Very low frequency signal amplitude distribution evaluation convolutional neural network model and training method thereof

Abstract

This invention discloses a convolutional neural network model and its training method for evaluating the amplitude distribution of very low frequency (VLF) signals, addressing the problems of long computation time and low efficiency in traditional physical models. The model includes: an input layer configured to receive data on the effective reflection height and electron density variation rate at the bottom of the ionosphere; a convolutional layer configured to extract feature information by performing a convolution operation on the data; a pooling layer configured to compress and extract the feature information extracted by the convolutional layer; a fully connected layer configured to map the extracted feature information to VLF signal amplitude; and an output layer configured to output the VLF signal amplitude. This invention significantly reduces computation time while ensuring the accuracy of the calculation results, and has significant value in the engineering application field of VLF detection.

Description

Technical Field

[0001] This invention relates to a convolutional neural network model for evaluating the amplitude distribution of very low frequency signals and its training method. Background Technology

[0002] The D layer of the ionosphere (lower ionosphere) is the partially ionized atmosphere at an altitude of 60-100 kilometers above the ground, with an electron density typically ranging from 1 to 1000 cm³. -3 This region of space is affected not only by solar activity from top to bottom, but also by atmospheric fluctuations from bottom to top. The electron density of the D layer of the ionosphere is crucial for radio communication, radio wave propagation, maritime navigation and positioning, and space environment monitoring.

[0003] Very Low Frequency (VLF) waves refer to electromagnetic waves with frequencies in the range of 3-30 kHz, primarily originating from thunderstorms or artificial VLF stations distributed globally. Due to their unique frequency characteristics, VLF waves are often confined within the Earth-ionospheric waveguide, consisting of the ground and the D layer of the ionosphere, enabling long-distance propagation. A typical attenuation coefficient is approximately ~2-3 dB / Mm. Artificial VLF stations typically have large antennas, requiring antenna arrays to be erected in open areas such as valleys or large plains, resulting in high costs. Currently, most existing artificial VLF transmitters worldwide are used in the military field, and publicly available information is limited. In engineering applications such as VLF navigation and positioning, it is often necessary to calculate the signal distribution around VLF transmitting stations, but these calculations are time-consuming. Even using a significantly simplified Long-Wavelength Propagation Capability (LWPC) model, the calculation time is still tens of minutes, failing to simultaneously meet the real-time and high-accuracy requirements of engineering applications.

[0004] Deep learning has demonstrated exceptional capabilities in revealing deeper underlying patterns and automatically extracting features, while leveraging graphics processing units (GPUs) for high accuracy and rapid computation. Deep learning-based prediction methods are gradually replacing traditional models, which typically require significant computational time on central processing units (CPUs). Given its powerful nonlinear modeling capabilities, deep learning has gained considerable attention in the field of computational mechanics. Summary of the Invention

[0005] Therefore, this invention proposes a fast evaluation model for the amplitude distribution of very low frequency (VLF) signals based on deep learning and its training method, which replaces the traditional VLF propagation model. While maintaining computational accuracy, it significantly reduces computation time to meet the needs of engineering applications.

[0006] Firstly, a convolutional neural network model for evaluating the amplitude distribution of very low frequency (VLF) signals is proposed, comprising: an input layer configured to receive data on the effective reflection height and the rate of change of electron density with height at the bottom of the ionosphere; a convolutional layer configured to perform a convolution operation on the data on the effective reflection height and the rate of change of electron density with height at the bottom of the ionosphere to extract feature information; a pooling layer configured to perform a pooling operation on the feature information extracted by the convolutional layer; a fully connected layer configured to map the feature information after the pooling operation to the VLF signal amplitude; and an output layer configured to output the VLF signal amplitude.

[0007] In some examples, the convolutional layer consists of multiple groups of convolutional kernels of different sizes, which are first grouped and then stacked to perform grouped convolution on the effective reflection height and the rate of electron density change with height at the bottom of the ionosphere. By using the grouped convolutional structure, the nonlinearity of the network is enhanced, improving the effectiveness of feature extraction for h' (electron reflection height) and β (sharpness parameter, the rate of electron density change with height).

[0008] In some examples, residual structures are used to span partial convolutional layers to deepen the depth of the convolutional neural network model for evaluating the amplitude distribution of the very low frequency signal while ensuring its stability, thus addressing the gradient vanishing problem.

[0009] Secondly, a training method for the very low frequency (VLF) signal amplitude distribution evaluation convolutional neural network model is proposed, comprising: establishing a VLF transmitter signal amplitude distribution dataset, wherein the dataset consists of several input and output combinations, wherein in each combination, the input includes the effective reflection height at the bottom of the ionosphere and the rate of change of electron density with height, and the output is the corresponding VLF signal amplitude; preprocessing the dataset to convert the input into an input format that meets the requirements of the VLF signal amplitude distribution evaluation convolutional neural network model; and inputting the preprocessed dataset into the VLF signal amplitude distribution evaluation convolutional neural network model for training.

[0010] In some examples, the LWPC physical model is used to calculate the effective reflection height at the bottom of the ionosphere and the very low frequency signal amplitude corresponding to the rate of change of electron density with height.

[0011] In some examples, the difference between the very low frequency (VLF) signal amplitude calculated by the LWPC physical model during training and the VLF signal amplitude amplitude predicted by the VLF signal amplitude distribution evaluation convolutional neural network model is used as the loss function.

[0012] In some examples, dropout (a special network layer, dropout layer) is used during training to enhance the ability of the convolutional neural network model to learn the rate of change of effective reflection height and electron density at the bottom of the ionosphere with height, thereby enhancing the robustness and generalization of the network model, effectively reducing model parameters and complexity, and avoiding overfitting.

[0013] Thirdly, a computer system is proposed, including a processor and a memory, wherein the very low frequency signal amplitude distribution evaluation convolutional neural network model is stored in the memory and configured to be executed by the processor.

[0014] Fourthly, a computer-readable storage medium is proposed for storing non-transitory computer-readable instructions that, when executed by a computer, enable the execution of the very low frequency signal amplitude distribution evaluation convolutional neural network model.

[0015] This invention utilizes a deep learning network to replace the LWPC model, accelerating the calculation of very low frequency (VLF) station signal distribution. A VLF amplitude distribution dataset is constructed using the LWPC model. A deep CNN network is then built, and the dataset is preprocessed to adapt to the input and output format of the CNN network. The dataset is shuffled and proportionally divided into training and validation sets, which are then input into the CNN network to obtain deep feature information. This information is then concatenated with the output data to establish a mapping relationship between the input and output data. The trained CNN network model is used for prediction, transforming the original physical model calculation problem into a prediction problem. Iterative output from the CNN network model allows for more efficient use of its model characteristics, improving prediction accuracy and efficiency. Finally, the CNN network model is deployed in a suitable format, exported, and integrated into the application system. This invention significantly reduces computation time while ensuring the accuracy of the calculation results, and has significant value in the engineering application field of VLF detection. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a convolutional neural network model for evaluating the amplitude distribution of very low frequency signals according to an embodiment of the present invention.

[0017] Figure 2 This is a flowchart of a training method for evaluating the amplitude distribution of very low frequency signals using a convolutional neural network model according to an embodiment of the present invention.

[0018] Figure 3 This is a graph showing the change in loss values ​​during different network training processes according to an embodiment of the present invention.

[0019] Figure 4 The amplitude results obtained by calculating a set of h' and β using the LWPC model according to an embodiment of the present invention are as follows. Figure 4 (a) is h'. Figure 4 (b) is β. Figure 4 (c) is the calculated amplitude.

[0020] Figure 5 This is a comparison chart of the prediction results using a trained network and the results using an LWPC model, according to an embodiment of the present invention. Figure 5 (a) is the amplitude calculated by the LWPC model. Figure 5 (b) The results obtained using network prediction. Figure 5 (c) is Figure 5 (a) and Figure 5 (b) Error value between the two sets of results. Detailed Implementation

[0021] First, we introduce a deep learning-based evaluation model for the amplitude distribution of very low frequency signals (hereinafter referred to as the "model"). It transforms the physical model calculation problem of very low frequency propagation into a prediction problem, that is, it realizes the calculation relationship between the effective reflection height (h') at the bottom of the ionosphere and the rate of change of electron density with height (β) and the output amplitude into a network mapping relationship.

[0022] like Figure 1 As shown, the very low frequency signal amplitude distribution evaluation model employs a deep convolutional neural network (CNN), including convolutional layers, pooling layers, and fully connected layers. For example, convolutional layers, pooling layers, and fully connected layers can be accumulated to 112 layers, forming a deep network. Furthermore, residual structures, grouped convolutional structures, and dropout structures are combined to enhance the network's ability to learn the features of h' and β. The introduction of residual structures, by connecting the preceding and following layers through shortcuts and spanning several intermediate convolutional layers, ensures network stability while increasing network depth and addresses the gradient vanishing problem. The use of grouped convolutional structures, formed by grouping and then stacking multiple sets of convolutional kernels of different sizes, enhances the network's nonlinearity and improves the effectiveness of extracting h' and β features. After parameter feature extraction, a dropout structure is added to discard some network feature parameters, addressing the overfitting problem of deep networks. For model training, the root mean square error (RMSE) was chosen as the loss function for the learning-based regression problem of the LWPC physics model. Specifically, the loss function is the error between the model's predicted amplitude and the amplitude calculated by the LWPC physics model. The optimizer was Adam, and the learning rate was chosen as lr(epoch) to adapt to the changing trends of parameters during the learning process.

[0023]

[0024] Figure 2 This is a flowchart illustrating the training method for a deep learning-based evaluation model of very low frequency signal amplitude distribution. The training method is explained in detail below.

[0025] Step 1: Using the Long-Wavelength Propagation Capability (LWPC) model, simulate and calculate the amplitude distribution of different VLF stations at different times, and establish a dataset of VLF transmitter signal amplitude distribution.

[0026] Step 11: Input the time of the very low frequency station's transmitted signal, as well as the station's frequency, power, and other parameters into the LWPC model, and output the amplitude distribution information of the station's transmitted signal.

[0027] Step 12: Establish the amplitude distribution data of very low frequency stations. The main influencing factor of amplitude is electron density. By using the formula (1) proposed by Wait and Spies, the electron density and h' are linked with β, thereby realizing the mapping relationship between amplitude and h' with β.

[0028] N e (z) = 1.43 × 10 13 exp(-0.15h)exp(β(z-h')) (1)

[0029] Where, N e Let be the electron density, z be the height, h' be the effective reflection height at the bottom of the ionosphere, and β be the rate at which the electron density changes with height.

[0030] Step 13: The dataset consists of several sets of input and output combinations. In each combination, the input consists of h' and β, and the output is the corresponding amplitude.

[0031] Step 2: Preprocess the dataset by shuffling it and dividing it into training and validation sets according to the specified proportions. This enhances the model's generalization ability and speeds up the training process.

[0032] Step 21: Preprocess the h' and β datasets, converting the network inputs h' and β into two-dimensional arrays to meet the input format requirements of the CNN network;

[0033] Step 22: Randomly shuffle the processed dataset to enhance the model's generalization ability and accelerate the training gradient descent rate;

[0034] Step 23: Divide the data into a dataset and a validation set in an 8:2 ratio. This is used to test the model configuration and training results, as well as whether the validation results are overfitting or underfitting, so that adjustments can be made and the model can be retrained.

[0035] Step 3: During the training phase of the neural network, the preprocessed data is input into the CNN network to obtain the deep feature information of the data. Then, the network and the output data are connected to realize the mapping relationship between the input data and the output data.

[0036] Step 31: Input the training data into the network according to the set number of training epochs and maximum number of input samples (batch size);

[0037] Step 32: During the training process, the convolutional layers of the CNN network acquire the feature information of h' and β, introduce nonlinearity through activation functions, and finally reduce the complexity of the feature information through pooling layers and extract the main features.

[0038] Step 33: The input data h' and β features are made one-dimensional through a fully connected layer to complete the target feature regression task; the output layer connects the network and the output data amplitude to realize the mapping and transformation relationship between the input h', β and the output amplitude data.

[0039] By using a trained CNN network model for very low frequency amplitude prediction, the original physical model calculation problem is transformed into a prediction problem. Through iterative model output, the model characteristics of the CNN network can be utilized more efficiently, improving the accuracy and efficiency of prediction.

[0040] Figure 3 This is a graph showing the change in loss values ​​during the training process of different networks. Net 1: No special structures such as residual structures are used; it consists of ordinary convolutional layers and pooling layers. Net 2: A deep network using special structures such as residual structures and dropout. Net 3: Compared to Net 2, it does not use the dropout structure.

[0041] Deploy the CNN network model in a suitable format and then integrate the resulting network into a very low frequency (VLF) signal amplitude calculation system for rapid evaluation of VLF signal amplitude distribution. For example, store the trained CNN network model parameters in Torch Script format, and port the code using the LibTorch extension package compatible with the C / C++ environment to enable the deep learning model to run in a C / C++ integrated system. When rapid computation using the CNN network is needed, simply input the corresponding h' and β into the network input port in the specified format to quickly obtain the corresponding amplitude distribution at the output port for subsequent evaluation and analysis.

[0042] The present invention also provides an embodiment of a computer system. The computer system includes a processor and a memory. The memory is used to store non-transitory computer-readable instructions (e.g., one or more computer program modules). The processor is used to execute the non-transitory computer-readable instructions, which, when executed by the processor, can perform one or more steps in the described very low frequency signal amplitude distribution evaluation model training method, or perform prediction of the very low frequency signal amplitude distribution. The memory and the processor can be interconnected via a bus system and / or other forms of connection mechanisms.

[0043] For example, a processor can be a central processing unit (CPU), a graphics processing unit (GPU), or other form of processing unit with data processing and / or program execution capabilities. A processor can be a general-purpose processor or a special-purpose processor, and it can control other components in the computer to perform desired functions.

[0044] For example, memory can include any combination of one or more computer program products, which can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory can include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory can include, for example, read-only memory (ROM), hard disk, erasable programmable read-only memory (EPROM), compact optical disc read-only memory (CD-ROM), USB storage, flash memory, etc. One or more computer program modules can be stored on the computer-readable storage medium, and the processor can run one or more computer program modules to implement various functions of the computer.

[0045] This invention also provides a computer-readable storage medium for storing non-transitory computer-readable instructions. When executed by a computer, these instructions can implement one or more steps in the above-described training method for the very low frequency (VLF) signal amplitude distribution evaluation model, or the prediction of the VLF signal amplitude distribution. When the VLF signal amplitude distribution evaluation model of this invention is implemented in software and sold or used as a standalone product, it can be stored in a computer-readable storage medium. For further details regarding the storage medium, please refer to the corresponding description of memory in computer systems above; further elaboration is not provided here.

[0046] To verify the accuracy of the training results, a set of h' and β was selected from the validation set, and the actual amplitude distribution was calculated using the LWPC physical model, such as... Figure 4 As shown, Figure 4 (a) is h'. Figure 4 (b) is β; Figure 4(c) The calculated amplitude; the selected h' and β are input into the CNN network used for training, and the predicted amplitude distribution is obtained. This is then compared with the results calculated by the LWPC physical model, such as... Figure 5 As shown, Figure 5 (a) is the amplitude calculated by the LWPC model. Figure 5 (b) The results obtained using network prediction. Figure 5 (c) is Figure 5 (a) and Figure 5 (b) Error values ​​between the two sets of results. Comparing the LWPC model and the CNN network model, the LWPC calculation time is 7226s, while the CNN network calculation time is 0.87s. The RMSE of the CNN network simulation results is 0.334. It is evident that this method can achieve high-precision calculation of the spatial distribution of very low frequency amplitudes while significantly reducing computation time, and has important application value in the field of very low frequency propagation observation.

Claims

1. A convolutional neural network model for evaluating the amplitude distribution of very low frequency signals, characterized in that, include: The input layer is configured to receive data on the effective reflection height at the bottom of the ionosphere and the rate of change of electron density with height. The convolutional layer is configured to extract feature information by performing convolution operations on the effective reflection height and the rate of change of electron density with height at the bottom of the ionosphere. The convolutional layer consists of multiple groups of convolutional kernels of different sizes, which are first grouped and then stacked. The grouped convolutions are performed on the effective reflection height and the rate of change of electron density with height at the bottom of the ionosphere, and the residual structure is used to cross some convolutional layers. Pooling layers are configured to compress and extract the feature information extracted by the convolutional layers; A fully connected layer, configured to map extracted feature information to very low frequency signal amplitude; as well as The output layer is configured to output very low frequency signal amplitude.

2. A training method for a convolutional neural network model for evaluating the amplitude distribution of very low frequency signals as described in claim 1, characterized in that, include: A dataset of the amplitude distribution of the transmitted signal from a very low frequency (VLF) transmitter station is established. The dataset consists of several sets of input and output combinations. In each combination, the input includes the effective reflection height at the bottom of the ionosphere and the rate of change of electron density with height, and the output is the corresponding VLF signal amplitude. The dataset is preprocessed to transform the input into an input format that meets the requirements for evaluating convolutional neural network models based on the very low frequency signal amplitude distribution. The preprocessed dataset is input into the very low frequency signal amplitude distribution evaluation convolutional neural network model and trained on it.

3. The method according to claim 2, characterized in that, The effective reflection height at the bottom of the ionosphere and the rate of change of electron density with height were calculated using the LWPC physical model.

4. The method according to claim 3, characterized in that, The amplitude of very low frequency signals is characterized by electron density. The relationship between electron density, effective reflection height at the bottom of the ionosphere, and the rate of change of electron density with height is as follows: in, N e For electron density, z For height, h' The effective reflection height at the bottom of the ionosphere. β This represents the rate at which electron density changes with altitude.

5. The method according to claim 3, characterized in that, The difference between the very low frequency (VLF) signal amplitude calculated by the LWPC physical model and the VLF signal amplitude amplitude predicted by the VLF signal amplitude distribution evaluation convolutional neural network model during training is used as the loss function.

6. The method according to claim 2, characterized in that, During training, dropout is used to enhance the ability of the convolutional neural network model to learn the rate of change of effective reflection height and electron density at the bottom of the ionosphere with height, based on the very low frequency signal amplitude distribution, thus avoiding overfitting.

7. A computer system, characterized in that, The very low frequency signal amplitude distribution evaluation convolutional neural network model of claim 1, including a processor and a memory, is stored in the memory and configured to be executed by the processor.

8. A computer-readable storage medium for storing non-transitory computer-readable instructions, characterized in that, When the non-transitory computer-readable instructions are executed by a computer, the convolutional neural network model for evaluating the amplitude distribution of very low frequency signals as described in claim 1 can be run.

Images