Sound field cloud picture prediction network training method based on data and physics sequential driving

Abstract

The invention provides a sound field cloud picture prediction network training method based on data and physics sequential driving, and belongs to the field of physical field modeling, and the training method comprises the steps: inputting sound field cloud picture training data into an initial sound field cloud picture prediction network to determine data driving item loss; iteratively training the initial sound field cloud picture prediction network according to the data driving item loss until the number of iterations reaches a first preset number of iterations, and obtaining a to-be-optimized sound field cloud picture prediction network; and inputting the obtained sound field cloud picture training data into the sound field cloud picture prediction network to be optimized to determine physical driving item loss, training the sound field cloud picture prediction network to be optimized according to the physical driving item loss until the number of iterations reaches a second preset number of iterations, and obtaining a completely trained sound field cloud picture prediction network. According to the method, the network parameters are quickly converged to be close to the optimal value through data driving, and then the network parameters are accurately converged through physical driving, so that the training precision can be improved, and the training time is greatly shortened.

Description

Technical Field

[0001] The present invention relates to the technical field of physical field modeling, and in particular to a sound field cloud map prediction network training method based on sequential driving of data and physics. Background Art

[0002] With the development of neural networks, they are widely used in various fields such as fluid mechanics, image and language recognition, autonomous driving, medical diagnosis, intelligent robots, etc. For the application research of neural networks in physical modeling, such as building sound field cloud maps, due to factors such as experimental costs, only a small amount of sample data can be obtained. How to achieve fast and accurate training of neural network models based on a small amount of sample data is the basis for improving the application value of neural networks.

[0003] The current training of neural networks is mainly divided into three driving methods: data-driven, physics-driven, and PINN (Physics Informed Neural Networks). The data-driven method is the most common neural network model training method, but it requires a large number of sample data sets with a certain degree of accuracy, so the model accuracy is low when the number of samples is small; the physical-driven method adds the control equation to the loss function of the neural network, which requires known prior physical knowledge, that is, to obtain the control equation that meets the requirements. The disadvantage is that the prediction accuracy is insufficient and the training time is long; the PINN drive divides the loss function of the neural network into two parts, namely the data-driven loss function term and the physical-driven loss function term, and adjusts the contribution ratio of the two loss terms through the weight coefficient, and takes the sum of the two loss terms as the total loss function. However, this method takes a long time to train and solve due to the simultaneous operation of the data-driven and physical-driven loss functions, and this method still has certain requirements for the number of sample sets and data accuracy.

[0004] Therefore, the existing technology has the problem that it is difficult to accurately and quickly train the neural network model when the sample size is small, and needs to be improved. Summary of the invention

[0005] In view of this, it is necessary to provide a sound field cloud map prediction network training method driven by data and physics in sequence to solve the technical problem in the prior art that it is difficult to accurately and quickly train the neural network model when the sample size is small.

[0006] In order to solve the above technical problems, on the one hand, the present invention provides a sound field cloud map prediction network training method, comprising: Input the acquired sound field cloud map training data into the initial sound field cloud map prediction network to obtain a first prediction output, determine the data-driven term loss according to the first prediction output, iteratively train the initial sound field cloud map prediction network according to the data-driven term loss until the number of iterations reaches a first preset number of iterations, and obtain the sound field cloud map prediction network to be optimized; The acquired sound field cloud map training data is input into the sound field cloud map prediction network to be optimized to obtain a second prediction output, the physical driving item loss is determined according to the second prediction output, and the sound field cloud map prediction network to be optimized is trained according to the physical driving item loss until the number of iterations reaches a second preset number of iterations, thereby obtaining a fully trained sound field cloud map prediction network.

[0007] In a possible implementation, the first preset number of iterations and the second preset number of iterations are determined based on a preset total number of iterations and a preset division ratio, and the first preset number of iterations is less than the second preset number of iterations.

[0008] In a possible implementation, the sound field cloud map training data includes sampling coordinates and corresponding sound pressure values ​​of a plurality of sampling points under a sound scattering physical simulation model.

[0009] In a possible implementation, the sound field cloud map prediction network includes four hidden layers, each hidden layer includes a number of neurons and a Tanh activation function.

[0010] In a possible implementation, determining a data-driven term loss according to the first prediction output includes: A mean square error is determined according to the first predicted output and the sound pressure value, and a data-driven term loss is determined according to the mean square error.

[0011] In a possible implementation, determining the physical driving item loss according to the second prediction output includes: Constructing a sound field control equation based on the physical information of the sound field to be predicted, and determining a physical driving term loss function according to the sound field control equation; A physical driven term loss is determined based on the second predicted output and a physical driven term loss function.

[0012] In one possible implementation, the physical drive loss function formula is expressed as:

[0013] in, Indicates the sound pressure, and Represents the horizontal and vertical coordinates of the prediction point, represents the wave number, Represents the pressure derivative.

[0014] On the other hand, the present invention also provides a network application method for predicting a sound field cloud image, comprising: Obtain the coordinates of the sampling points to be predicted; The coordinates of the sampling points to be predicted are input into a well-trained sound field cloud map prediction network to obtain a sound pressure prediction result, and the sound field cloud map prediction result is determined according to the sound pressure prediction result; The fully trained sound field cloud map prediction network is determined according to the above-mentioned sound field cloud map prediction network training method.

[0015] The beneficial effects of the present invention are as follows: the sound field cloud map prediction network training method provided by the embodiment of the present invention first inputs the acquired sound field cloud map training data into the initial sound field cloud map prediction network to obtain a first prediction output, determines the data-driven term loss according to the first prediction output, and iteratively trains the initial sound field cloud map prediction network according to the data-driven term loss until the number of iterations reaches a first preset number of iterations, thereby obtaining the sound field cloud map prediction network to be optimized; then inputs the acquired sound field cloud map training data into the sound field cloud map prediction network to be optimized to obtain a second prediction output, determines the physical-driven term loss according to the second prediction output, and trains the sound field cloud map prediction network to be optimized according to the physical-driven term loss until the number of iterations reaches a second preset number of iterations, thereby obtaining a fully trained sound field cloud map prediction network. The present invention first iterates the data-driven term loss to quickly converge the network parameters of the sound field cloud map prediction network to the vicinity of the optimal value, and then iterates the training through the physical-driven term loss to accurately converge the network parameters. The present invention sequentially performs data-driven and physical-driven methods to improve training accuracy and greatly reduce training time. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 A schematic diagram of a flow chart of an embodiment of a sound field cloud image prediction network training method provided by the present invention; Figure 2 A schematic diagram of a two-dimensional rectangular sound field model according to an embodiment of the present invention; Figure 3 A schematic diagram of a process for calculating the loss of a physical driving item according to an embodiment of the present invention; Figure 4 A schematic diagram of a flow chart of an embodiment of a network application method for predicting sound field cloud images provided by the present invention; Figure 5 Solve cloud diagrams for fluid simulation software; Figure 6 This is the prediction cloud diagram of the sequential driving method when the number of samples is 134 groups; Figure 7This is the prediction cloud diagram of the sequential driving method when the number of samples is 67 groups; Figure 8 This is the prediction cloud diagram of the sequential driving method when the number of samples is 14 groups; Fig. 9 Prediction cloud diagram for data-driven approach; Fig.10 For physics-driven forecast cloud maps; Fig.11 This is the prediction cloud map driven by PINN; Fig.12 The following is a comparison chart of training time for different driving modes. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0019] In the description of the embodiments of the present invention, unless otherwise specified, "multiple" means two or more than two. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" may mean: A exists alone, A and B exist at the same time, and B exists alone.

[0020] The descriptions of "first", "second", etc. involved in the embodiments of the present invention are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the technical features defined as "first" or "second" may explicitly or implicitly include at least one of the features.

[0021] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0022] The present invention provides a sound field cloud map prediction network training method and application method driven sequentially based on data and physics, which are respectively described below.

[0023] Figure 1 A flow chart of an embodiment of the sound field cloud image prediction network training method provided by the present invention is as follows Figure 1As shown, the sound field cloud image prediction network training method includes: S101, inputting the acquired sound field cloud map training data into the initial sound field cloud map prediction network to obtain a first prediction output, determining the data-driven term loss according to the first prediction output, iteratively training the initial sound field cloud map prediction network according to the data-driven term loss until the number of iterations reaches a first preset number of iterations, and obtaining the sound field cloud map prediction network to be optimized; S102, input the acquired sound field cloud map training data into the sound field cloud map prediction network to be optimized to obtain a second prediction output, determine the physical driving term loss according to the second prediction output, and train the sound field cloud map prediction network to be optimized according to the physical driving term loss until the number of iterations reaches a second preset number of iterations, so as to obtain a fully trained sound field cloud map prediction network.

[0024] Compared with the prior art, the sound field cloud map prediction network training method provided by the embodiment of the present invention first inputs the acquired sound field cloud map training data into the initial sound field cloud map prediction network to obtain a first prediction output, determines the data-driven term loss according to the first prediction output, and iteratively trains the initial sound field cloud map prediction network according to the data-driven term loss until the number of iterations reaches a first preset number of iterations, thereby obtaining the sound field cloud map prediction network to be optimized; then inputs the acquired sound field cloud map training data into the sound field cloud map prediction network to be optimized to obtain a second prediction output, determines the physical-driven term loss according to the second prediction output, and trains the sound field cloud map prediction network to be optimized according to the physical-driven term loss until the number of iterations reaches a second preset number of iterations, thereby obtaining a fully trained sound field cloud map prediction network. The present invention first iteratively trains the network parameters of the sound field cloud map prediction network through data-driven term loss, and then iterates the training through physical-driven term loss to accurately converge the network parameters. The present invention sequentially performs data-driven and physical-driven methods to improve training accuracy and greatly reduce training time.

[0025] In some embodiments of the present invention, the first preset number of iterations and the second preset number of iterations are determined based on a preset total number of iterations and a preset division ratio, and the first preset number of iterations is less than the second preset number of iterations.

[0026] Specifically, in order to ensure training accuracy and training efficiency, the present invention combines the advantages of both physical-driven and data-driven driving modes. For data-driven, the network parameters of the sound field cloud map prediction network can be quickly converged to near the optimal value in fewer rounds, but it is difficult to find accurate network parameters, so the training accuracy is poor when the amount of data is small; for physical-driven, although there are disadvantages of insufficient prediction accuracy and long training time, under the premise that the network parameters have converged to near the optimal value, physical-driven can effectively bring the network parameters closer to the optimal value, so that the network parameters of the sound field cloud map prediction network converge accurately.

[0027] Therefore, for the data-driven and physically-driven training network parameters in sequence, the first 10% to 30% of the iterations can be defined to call the data-driven item loss function, and the last 70% to 90% of the iterations can be defined to call the physically-driven item loss function. If the data-driven ratio is too small, the weights and deviation values ​​of the neural network have not converged to the vicinity of the optimal value, resulting in an increase in the solution time. If the data-driven ratio is too large, the solution accuracy will decrease due to the small proportion of physically-driven. In an embodiment, the specific preset division ratio can be determined within this range according to the empirical model. In the sound field cloud map prediction network training, the first 20% of the total iterations are used as the first preset iteration number of data-driven, and the last 80% of the total iterations are used as the second preset iteration number of physically-driven.

[0028] In some embodiments of the present invention, the sound field cloud map training data includes sampling coordinates and corresponding sound pressure values ​​of a plurality of sampling points under the sound scattering physical simulation model.

[0029] Specifically, Figure 2 Schematic diagram of a two-dimensional rectangular sound field model according to an embodiment of the present invention. Figure 2 As shown, in the embodiment, a physical model is constructed by using a two-dimensional point sound source sound scattering. The selected two-dimensional rectangular sound field model has an aspect ratio of 7:4. A point sound source radiating outward is set at (-2, 0). The four boundary conditions of the rectangle are all total reflection boundary conditions. The medium in the rectangle is air. After physical derivation, it can be obtained that the sound pressure satisfies the following equation:

[0030] in, represents the upward partial derivative of pressure, Indicates the sound pressure, represents the wave number, in the embodiment After combining the model derivation, the value is 4.5563, which corresponds to a sound frequency of 250Hz in an environment of 20℃.

[0031] The embodiment inputs the above physical information into the acoustic simulation software COMSOL for simulation calculation, and randomly sets sampling points, takes the sampling coordinates of the sampling points and the corresponding sound pressure values ​​of the points as a set of training data, and normalizes the acquired data to the range of 0 to 1, thereby obtaining a sound field cloud map training data set.

[0032] In some embodiments of the present invention, the sound field cloud map prediction network includes four hidden layers, each hidden layer includes a number of neurons and a Tanh activation function.

[0033] Specifically, the sound field cloud map prediction network in the embodiment includes four hidden layers, each layer has 32 neurons, the activation function is the Tanh activation function, and the Adam (adaptive matrix estimation) optimizer is selected for iterative optimization, the learning rate is set to 0.001, and the batch size is set to 32. The embodiment sets three cases of 134 samples, 67 samples and 14 samples to verify the training effect of the present invention. When the number of samples is 134, the samples are automatically divided into 5 batches and input into the sound field cloud map prediction network. Similarly, when the number of samples is 67, it is automatically divided into 3 batches and input into the sound field cloud map prediction network. When the number of samples is 14, it is automatically divided into 1 batch and input into the sound field cloud map prediction network.

[0034] In some embodiments of the present invention, determining the data-driven term loss according to the first prediction output includes: A mean square error is determined according to the first predicted output and the sound pressure value, and a data-driven term loss is determined according to the mean square error.

[0035] Specifically, the total number of training iterations set in the embodiment is 1000, and the first 200 are used for the data-driven term loss function according to the preset division ratio. In the data-driven term loss function, the embodiment calculates the mean square error between the first predicted output of the model output and the actual sound pressure value as the data-driven term loss, and the formula is expressed as:

[0036] in, represents the data-driven term loss, is the total number of samples, represents the sound pressure value of the first predicted output, Indicates the actual sound pressure value.

[0037] The embodiment calculates the data-driven term loss through the mean square error, and updates the network parameters through Adam adaptive matrix estimation, so that the weight deviation values ​​of each item of the sound field cloud map prediction network can be quickly iterated to near the convergence value.

[0038] In some embodiments of the present invention, Figure 3 FIG. 1 is a flow chart of calculating the loss of physical driving items according to an embodiment of the present invention. Figure 3 As shown, determining the physical driving item loss according to the second prediction output includes: S301, constructing a sound field control equation based on physical information of the sound field to be predicted, and determining a physical driving term loss function according to the sound field control equation; S302: Determine a physical driving item loss according to the second prediction output and a physical driving item loss function.

[0039] In some embodiments of the present invention, the physical drive term loss function formula is expressed as:

[0040] in, Indicates the sound pressure, and Represents the horizontal and vertical coordinates of the prediction point, represents the wave number, Represents the pressure derivative.

[0041] Specifically, for a two-dimensional sound field, the embodiment uses the Helmholtz equation to construct a sound field control equation based on the physical information of the sound field to be predicted. The formula is expressed as:

[0042] in, represents the Laplace operator, Indicates the sound pressure, Indicates the wave number.

[0043] According to the sound field control equation, the physical driving term loss function can be constructed:

[0044] in, and Represents the horizontal and vertical coordinates of the prediction point, Represents the pressure derivative.

[0045] Then, after determining the physical drive item loss function, the embodiment uses the last 800 times for the physical drive item loss function according to the preset division ratio, and also updates the network parameters through Adam adaptive matrix estimation, so that the weight deviation values ​​of each item of the sound field cloud map prediction network converge accurately.

[0046] In addition, it should be noted that the above is only an embodiment of the present invention. According to the idea of ​​the present invention, it can also be used for rapid training and prediction of other physical fields such as flow fields. The constructed neural network and parameter settings can also be adaptively changed to other types of neural networks. The data-driven item loss function and the physical-driven item loss function can also be adaptively constructed using other methods. The preset total number of iterations and the preset division ratio can also be adjusted according to actual conditions.

[0047] In summary, in order to accurately and quickly train the sound field cloud map prediction network when the sample size is small, the present invention first inputs the acquired sound field cloud map training data into the initial sound field cloud map prediction network to obtain a first prediction output, determines the data-driven term loss according to the first prediction output, and iteratively trains the initial sound field cloud map prediction network according to the data-driven term loss until the number of iterations reaches the first preset number of iterations, and obtains the sound field cloud map prediction network to be optimized; then the acquired sound field cloud map training data is input into the sound field cloud map prediction network to be optimized to obtain a second prediction output, determines the physical-driven term loss according to the second prediction output, and trains the sound field cloud map prediction network to be optimized according to the physical-driven term loss until the number of iterations reaches the second preset number of iterations, and obtains a fully trained sound field cloud map prediction network. The present invention first iteratively trains the network parameters of the sound field cloud map prediction network through data-driven term loss, and then iterates the training through physical-driven loss to accurately converge the network parameters. The present invention sequentially performs data-driven and physical-driven methods to improve training accuracy and greatly reduce training time.

[0048] The embodiment of the present invention also provides a network application method for predicting a sound field cloud image, combining Figure 4 Come and see, Figure 4 A schematic diagram of a flow chart of an embodiment of a network application method for predicting sound field cloud images provided by the present invention is shown in FIG. Figure 4 As shown, the network application method for predicting the sound field cloud image includes: S401, obtaining the coordinates of the sampling points to be predicted; S402, inputting the coordinates of the sampling points to be predicted into a well-trained sound field cloud map prediction network to obtain a sound pressure prediction result, and determining a sound field cloud map prediction result according to the sound pressure prediction result; The fully trained sound field cloud map prediction network is determined according to the above-mentioned sound field cloud map prediction network training method.

[0049] In an embodiment of the present invention, first, the coordinates of the sampling points to be predicted are effectively acquired, and then the sound pressure is effectively predicted using the above-mentioned fully trained sound field cloud map prediction network, so that the sound pressure prediction result can be output, and the sound field cloud map prediction result is constructed based on the sound pressure prediction result.

[0050] To verify the effectiveness of the present invention, the present invention also conducts an experimental comparison between the data-driven and physical-driven method of the present invention and the existing data-driven, physical-driven and PINN-driven methods. Figure 5 To solve the cloud diagram for fluid simulation software, Figure 6 This is the prediction cloud diagram of the sequential driving method when the number of samples is 134. Figure 7 This is the prediction cloud diagram of the sequential driving method when the number of samples is 67 groups. Figure 8 This is the prediction cloud diagram of the sequential driving method when the number of samples is 14 groups. Fig. 9This is the prediction cloud diagram of the data-driven method. Fig.10 For the physics-driven prediction cloud map, Fig.11 This is the prediction cloud map driven by PINN. Fig.12 The following is a comparison chart of the training time of different driving modes. The number of iterations of data driving is 3000, and both data driving and physical driving have 134 sets of samples. Figures 5 to 12 It can be seen that the sequential driving method of the present invention has a similar prediction effect to the PINN method when trained with 134 sets of samples, and is significantly better than the data-driven or physical-driven method. The training time of the present invention is only 17.438s, which is about 87s less than the PINN method. Figure 7 and Figure 8 It can be seen that this scheme can still achieve good prediction results when the number of samples is 67 groups, and the prediction results are average but still better than the physical drive method when the number of samples is 14 groups. It can be seen that for small sample training, the present invention is superior to the existing three methods in terms of accuracy and time cost.

[0051] The above is a detailed introduction to the sound field cloud map prediction network training method based on data and physics driven in sequence provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A sound field cloud map prediction network training method, characterized in that: include: Inputting the acquired sound field cloud map training data into the initial sound field cloud map prediction network to obtain a first prediction output, determining a data-driven term loss according to the first prediction output, iteratively training the initial sound field cloud map prediction network according to the data-driven term loss until the number of iterations reaches a first preset number of iterations, and obtaining a sound field cloud map prediction network to be optimized; The acquired sound field cloud map training data is input into the sound field cloud map prediction network to be optimized to obtain a second prediction output, the physical driving item loss is determined according to the second prediction output, and the sound field cloud map prediction network to be optimized is trained according to the physical driving item loss until the number of iterations reaches a second preset number of iterations, so as to obtain a fully trained sound field cloud map prediction network.

2. The sound field cloud image prediction network training method according to claim 1, characterized in that: The first preset number of iterations and the second preset number of iterations are determined based on a preset total number of iterations and a preset division ratio, and the first preset number of iterations is less than the second preset number of iterations.

3. The sound field cloud image prediction network training method according to claim 1, characterized in that: The sound field cloud map training data includes sampling coordinates and corresponding sound pressure values ​​of a plurality of sampling points under a sound scattering physical simulation model.

4. The sound field cloud image prediction network training method according to claim 1, characterized in that: The sound field cloud map prediction network includes four hidden layers, each of which contains a number of neurons and a Tanh activation function.

5. The sound field cloud image prediction network training method according to claim 3, characterized in that: The determining of the data-driven term loss according to the first prediction output comprises: A mean square error is determined according to the first predicted output and the sound pressure value, and a data-driven term loss is determined according to the mean square error.

6. The sound field cloud image prediction network training method according to claim 1, characterized in that: The determining of the physical driving item loss according to the second prediction output comprises: Constructing a sound field control equation based on physical information of the sound field to be predicted, and determining a physical driving term loss function according to the sound field control equation; A physical driven term loss is determined based on the second predicted output and the physical driven term loss function.

7. The sound field cloud image prediction network training method according to claim 6, characterized in that: The physical drive loss function formula is expressed as: in, Indicates the sound pressure, and Represents the horizontal and vertical coordinates of the prediction point, represents the wave number, Represents the pressure derivative.

8. A network application method for predicting sound field cloud images, characterized in that: include: Obtain the coordinates of the sampling points to be predicted; Inputting the coordinates of the sampling points to be predicted into a well-trained sound field cloud map prediction network to obtain a sound pressure prediction result, and determining a sound field cloud map prediction result according to the sound pressure prediction result; Wherein, the fully trained sound field cloud map prediction network is determined according to the sound field cloud map prediction network training method according to any one of claims 1 to 7.

Images