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

Abstract

The invention discloses a very-low-frequency signal amplitude distribution evaluation convolutional neural network model and a training method thereof, and aims to solve the problems of long calculation time and low efficiency of a traditional physical model. The model comprises: an input layer configured to receive effective reflection height of the bottom of an ionized layer and rate data of electron density varying with height; the convolution layer is configured to perform convolution operation on the effective reflection height of the bottom of the ionized layer and the rate data of the electron density changing along with the height to extract feature information; the pooling layer is configured to perform compression extraction on the feature information extracted by the convolution layer; the full connection layer is configured to map the extracted feature information to a very low frequency signal amplitude; and an output layer configured to output the very low frequency signal amplitude. According to the method, the calculation time can be greatly reduced, the accuracy of a calculation result is ensured, and the method has important value in the engineering application field of very low frequency detection.

Description

Technical Field

[0001] The present invention relates to a very low frequency signal amplitude distribution evaluation convolutional neural network model and a training method thereof. Background Art

[0002] The D layer (lower ionosphere) is the partially ionized atmosphere at an altitude of 60-100 km above the ground, with an electron density of 1-1000 cm -3 This space region is not only affected 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 communications, radio wave propagation, marine navigation positioning and space environment monitoring.

[0003] Very Low Frequency (VLF) waves refer to electromagnetic waves with a frequency range of 3-30kHz, mainly originating from thunderstorm activities or artificial VLF stations around the world. Due to its unique frequency characteristics, VLF waves are often bound in the Earth-ionosphere waveguide composed of the ground and the D layer of the ionosphere, which can achieve long-distance propagation, and the typical attenuation coefficient is about ~2-3dB / Mm. The antenna size of artificial VLF stations is usually large, and it is necessary to rely on valleys or open areas of large plains to set up antenna arrays, which is costly. Currently, most of the existing artificial station transmitters in the world are used in the military field, and public information is limited. In engineering applications such as VLF navigation and positioning, it is often necessary to calculate the signal distribution around the VLF transmitting station, but the related calculations are time-consuming. Even if a greatly simplified VLF propagation capability (Long-Wavelength Propagation Capability, LWPC) model is used, the calculation takes tens of minutes, which cannot meet the real-time and high-accuracy engineering application requirements at the same time.

[0004] Deep learning has shown extraordinary ability in revealing deeper potential patterns and automatically extracting features, and can use graphics processing units (GPUs) to achieve high accuracy and fast calculations. Prediction methods based on deep learning have gradually begun to replace traditional models, which usually require a lot of computing time on central processing units (CPUs). Given its powerful nonlinear modeling capabilities, deep learning has received great attention in the field of computational mechanics. Summary of the invention

[0005] Therefore, the present invention proposes a deep learning-based very low frequency signal amplitude distribution rapid evaluation model and its training method, which replaces the traditional very low frequency propagation model and greatly shortens the calculation time while retaining the calculation accuracy to meet the needs of engineering applications.

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

[0007] In some examples, the convolution layer is composed of multiple groups of convolution kernels of different sizes, which are grouped and then superimposed to perform group convolution on the effective reflection height at the bottom of the ionosphere and the rate at which the electron density changes with height. The grouped convolution structure is used to enhance the nonlinearity of the network and improve the effectiveness of feature extraction for h' (electron reflection height) and β (sharpness parameter, the rate at which the electron density changes with height).

[0008] In some examples, a residual structure is used to span some convolutional layers to deepen the depth of the VLF signal amplitude distribution evaluation convolutional neural network model while ensuring its stability to solve the problem of gradient disappearance.

[0009] In the second aspect, a training method for the very low frequency signal amplitude distribution evaluation convolutional neural network model is proposed, including: establishing a very low frequency transmitting station transmitted signal amplitude distribution data set, the data set is a plurality of input and output combinations, in each combination, the input includes the effective reflection height of the bottom of the ionosphere and the rate of change of the electron density with the height, and the output is the corresponding very low frequency signal amplitude; preprocessing the data set to convert the input into an input format that meets the input format requirements of the very low frequency signal amplitude distribution evaluation convolutional neural network model; inputting the preprocessed data set into the very low frequency signal amplitude distribution evaluation convolutional neural network model to train it.

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

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

[0012] In some examples, dropout (special network layer, dropout layer) is used during training to enhance the ability of the very low frequency signal amplitude distribution evaluation convolutional neural network model to learn the effective reflection height at the bottom of the ionosphere and the rate data of change of electron density with altitude, enhance the robustness and generalization of the network model, effectively reduce model parameters and complexity, and avoid overfitting.

[0013] In a third aspect, a computer system is proposed, comprising 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] In a fourth aspect, a computer-readable storage medium is proposed for storing non-temporary computer-readable instructions, which, when executed by a computer, can run the very low frequency signal amplitude distribution evaluation convolutional neural network model.

[0015] The present invention uses a deep learning network to replace the LWPC model to accelerate the calculation of the signal distribution of very low frequency stations. The LWPC model is used to calculate and construct a very low frequency amplitude distribution data set, build a deep CNN network, pre-process the data set to adapt to the input and output form of the CNN network, and the data set is shuffled and divided into a training set and a verification set according to a proportion, and input into the CNN network to obtain the deep feature information of the data, which is connected with the output data to realize the mapping relationship between the input data and the output data; the trained CNN network model is used for prediction, and the original physical model calculation problem is converted into a prediction problem. Through the iterative output of the CNN network model, the model characteristics of the CNN network can be more efficiently utilized to improve the accuracy and efficiency of the prediction; finally, the CNN network model is deployed and exported in a suitable format and integrated into the application system. The present invention can greatly reduce the calculation time and ensure the accuracy of the calculation results, which is of great value in the engineering application field of very low frequency detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 4 is a schematic diagram of a convolutional neural network model for evaluating very low frequency signal amplitude distribution according to an embodiment of the present invention.

[0017] Figure 2 It is a flow chart of a training method for a very low frequency signal amplitude distribution evaluation convolutional neural network model according to an embodiment of the present invention.

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

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

[0020] Figure 5 is a comparison diagram of the prediction results using the trained network and the results using the LWPC model according to an embodiment of the present invention, Figure 5 (a) is the amplitude calculated by LWPC model, Figure 5 (b) is the result obtained using network prediction. Figure 5 (c) Figure 5 (a) with Figure 5 (b) The error value between the two sets of results. DETAILED DESCRIPTION

[0021] Firstly, a very low frequency signal amplitude distribution evaluation model based on deep learning (referred to as the "model") is introduced. It transforms the calculation problem of the physical model of very low frequency propagation into a prediction problem, that is, it transforms the calculation relationship between the input parameters of the effective reflection height (h') of the bottom of the ionosphere and the rate of change of electron density with height (β) and the output result amplitude into a mapping relationship of the network.

[0022] like Figure 1 As shown in the figure, the VLF signal amplitude distribution evaluation model adopts a deep convolutional neural network (CNN), including convolutional layers, pooling layers, and fully connected layers. For example, the convolutional layers, pooling layers, and fully connected layers can accumulate to 112 layers to form a deep network. In addition, the residual structure, grouped convolution structure, and dropout structure are combined to enhance the network's ability to learn the features of h' and β. The residual structure is introduced to connect the front layer and the back layer through shortcuts, spanning several convolution layers in the middle, so as to achieve the stability of the network while deepening the network and solve the problem of gradient disappearance. The grouped convolution structure is adopted, and the network structure formed by grouping and superimposing multiple groups of convolution kernels of different sizes is used to enhance the nonlinearity of the network and improve the effectiveness of h' and β feature extraction; after the parameter feature extraction is completed, the dropout structure is added to discard some network feature parameters to solve the problem of overfitting of the deep network. In the training of the model, the root mean square error (RMSE) is selected as the loss function of the LWPC physical model learning regression problem, that is, the error between the amplitude predicted by the calculation model and the amplitude calculated by the LWPC physical model is used as the loss function. The optimizer is Adam, and the learning rate is selected to use the learning rate function lr (epoch) to adapt to the trend of parameter changes in the learning process.

[0023]

[0024] Figure 2 This is a flow chart of the training method of the VLF signal amplitude distribution evaluation model based on deep learning. The training method is described in detail below.

[0025] Step 1: Use the Long-Wavelength Propagation Capability (LWPC) model to simulate and calculate the amplitude distribution of different VLF stations at different times, and establish a VLF transmitting station signal amplitude distribution data set.

[0026] Step 11: Input the time when the very low frequency station transmits the signal and the frequency, power and other parameters of the station 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 the VLF station. The main influencing factor of the amplitude is the electron density. Through the formula (1) proposed by Wait and Spies, the electron density and h' are connected with β, so as to realize the mapping relationship between the amplitude and h' and β.

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

[0029] Among them, N e is the electron density, z is the altitude, h' is the effective reflection altitude at the bottom of the ionosphere, and β is the rate at which the electron density changes with altitude.

[0030] Step 13: The data set is a number 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 data set, shuffle the data set and divide it into training set and validation set in proportion to enhance the generalization ability of the model and speed up the training rate.

[0032] Step 21, preprocessing h' and β data sets, converting network input h' and β into a two-dimensional array to meet the input format requirements of the CNN network;

[0033] Step 22: Randomly shuffle the processed data set to enhance the generalization ability of the model and speed up the rate of training gradient descent;

[0034] Step 23. Divide the data into a data set and a validation set in a ratio of 8:2 to test the model configuration and the quality of the training results, and to verify whether the results are overfitting or underfitting, so as to make adjustments and retrain.

[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, and 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 training times (Epoch) and the maximum number of input samples (Batch size);

[0037] Step 32: The convolution layer of the CNN network obtains the feature information of h' and β during the training process, introduces nonlinearity through the activation function, and finally reduces the complexity of the feature information through the pooling layer to extract the main features.

[0038] Step 33: The input data h' and β features are transformed into one dimension through the fully connected layer to complete the target feature regression task; the network and the output data amplitude are connected through the output layer to realize the mapping conversion relationship between the input h', β and the output amplitude data.

[0039] The trained CNN network model is used to predict the very low frequency amplitude, converting the original physical model calculation problem into a prediction problem. Through model iterative output, the model characteristics of the CNN network can be more efficiently utilized to improve the accuracy and efficiency of the prediction.

[0040] Figure 3 This is a graph of the loss value changes during the training process of different networks. Net 1: does not use special structures such as residual structures, and is composed of ordinary convolutional layers and pooling layers; Net 2: uses a deep network with special structures such as residual structures and dropout; Net 3: does not use the dropout structure compared to Net 2.

[0041] Select the appropriate format to deploy the CNN network model and bring it out, and further integrate the obtained network into the calculation system of the very low frequency signal amplitude for the rapid evaluation of the very low frequency signal amplitude distribution. For example, store the trained CNN network model parameters in the form of Torch Script, select the LibTorch expansion package that conforms to the C / C++ environment for code transplantation, and realize the operation of the deep learning model in the C / C++ integrated system. When the CNN network needs to be used for fast calculation, just input the corresponding h' and β in the specified format to the network input port, and the corresponding amplitude distribution can be quickly obtained 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-temporary computer-readable instructions (e.g., one or more computer program modules). The processor is used to run the non-temporary computer-readable instructions, and when the non-temporary computer-readable instructions are run by the processor, one or more steps in the very low frequency signal amplitude distribution evaluation model training method can be executed, or a very low frequency signal amplitude distribution prediction can be performed. The memory and the processor can be interconnected via a bus system and / or other forms of connection mechanisms.

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

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

[0045] The present invention also provides a computer-readable storage medium for storing non-temporary computer-readable instructions. When the non-temporary computer-readable instructions are executed by a computer, one or more steps in the above-mentioned very low frequency signal amplitude distribution evaluation model training method or the prediction of very low frequency signal amplitude distribution can be implemented. When the very low frequency signal amplitude distribution evaluation model of the present invention is implemented in the form of software and sold or used as an independent product, it can be stored in a computer-readable storage medium. For relevant descriptions of the storage medium, reference can be made to the corresponding description of the memory in the computer system above, which will not be repeated here.

[0046] In order to verify the accuracy of the training results, a set of h' and β is selected from the validation set, and the actual amplitude distribution is obtained using the LWPC physical model. Figure 4 As shown, Figure 4 (a) is h', Figure 4 (b) is β; Figure 4(c) is the calculated amplitude; the selected h' and β are input into the CNN network of the training network to obtain the predicted amplitude distribution, which is compared with the results calculated by the LWPC physical model, as shown in Figure 5 As shown, Figure 5 (a) is the amplitude calculated by LWPC model, Figure 5 (b) is the result obtained using network prediction. Figure 5 (c) Figure 5 (a) with Figure 5 (b) The error between the two sets of results. Comparing the LWPC model and the CNN network model, the calculation time required for LWPC is 7226s, the time required for CNN network is 0.87s, and the RMSE of the simulation results of CNN network is 0.334. It can be seen that this method can achieve high-precision calculation of the spatial distribution of VLF amplitude while greatly reducing the calculation time, and has important application value in the field of VLF propagation observation.

Claims

1. A convolutional neural network model for very low frequency signal amplitude distribution assessment, characterized in that: include: An input layer configured to accept data on the effective reflection height at the bottom of the ionosphere and the rate at which the electron density changes with height; A convolution layer, which is configured to perform a convolution operation on the effective reflection height at the bottom of the ionosphere and the rate data of the electron density changing with the height to extract feature information; A pooling layer, which is configured to compress and extract feature information extracted by the convolutional layer; A fully connected layer configured to map the extracted feature information to a very low frequency signal amplitude; as well as The output layer is configured to output the VLF signal amplitude.

2. The very low frequency signal amplitude distribution evaluation convolutional neural network model according to claim 1, characterized in that: The convolution layer is composed of multiple groups of convolution kernels of different sizes, which are first grouped and then superimposed. The effective reflection height at the bottom of the ionosphere and the rate data of electron density changing with height are grouped and convolved to enhance the nonlinearity of the very low frequency signal amplitude distribution evaluation convolution neural network model.

3. The very low frequency signal amplitude distribution evaluation convolutional neural network model according to claim 2, characterized in that: The residual structure is utilized to cross some convolutional layers, and the depth of the VLF signal amplitude distribution evaluation convolutional neural network model is deepened while ensuring its stability to solve the problem of gradient disappearance.

4. A method for training a very low frequency signal amplitude distribution evaluation convolutional neural network model according to any one of claims 1 to 3, characterized in that: include: Establish a data set of amplitude distribution of signals transmitted by a very low frequency transmitting station. The data set consists of several sets of input and output combinations. In each set, the input includes the effective reflection height at the bottom of the ionosphere and the rate at which the electron density changes with the height, and the output is the corresponding very low frequency signal amplitude. Preprocessing the data set to convert the input into an input format that meets the input format requirements of the very low frequency signal amplitude distribution assessment convolutional neural network model; The preprocessed data set is input into the very low frequency signal amplitude distribution evaluation convolutional neural network model for training.

5. The method according to claim 4, characterized in that 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 at which the electron density changes with altitude.

6. The method according to claim 5, characterized in that The electron density is used to characterize the amplitude of the very low frequency signal. The relationship between the electron density and the effective reflection height at the bottom of the ionosphere and the rate at which the electron density changes with height is as follows: N e (z)=1.43×10 13 exp(-0.15h)exp(β(z-h')) Among them, N e is the electron density, z is the altitude, h' is the effective reflection altitude at the bottom of the ionosphere, and β is the rate at which the electron density changes with altitude.

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

8. The method according to claim 4, characterized in that Dropout is used during training to enhance the ability of the very low frequency signal amplitude distribution evaluation convolutional neural network model to learn the effective reflection height at the bottom of the ionosphere and the rate data of electron density changing with altitude, so as to avoid overfitting.

9. A computer system, characterized in that: It comprises a processor and a memory, wherein the very low frequency signal amplitude distribution evaluation convolutional neural network model described in any one of claims 1 to 3 is stored in the memory and is configured to be executed by the processor.

10. 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 very low frequency signal amplitude distribution evaluation convolutional neural network model described in any one of claims 1 to 3 can be run.

Images