the neural network ResNet trained only using the fine-tuning data set are used for comparison. Among them, ResNet still uses the neural network constructed by the method as the skeleton, and only uses the distance prediction linear layer for training. According to Table 2, it can be seen that MTL-TL achieves the best result, which greatly improves the performance of MTL. STL-TL fails to exceed MTL-TL due to the low performance of the pre-trained model, but also exceeds MP-MIR. ResNet cannot effectively predict the distance due to the over-scarcity of the training data.

[0088] Table 2, comparison of pre-trained model performance

[0089]

[0090] The above performance comparison proves the effectiveness of the present application.

[0091] 3, symbols related in this paper

[0092]

Claims

1. A method for estimating the distance to an underwater sound source based on normal mode phase and a multi-task neural network, characterized in that... The steps are as follows: Step 1: The formula for calculating the phase of an adiabatic normal mode wave is shown below: in: Indicates sound pressure phase. It is the real part of the horizontal wavenumber; Indicates frequency point, Indicates horizontal distance. Represents a mode; The sound pressure phase of multiple modes is represented by the following formula: in: This indicates the modal order contained in the signal; Broadband sound pressure phase calculated using the above formula: Normalization is performed using the following formula: in and These represent taking the real and imaginary parts of a complex vector, respectively. For the measured sound field, broadband complex sound pressure is used to represent it: Normalize: To ensure that the labels for the regression task have a uniform scale, the labels are normalized using the following formula: , , in Indicates distance label, This indicates the sound velocity label of the substrate. Indicates the number representing the modal combination. This indicates the number of training samples; the final result is the normalized simulated sound field phase, measured complex sound pressure, and corresponding labels for distance, depth, and ground sound velocity of the simulated sound field. Step 2: Construct a multi-task neural network model. The network uses stacked residual blocks: the first layer is a 64-channel layer with a size of [missing information]. The convolutional layer Conv; then the main connection part of each residual block uses two convolutional layers, with a kernel size of . Following the convolutional layers are modified linear units (ReLU) activation functions and batch normalization layers. The first two residual blocks have 64 channels, and the last two have 128 channels. In the skip connection part, the first, second, and fourth residual blocks directly add the input data and the main connection mapping data. The third residual block first uses a 128-channel convolutional layer to map the data, and then adds it to the main connection mapping data. The data is encoded by the above neural network and flattened to obtain 1024-dimensional depth features, which are then passed through three linear Fc layers to obtain 1-dimensional prediction results, corresponding to distance, bottom sound velocity, and mode, respectively. The multi-task neural network model is trained using the data obtained in step 1. During training, the loss function for each prediction result is the mean squared error (MSE), and the task weights are all set to 1. The final loss function: in , and These are the predictions of distance, sound velocity in the substrate, and modality by the neural network, respectively. Step 3: Perform transfer learning on the obtained neural network. During transfer learning, discard the linear layers for predicting the background sound velocity and mode, and only fine-tune the linear layers for predicting the distance. Step 4: Apply the neural network learned from step 3 to the measured data to obtain the distance estimation results.

2. The underwater sound source distance estimation method based on normal mode phase and multi-task neural network according to claim 1, characterized in that: In step 3, the neural network performs transfer learning by first training the model using the simulated sound pressure phase, and then fine-tuning the model using the measured sound field complex sound pressure to alleviate the distribution difference between the simulated data and the measured data.

Images