Ship full-band underwater noise forecasting method based on multi-forecasting method fusion

Abstract

The invention discloses a ship full-band underwater noise forecasting method based on multi-forecasting method fusion, and relates to the technical field of ship noise forecasting. According to the method, underwater noise forecasting is carried out on a target ship through multiple different underwater noise forecasting methods. And then inputting the plurality of obtained underwater noise forecasting results into a fusion forecasting network constructed based on a neural network model, and performing attention enhancement on spectrum signals of different characteristic frequency bands in the plurality of underwater noise forecasting results by the fusion forecasting network based on a flexible attention mechanism to realize data fusion. And obtaining a full-band underwater noise fusion forecast result. The method can have the advantages of various underwater noise forecasting methods, thereby being beneficial to improving the reliability and accuracy of ship full-band underwater noise forecasting.

Description

Technical Field

[0001] The present application relates to the technical field of ship noise prediction, and in particular to a method for predicting full-band underwater noise of ships based on the fusion of multiple prediction methods. Background Art

[0002] Underwater noise radiation caused by vibrations from ship machinery not only impacts the ship's environmental performance but also the comfort, physical and mental health of crew and passengers. Classification societies around the world, including the International Maritime Organization, have established stringent regulations for underwater noise emissions. Therefore, underwater noise prediction is essential during the ship design phase to facilitate noise reduction design.

[0003] Currently, commonly used methods for predicting underwater ship noise include the finite element method, boundary element method, statistical energy method, hybrid finite element and statistical energy method, neural network method, and others. However, these existing methods all have their limitations. Given the wide spectrum of underwater ship noise signals, reaching over 10 kHz, existing methods are unable to accurately predict the full frequency range of underwater ship noise. Summary of the Invention

[0004] In response to the above-mentioned problems and technical needs, this application proposes a method for predicting full-band underwater noise of ships based on the fusion of multiple prediction methods. The technical solution of this application is as follows:

[0005] A ship full-band underwater noise prediction method based on the fusion of multiple prediction methods, the ship full-band underwater noise prediction method includes:

[0006] Underwater noise prediction is performed on the target ship using M different underwater noise prediction methods, and M underwater noise prediction results are obtained at the target ship's predicted position.

[0007] M underwater noise forecast results are input into a fusion forecast network. The fusion forecast network is pre-built based on a neural network model. The fusion forecast network uses a flexible attention mechanism to enhance the attention of the spectrum signals of different characteristic frequency bands in the M underwater noise forecast results, and then realizes data fusion to obtain the full-band underwater noise fusion forecast result at the location to be forecasted; wherein, the integer parameter M ≥ 2.

[0008] Its further technical solution is that the fusion prediction network includes M flexible attention modules and a feature fusion module;

[0009] Each underwater noise prediction result is input into a corresponding flexible attention module. The flexible attention module enhances the attention of the spectrum signals of different characteristic frequency bands in the input underwater noise prediction result based on the flexible attention mechanism and outputs the attention enhancement result.

[0010] The attention enhancement results output by M flexible attention modules are all output to the feature fusion module. The feature fusion module fuses the M attention enhancement results and outputs the full-band underwater noise fusion prediction result.

[0011] Its further technical solution is that each flexible attention module includes a feature extraction unit, K attention units and a multiplication unit, and the integer parameter K ≥ 2;

[0012] The feature extraction unit extracts features from the spectrum signals of K characteristic frequency bands in the underwater noise forecast result to obtain feature vectors of the K characteristic frequency bands. The feature vector of each characteristic frequency band is input into an attention unit, and the attention unit outputs the attention weight corresponding to the characteristic frequency band; the multiplication unit multiplies the spectrum signal of each characteristic frequency band of the underwater noise forecast result with the attention weight corresponding to the characteristic frequency band and outputs the attention enhancement result.

[0013] Its further technical solution is that in the full frequency band of ship underwater noise, the division density of characteristic frequency bands in the high frequency band, medium frequency band and low frequency band increases successively. The lower the frequency band in which the characteristic frequency band is located, the smaller the frequency range covered.

[0014] A further technical solution is to divide the characteristic frequency band based on the 1 / 3 octave principle.

[0015] A further technical solution is that the characteristic frequency band division methods of any two underwater noise prediction results are the same or different.

[0016] Its further technical solution is to use the loss function Loss=L+λ×R(θ) when training the fusion prediction network, where L is the prediction error of the fusion prediction network and β n is the noise data at sampling frequency n in the full-band underwater noise measurement results of the training sample, is the noise data at sampling frequency n in the full-band underwater noise fusion prediction result obtained by inputting the training sample into the fusion prediction network, N is the total number of sampling frequency points in the full-band ship underwater noise; λ is the regularization coefficient, which is used to control the intensity of regularization, R(θ) is the L2 norm, θ=<θ a ,θ l > represents the neural network parameters θ representing the attention module a and the neural network parameters θ of the feature fusion module l Collection of.

[0017] A further technical solution is that each attention unit includes multiple hidden layers cascaded in sequence and a Softmax function.

[0018] The beneficial technical effects of this application are:

[0019] The present application discloses a method for predicting full-band underwater noise of ships based on the fusion of multiple prediction methods. The method implements data fusion after enhancing the attention of spectral signals of different characteristic frequency bands in the underwater noise prediction results based on a flexible attention mechanism, and obtains the full-band underwater noise fusion prediction results based on a deep learning method. The method can combine the advantages of multiple underwater noise prediction methods, thereby improving the reliability and accuracy of full-band underwater noise prediction of ships.

[0020] This method analyzes the frequency band contribution of each underwater noise prediction method based on the flexible attention mechanism, thereby quantifying the advantageous prediction interval of each underwater noise prediction method and forming an attention matrix. This matrix can visualize the advantageous prediction intervals of underwater noise prediction methods in different characteristic frequency bands, and also helps to analyze and understand the prediction characteristics of various underwater noise prediction methods when realizing data fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a method for predicting full-band underwater noise of a ship according to an embodiment of the present application.

[0022] Figure 2 Figure 2 is a network structure diagram of a single flexible attention module in one embodiment.

[0023] Figure 3 FIG. 1 is a schematic diagram of feature fusion in an embodiment. DETAILED DESCRIPTION

[0024] The specific implementation of this application will be further described below with reference to the accompanying drawings.

[0025] This application discloses a method for predicting ship underwater noise in all frequency bands based on the fusion of multiple prediction methods. Figure 1 As shown in the process framework diagram, the ship full-band underwater noise prediction method includes:

[0026] Step 1: Use M different underwater noise prediction methods to predict the underwater noise of the target ship, and obtain M underwater noise prediction results at the target ship's predicted position. The integer parameter M is ≥ 2, but usually, the value of M is at least 3. Figure 1 Taking underwater noise prediction methods 1 to M as an example, the corresponding underwater noise prediction results X1 to X M .

[0027] The underwater noise prediction methods used in this step include finite element method, boundary element method, statistical energy method, finite element-statistical energy hybrid method, neural network method, and transfer function matrix method based on actual ship test data. The specific implementation of these methods will not be repeated in this application.

[0028] Each underwater noise prediction method has its own corresponding dominant frequency band, and the underwater noise prediction results obtained by the underwater noise prediction method are highly accurate within the dominant frequency band. The focus of this application is to fuse the underwater noise prediction results obtained by multiple underwater noise prediction methods, so as to combine the advantages of different underwater noise prediction methods. Therefore, the dominant frequency bands of the M underwater noise prediction methods selected overlap and cover the low-frequency band, mid-frequency band, and high-frequency band of the full frequency band of ship underwater noise. For example, one method is to combine the finite element method, the finite element-statistical energy hybrid method, and the statistical energy method. The dominant frequency band of the finite element method is the low-frequency band, the dominant frequency band of the finite element-statistical energy hybrid method is the mid-frequency band, and the dominant frequency band of the statistical energy method is the high-frequency band, which can meet the above conditions. The dominant frequency bands of various other underwater noise prediction methods are all known. Based on the above principles, the combination of the M underwater noise prediction methods can be determined. As long as the combination meets the above requirements, this method can be used.

[0029] Step 2: Input the M underwater noise prediction results into the fusion prediction network to obtain the full-band underwater noise fusion prediction result at the location to be predicted.

[0030] Although the dominant frequency bands of the M underwater noise prediction methods overlap, the boundaries between these dominant frequency bands are not clear. Therefore, it is impossible to simply hard-splice the prediction results of the M underwater noise prediction methods in the dominant frequency bands, and the results are not accurate enough.

[0031] Random forests can achieve multi-strategy fusion by building decision trees. However, random forests excel at solving classification problems, while noise prediction is essentially a regression mapping or prediction problem. These two are fundamentally different. Random forests are typically only able to rank the importance (contribution) of features, which are often discrete, making their application to continuous noise signal processing somewhat challenging. DS evidence theory can also be used to achieve strategy fusion, but it lacks self-learning capabilities and relies on prior knowledge. Therefore, these two commonly used methods cannot be used to fuse underwater noise prediction results.

[0032] Based on the above considerations, this application pre-constructs a fusion prediction network based on a neural network model, and uses the fusion prediction network to enhance the attention of the spectral signals of different characteristic frequency bands in the M underwater noise forecast results based on the flexible attention mechanism, and then realizes data fusion, thereby obtaining a comprehensive full-band underwater noise fusion prediction result.

[0033] Please refer to Figure 1 As shown in the network structure diagram, the fusion prediction network includes M flexible attention modules and a feature fusion module. Figure 1The M flexible attention modules are respectively denoted as flexible attention module 1 to flexible attention module M.

[0034] Each underwater noise prediction result is input into a corresponding flexible attention module, and the attention enhancement results output by M flexible attention modules are output to the feature fusion module. Figure 1 In the underwater noise prediction results X1~X M They correspond to the input flexible attention module 1 to flexible attention module M respectively, and the outputs of flexible attention module 1 to flexible attention module M are all connected to the feature fusion module.

[0035] The flexible attention module uses a flexible attention mechanism to enhance the spectral signals of different characteristic frequency bands in the input underwater noise prediction results and outputs the enhanced attention results. Each flexible attention module consists of a feature extraction unit, K attention units, and a multiplication unit. The integer parameter K represents the number of characteristic frequency bands divided into the full frequency band of ship underwater noise. The K characteristic frequency bands are continuous and non-overlapping, each covering a corresponding frequency range, and K ≥ 2. Each attention unit consists of multiple hidden layers cascaded in sequence and a softmax function.

[0036] Taking the flexible attention module i as an example, its network structure is as follows Figure 2 As shown, the integer parameter 1≤i≤M. The feature extraction unit in the flexible attention module i respectively performs the underwater noise prediction result X i The feature vectors of K characteristic frequency bands are obtained by extracting the feature signals of the K characteristic frequency bands. Figure 2 In the middle, x1~x K They represent the underwater noise prediction results X i The spectrum signal of the 1st characteristic frequency band to the Kth characteristic frequency band, Respectively represent x1~x K The extracted eigenvectors of the 1st to Kth eigenfrequency bands, eigenvectors Construct the eigenvector matrix E i .

[0037] The feature vector of each characteristic frequency band is input into an attention unit, and the attention unit outputs the attention weight corresponding to the characteristic frequency band. Figure 2 In the eigenvector Corresponding to the input attention unit 1 to attention unit K, attention unit 1 to attention unit K respectively output the attention weights of the 1st feature frequency band to the Kth feature frequency band Attention weight Construct the attention matrix A i The larger the attention weight of a characteristic frequency band is, the better the underwater noise prediction result X is. iThe greater the prediction advantage in this characteristic frequency band. In one example, the underwater noise prediction result X 1～ The visualization diagram of the attention weight of X3 in different feature frequency bands is as follows Figure 3 As shown, according to the underwater noise prediction result X1, we can get According to the underwater noise prediction result X2, we can get According to the underwater noise prediction result X3, we can get The color depth indicates the attention weight corresponding to each characteristic frequency band in each underwater noise prediction result. The darker the color of the characteristic frequency band, the greater the attention weight corresponding to the characteristic frequency band, indicating that the advantage of the characteristic frequency band is higher. Figure 3 It can be seen that although various underwater prediction methods have their own theoretical advantage frequency bands, the flexible attention enhancement operation of this application can further subdivide and quantify the attention weight of each characteristic frequency band. For example, the theoretical advantage range of an underwater noise prediction method is the low frequency band, but according to the underwater noise prediction result X1 obtained, it can be determined that it does not have a large advantage in all low frequency bands. In fact, the advantage is more obvious in some characteristic frequency bands of the low frequency band. Figure 3 It can be seen from the example that each underwater noise prediction result X i The attention weights in different characteristic frequency bands do not change monotonically, which leads to the difference in the prediction results of each underwater noise X. i The advantage interval of is not necessarily a continuous and monotonous change, which also indirectly shows that it is not feasible to hard-splice multiple underwater noise prediction results.

[0038] Then the multiplication unit converts the underwater noise prediction result X i The spectral signal of each characteristic frequency band is multiplied by the attention weight corresponding to the characteristic frequency band and the output attention enhancement result Y is obtained. i By this operation, the underwater noise prediction result X can be further enhanced. i The spectrum signal in the advantageous prediction interval is weakened by the underwater noise prediction result X i Spectral signal in the non-dominant prediction interval.

[0039] In one embodiment, to better meet the needs of ship underwater noise prediction, the density of the characteristic frequency bands in the high-frequency, mid-frequency, and low-frequency regions of the full frequency band of ship underwater noise increases in sequence. The lower the frequency band, the smaller the frequency range covered. In other words, the characteristic frequency bands are more densely divided in the low-frequency region, while the characteristic frequency bands in the high-frequency region are relatively sparse. Based on this principle, one implementation method is to divide the full frequency band of ship underwater noise into K characteristic frequency bands based on the 1 / 3 octave principle.

[0040] In addition, considering that the characteristics of different ship underwater noise prediction methods are different, the underwater noise prediction results obtained also have different characteristics. Therefore, the characteristic frequency band division methods of any two underwater noise prediction results are the same or different, and different underwater noise prediction results can be divided into characteristic frequency bands according to their own characteristics.

[0041] Finally, the feature fusion module fuses the M attention enhancement results and outputs a full-band underwater noise fusion prediction result. The feature fusion module can use a simple deep learning model.

[0042] The present application utilizes a fusion prediction network to integrate the advantages of multiple underwater noise predictions, thereby achieving flexible fusion to obtain a full-band underwater noise fusion prediction result with higher accuracy. After the fusion prediction network is built, it is necessary to use training samples for model training, including obtaining M underwater noise prediction results at a position of the target ship under M different underwater noise prediction methods as training samples, and using sensor acquisition and other methods to collect the full-band underwater noise measurement results at this position. The M underwater noise prediction results of the training samples are then input into the fusion prediction network to obtain the full-band underwater noise fusion prediction result of the training sample, and the loss function is calculated to perform back propagation to update the network parameters to achieve network training. The loss function Loss = L + λ × R (θ) used for network training of the fusion prediction network, where L is the prediction error of the fusion prediction network and Among them, β n is the noise data at sampling frequency n in the full-band underwater noise measurement results of the training sample, is the noise data at sampling frequency n in the full-band underwater noise fusion prediction result obtained by inputting the training sample into the fusion prediction network, and N is the total number of sampling frequencies in the full-band ship underwater noise. λ is the regularization coefficient, which is used to control the strength of regularization. R(θ) is the L2 norm, which helps to speed up the optimization process and prevent overfitting. θ=<θ a ,θ l > represents the neural network parameters θ representing the attention module a and the neural network parameters θ of the feature fusion module l Collection of.

[0043] The above description is only a preferred embodiment of the present application, and the present application is not limited to the above embodiments. It is understood that other improvements and variations directly derived or imagined by those skilled in the art without departing from the spirit and concept of the present application should be considered to be included in the scope of protection of the present application.

Claims

1. A method for predicting full-band underwater noise of ships based on the fusion of multiple prediction methods, characterized in that: The method for predicting full-band underwater noise of a ship comprises: Underwater noise prediction is performed on a target ship using M different underwater noise prediction methods, respectively, to obtain M underwater noise prediction results at the target ship's position to be predicted; M underwater noise prediction results are input into a fusion prediction network. The fusion prediction network is pre-constructed based on a neural network model. The fusion prediction network performs data fusion after enhancing the attention of the spectrum signals of different characteristic frequency bands in the M underwater noise prediction results based on a flexible attention mechanism, thereby obtaining a full-band underwater noise fusion prediction result at the position to be predicted; wherein the integer parameter M ≥ 2.

2. The method for predicting full-band underwater noise of a ship according to claim 1, characterized in that: The fusion prediction network includes M flexible attention modules and a feature fusion module; Each underwater noise prediction result is input into a corresponding flexible attention module. The flexible attention module performs attention enhancement on the spectrum signals of different characteristic frequency bands in the input underwater noise prediction result based on the flexible attention mechanism and outputs the attention enhancement result. The attention enhancement results output by the M flexible attention modules are all output to the feature fusion module, and the feature fusion module performs feature fusion on the M attention enhancement results and outputs the full-band underwater noise fusion prediction result.

3. The method for predicting full-band underwater noise of a ship according to claim 2, characterized in that: Each flexible attention module consists of a feature extraction unit, K attention units and a multiplication unit, with an integer parameter K ≥ 2; The feature extraction unit extracts features from the spectrum signals of K characteristic frequency bands in the underwater noise prediction result respectively to obtain feature vectors of the K characteristic frequency bands. The feature vector of each characteristic frequency band is input into an attention unit, and the attention unit outputs the attention weight corresponding to the characteristic frequency band; the multiplication unit multiplies the spectrum signal of each characteristic frequency band of the underwater noise prediction result by the attention weight corresponding to the characteristic frequency band and outputs the attention enhancement result.

4. The method for predicting full-band underwater noise of a ship according to claim 3, characterized in that: In the full frequency band of ship underwater noise, the division density of characteristic frequency bands in the high frequency band, medium frequency band and low frequency band increases successively. The lower the frequency band of the characteristic frequency band is, the smaller the frequency range covered is.

5. The method for predicting full-band underwater noise of a ship according to claim 4, characterized in that: The characteristic frequency bands are divided based on the 1 / 3 octave principle.

6. The method for predicting full-band underwater noise of a ship according to claim 4, characterized in that: The characteristic frequency band division methods of any two underwater noise prediction results are the same or different.

7. The method for predicting full-band underwater noise of a ship according to claim 3, characterized in that: The loss function used when training the fusion prediction network is Loss=L+λ×R(θ), where L is the prediction error of the fusion prediction network and β n is the noise data at sampling frequency n in the full-band underwater noise measurement results of the training sample, is the noise data at sampling frequency n in the full-band underwater noise fusion prediction result obtained by inputting the training sample into the fusion prediction network, N is the total number of sampling frequency points in the full-band ship underwater noise; λ is the regularization coefficient, which is used to control the intensity of regularization; R(θ) is the L2 norm, θ=<θ a ,θ l > represents the neural network parameters θ representing the attention module a and the neural network parameters θ of the feature fusion module l Collection of.

8. The method for predicting full-band underwater noise of a ship according to claim 3, characterized in that: Each attention unit consists of multiple hidden layers cascaded in sequence and a Softmax function.

Images