Artificial Intelligence-Based Lightning Electromagnetic Pulse Waveform Compression Method and System

Abstract

The present invention belongs to the fields of lightning waveform compression, artificial intelligence, and deep learning, and specifically relates to a lightning electromagnetic pulse waveform compression method and system based on artificial intelligence, aiming to solve the problems that existing data compression methods cannot process various types of lightning electromagnetic pulse waveform signals with large differences, and are incompatible between reducing compression calculation complexity and increasing the compression ratio. The present invention includes: constructing a lightning data compression model based on an autoencoder, the model including an input layer, an encoding compression module, a variable ratio adjustment module, a decoding reconstruction module, and an output layer; training the model based on various types of lightning electromagnetic pulse waveform data; adjusting the model parameters through the loss between the original waveform and the reconstructed waveform; and compressing the real-time lightning electromagnetic pulse waveform data through the trained model, with the compression ratio controlled by the variable ratio adjustment module. The present invention can achieve high-quality compression of various types of lightning electromagnetic pulse waveform signals, and the compression ratio is flexibly adjustable.

Description

Background Art

[0002] Lightning is a spectacular and shocking natural phenomenon. China is located in the eastern part of Asia with a significant monsoon climate. Lightning occurs frequently in summer. In today's era of rapid technological development, accidents caused by lightning still occur from time to time. Therefore, the research on lightning is particularly important. However, continuous detection and data research on lightning have led to an increasingly large amount of data. Based on the current huge amount of lightning electromagnetic pulse waveform data, a data compression technology is particularly important.

[0003] With the progress of technology, artificial intelligence has returned to the public eye after several rises and falls and has become a major hot field. Especially in the field of deep learning of artificial intelligence, models based on neural networks, due to their brain-like structure with great flexibility, have super learning ability when facing different data. They can not only learn the internal laws of different data, but also their learning ability will play different roles according to the different structures of the built models.

[0004] Therefore, due to the powerful data representation ability of neural network models based on artificial intelligence, more and more research applies artificial intelligence to the compression of lightning electromagnetic pulse waveforms. However, there are many types of lightning electromagnetic pulse signals, and existing methods are still difficult to have good compression and decompression performance for various types of lightning electromagnetic pulse waveform signals with large differences. In addition, with the increase of the data compression ratio, the computational complexity of the model also increases, and existing methods cannot be compatible between reducing the compression computational complexity and increasing the compression ratio. Summary of the Invention

[0005] To solve the above problems in the prior art, that is, existing data compression methods cannot process various types of lightning electromagnetic pulse waveform signals with large differences and cannot be compatible between reducing the compression computational complexity and increasing the compression ratio, the present invention provides a lightning electromagnetic pulse waveform compression method based on artificial intelligence. The method includes:

[0006] Step S10, collecting a set number of lightning electromagnetic pulse waveform data through a lightning radiation field signal acquisition device as a model training data set;

[0007] Step S20, constructing a lightning data compression model based on a stacked autoencoder; the lightning data compression model includes an input layer, an encoding compression module, a variable ratio adjustment module, a decoding reconstruction module, and an output layer;

[0008] Step S30, selecting a training data from the model training data set, inputting the data into the current lightning data compression model, and obtaining the compressed data output by the encoding compression module and the reconstructed data output by the decoding reconstruction module;

[0009] Step S40, calculate the loss value between the compressed data and the reconstructed data, and update the parameters of the lightning data compression model by optimizing the function using the stochastic gradient descent method, serving as the current lightning data compression model;

[0010] Step S50, repeat steps S30 - S40 for model iterative training until the set training end condition is reached, obtaining the trained lightning data compression model;

[0011] Step S60, compress the real - time collected lightning electromagnetic pulse waveform data through the trained lightning data compression model.

[0012] In some preferred embodiments, the encoding and compression module includes a multi - layer one - dimensional convolutional neural network and a max - pooling layer, and the decoding and reconstruction module includes a one - dimensional transposed convolutional neural network corresponding one - to - one with the multi - layer one - dimensional convolutional neural network and a max - upsampling layer corresponding to the max - pooling layer.

[0013] In some preferred embodiments, the one - dimensional convolutional neural network is expressed as:

[0014]

[0015] where, \(X\) j is the \(j\) - th output feature vector of the output feature \(X\) of the one - dimensional convolutional neural network, \(S\) i is the \(i\) - th input feature vector of the input feature \(S\), \(K\) ij is the convolutional kernel vector in the convolutional kernel \(K\) of the one - dimensional convolutional neural network that connects the \(i\) - th input feature vector and the \(j\) - th output feature vector, \(b\) j is the \(j\) - th bias vector of the bias \(b\) of the one - dimensional convolutional neural network, \(f(·)\) represents the non - linear activation calculation, and \(M\) j is the number of mappings connecting the input feature and the output feature.

[0016] In some preferred embodiments, the max - pooling layer is expressed as:

[0017]

[0018] where, \(P\) j (n) is the \(n\) - th index value of the \(j\) - th output feature vector of the output feature \(P\) of the max - pooling layer, \(R\) represents the size of the pooling window, \(r\) represents the element index of the pooling window, \(L\) represents the stride of the pooling, and \(max[·]\) represents the maximum value calculation.

[0019] In some preferred embodiments, the one - dimensional transposed convolutional neural network is expressed as:

[0020]

[0021] Among them, Y j is the j-th output feature vector of the output feature Y of the one-dimensional inverse convolutional neural network, and T represents the transpose calculation.

[0022] In some preferred embodiments, the maximum upsampling layer is expressed as:

[0023]

[0024] Among them, U j (n×L+r) is the (n×L+r)-th index value of the j-th output feature vector of the output feature U of the maximum upsampling layer, and S i (n) is the n-th index value of the i-th input feature vector of the input feature S.

[0025] In some preferred embodiments, the variable ratio adjustment module includes a set number of residual blocks connected in sequence, a fully connected layer, and a linear activation layer;

[0026] The residual block includes a set number of fully connected layers and a Tanh activation layer.

[0027] In some preferred embodiments, the linear activation layer is expressed as:

[0028] A j = S i × (K ij ) T + b j

[0029] Among them, A j is the j-th output feature vector of the output feature A of the linear activation layer.

[0030] In some preferred embodiments, the Tanh activation layer is expressed as:

[0031]

[0032] Among them, T j (n) is the n-th index value of the j-th output feature vector of the output feature T of the Tanh non-linear activation layer.

[0033] On the other hand, the present invention proposes an artificial intelligence-based lightning electromagnetic pulse waveform compression system, which includes the following modules:

[0034] A data acquisition module configured to collect a set number of lightning electromagnetic pulse waveform data through a lightning radiation field signal acquisition device as a model training data set; and to collect lightning electromagnetic pulse waveform data in real time as data to be compressed;

[0035] A model construction module, configured to construct a lightning data compression model based on a stacked autoencoder; the lightning data compression model includes an input layer, an encoding and compression module, a variable ratio adjustment module, a decoding and reconstruction module, and an output layer;

[0036] A model training module, configured to select a training data from the model training dataset, input the data into the current lightning data compression model, obtain the compressed data output by the encoding and compression module and the reconstructed data output by the decoding and reconstruction module, calculate the loss value between the compressed data and the reconstructed data, and update the parameters of the lightning data compression model by optimizing the function using the stochastic gradient descent method, as the current lightning data compression model, perform model iterative training until the set training end condition is reached, and obtain a trained lightning data compression model;

[0037] A data compression module, configured to compress the data to be compressed collected in real time through the trained lightning data compression model.

[0038] Advantages of the present invention:

[0039] (1) The lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention constructs a lightning data compression model based on a stacked autoencoder, uses the model to extract features and encode the lightning electromagnetic pulse waveform data. The length of each originally collected waveform is 1000 floating-point data. After feature extraction and encoding compression, the compression ratio can reach 6.4% and below, and the highest compression ratio can reach 0.1% (that is, using 1 floating-point number to represent the original signal).

[0040] (2) The lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention sets a decoding and reconstruction module corresponding to the structure of the encoding and compression module in the model. The peak points of the reconstructed waveform are clear, the rising and falling edges are steep, the waveform shape is complete, the fitting effect is good compared with the original waveform, and the waveform loss is small.

[0041] (3) The lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention can achieve good results in data encoding compression and decoding reconstruction for single-category lightning electromagnetic pulse waveform data and waveform data of multiple types such as mixed cloud flashes and ground flashes.

[0042] (4) The lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention sets a variable ratio adjustment module in the model, which can realize multi-level compression ratio adjustment. When using 5 residual blocks as the variable ratio adjustment module, 7-level decreasing of the compression ratio can be realized, that is, the model data compression ratios are 6.4%, 3.2%, 1.6%, 0.8%, 0.4%, 0.2%, and 0.1% respectively. Description of the Drawings

[0043] Other features, objects, and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments read in conjunction with the accompanying drawings:

[0044] Figure 1 is a schematic flowchart of the method for compressing lightning electromagnetic pulse waveforms based on artificial intelligence of the present invention

[0045] Figure 2 is a schematic flowchart of the working process of the autoencoder in an embodiment of the method for compressing lightning electromagnetic pulse waveforms based on artificial intelligence of the present invention;

[0046] Figure 3 is a schematic flowchart of waveform compression and reconstruction in an embodiment of the method for compressing lightning electromagnetic pulse waveforms based on artificial intelligence of the present invention;

[0047] Figure 4 is a structural diagram of the encoding compression module and the decoding reconstruction module in an embodiment of the method for compressing lightning electromagnetic pulse waveforms based on artificial intelligence of the present invention;

[0048] Figure 5 is a structural diagram of the variable ratio adjustment module in an embodiment of the method for compressing lightning electromagnetic pulse waveforms based on artificial intelligence of the present invention;

[0049] Figure 6 is a comparison diagram of the original waveforms and reconstructed waveforms of various categories in an embodiment of the method for compressing lightning electromagnetic pulse waveforms based on artificial intelligence of the present invention;

[0050] Figure 7 is a mean squared error (MSE) waveform diagram of applying the model to unclassified data in an embodiment of the method for compressing lightning electromagnetic pulse waveforms based on artificial intelligence of the present invention;

[0051] Figure 8 is a comparison diagram of the original waveform and the reconstructed waveform of applying the model to unclassified data in an embodiment of the method for compressing lightning electromagnetic pulse waveforms based on artificial intelligence of the present invention. Detailed Embodiments

[0052] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only for explaining the relevant invention and not for limiting the invention. Additionally, it should be noted that for the sake of description, only parts related to the relevant invention are shown in the drawings.

[0053] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0054] The present invention provides an artificial intelligence-based lightning electromagnetic pulse waveform compression method, which can achieve the purpose of waveform compression for existing lightning electromagnetic pulse waveform data, and can reconstruct (decompress) the original waveform from the compressed data to achieve the purpose of reducing storage space. It solves the problem that existing data compression technologies do not all have good compression and decompression performance in compressing a wide variety of lightning electromagnetic pulse signals with large differences, and cannot be compatible between reducing compression calculation complexity and increasing the compression ratio.

[0055] An artificial intelligence-based lightning electromagnetic pulse waveform compression method of the present invention, the method comprising:

[0056] Step S10, collecting a set number of lightning electromagnetic pulse waveform data through a lightning radiation field signal acquisition device as a model training data set;

[0057] Step S20, constructing a lightning data compression model based on a stacked autoencoder; the lightning data compression model includes an input layer, an encoding compression module, a variable ratio adjustment module, a decoding reconstruction module, and an output layer;

[0058] Step S30, selecting a training data from the model training data set, inputting the data into the current lightning data compression model, and obtaining the compressed data output by the encoding compression module and the reconstructed data output by the decoding reconstruction module;

[0059] Step S40, calculating the loss value between the compressed data and the reconstructed data, and updating the parameters of the lightning data compression model by a stochastic gradient descent method optimization function as the current lightning data compression model;

[0060] Step S50, repeating steps S30 - S40 for model iterative training until a set training end condition is reached to obtain a trained lightning data compression model;

[0061] Step S60, compressing the lightning electromagnetic pulse waveform data collected in real time through the trained lightning data compression model.

[0062] For a clearer description of the artificial intelligence-based lightning electromagnetic pulse waveform compression method of the present invention, the following is a detailed description of each step in Figure 1 each embodiment of the present invention.

[0063] The artificial intelligence-based lightning electromagnetic pulse waveform compression method of the first embodiment of the present invention includes steps S10 - S60, and each step is described in detail as follows:

[0064] Step S10, collecting a set number of lightning electromagnetic pulse waveform data through a lightning radiation field signal acquisition device as a model training data set.

[0065] In one embodiment of the present invention, existing real-time very low frequency / low frequency (VLF / LF) lightning radiation field signal acquisition equipment is used to collect one-dimensional lightning electromagnetic pulse waveform data. The single sampling time length of the equipment is 1 ms, and the sampling rate is 1 MSPS, that is, the number of sampling points for each sample data is 1000.

[0066] The lightning electromagnetic pulse waveform data used as the model training dataset are a total of 131,965 pieces of data collected by the three-dimensional lightning detection network of the Institute of Electrical Engineering, Chinese Academy of Sciences, including two categories: cloud flashes and cloud-to-ground flashes. Specifically, it includes 11,450 negative cloud-to-ground flashes (-CG), 18,596 ordinary cloud flashes (IC), 5,659 positive cloud-to-ground flashes (+CG), 5,002 negative bipolar narrow pulses (-NBE), and 10,206 positive bipolar narrow pulses (+NBE).

[0067] Step S20: Construct a lightning data compression model based on a stacked autoencoder; the lightning data compression model includes an input layer, an encoding and compression module, a variable ratio adjustment module, a decoding and reconstruction module, and an output layer.

[0068] As an artificial neural network model, the autoencoder can learn the latent representation of unlabeled data, which is a typical unsupervised learning method. The structure of the autoencoder is divided into two parts: an encoder and a decoder. The main task of the encoder part is to compress the input data in terms of features, and the task of the decoder part is to reconstruct the original data based on the features obtained after compression by the encoder.

[0069] As Figure 2 shown, it is a schematic diagram of the working process of the autoencoder in one embodiment of the lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention. The encoder transforms the input signal x into the encoded signal y, and the decoder converts the encoded signal y into the output signal The conversion process formula is as follows. In the formula, f(·) represents the encoding rule function of the encoder, g(·) represents the decoding rule function of the decoder, and the encoding and decoding process is as shown in Equation (1):

[0070]

[0071] The simplest autoencoder consists of only three layers: an input layer, an output layer, and a hidden layer. The stacked autoencoder adds multiple hidden layers on the basis of the autoencoder. Compared with the traditional autoencoder, the stacked autoencoder has stronger data representation ability.

[0072] As Figure 3As shown in the figure, it is a schematic flowchart of waveform compression and reconstruction of an embodiment of the lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention. The left side represents the input waveform data, and the right side represents the output waveform data. Ideally, the input and output waveforms should be exactly the same. The left side of the i-th hidden layer of the middle stacked autoencoder model is the encoder part, and the right side is the decoder part. The input layer on the left side inputs the lightning electromagnetic pulse waveform data. The input waveform data is encoded and compressed layer by layer through the encoder part of the stacked autoencoder. Subsequently, the hidden layer of the decoder part on the right decodes and reconstructs the compressed waveform data. Finally, the i of the middle hidden layer is obtained as the final compression result.

[0073] As Figure 4 shown, it is a structural diagram of the encoding and compression module and the decoding and reconstruction module of an embodiment of the lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention. The encoding and compression module includes a multi-layer one-dimensional convolutional neural network and a max pooling layer. The decoding and reconstruction module includes a one-dimensional transposed convolutional neural network corresponding to the multi-layer one-dimensional convolutional neural network one by one and a max upsampling layer corresponding to the max pooling layer.

[0074] The one-dimensional convolutional neural network is represented as shown in Equation (2):

[0075]

[0076] where X j is the j-th output feature vector of the output feature X of the one-dimensional convolutional neural network, S i is the i-th input feature vector of the input feature S, K ij is the convolutional kernel vector in the convolutional kernel K of the one-dimensional convolutional neural network that connects the i-th input feature vector and the j-th output feature vector, b j is the j-th bias vector of the bias b of the one-dimensional convolutional neural network, f(·) represents the non-linear activation calculation, and M j is the number of mappings connecting the input feature and the output feature.

[0077] The max pooling layer is represented as shown in Equation (3):

[0078]

[0079] where P j (n) is the n-th index value of the j-th output feature vector of the output feature P of the max pooling layer, R represents the size of the pooling window, r represents the element index of the pooling window, L represents the step size of the pooling, and max[·] represents the maximum value calculation.

[0080] The one-dimensional transposed convolutional neural network is represented as shown in Equation (4):

[0081]

[0082] Among them, Y is the j-th output feature vector of the output feature Y of the one-dimensional inverse convolutional neural network, and T represents the transpose calculation. The one-dimensional inverse convolutional neural network corresponds one-to-one with the one-dimensional convolutional neural network, and their convolutional kernels and biases are the same.

[0083] The max pooling upsampling layer is expressed as shown in Equation (5):

[0084]

[0085] Among them, U j (n×L+r) is the (n×L+r)-th index value of the j-th output feature vector of the output feature U of the max pooling upsampling layer, and S i (n) is the n-th index value of the i-th input feature vector of the input feature S.

[0086] As Figure 5 shown, it is the structural diagram of the variable ratio adjustment module of an embodiment of the lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention. The variable ratio adjustment module includes a set number of residual blocks, a fully connected layer, and a linear activation layer connected in sequence. The residual block includes a set number of fully connected layers and a Tanh activation layer. In this embodiment, 5 residual blocks are set in the variable ratio adjustment module, and 3 fully connected layers are set in each residual block, which can realize a 7-level decrease in the compression ratio, that is, the model data compression ratios are 6.4%, 3.2%, 1.6%, 0.8%, 0.4%, 0.2%, and 0.1% respectively.

[0087] The linear activation layer is expressed as shown in Equation (6):

[0088] A j = S i × (K ij ) T + b j (6)

[0089] Among them, A j is the j-th output feature vector of the output feature A of the linear activation layer.

[0090] The Tanh activation layer is expressed as shown in Equation (7):

[0091]

[0092] Among them, T j (n) is the n-th index value of the j-th output feature vector of the output feature T of the Tanh non-linear activation layer.

[0093] Step S30: Select a training data from the model training dataset, input the data into the current lightning data compression model, and obtain the compressed data output by the encoding and compression module and the reconstructed data output by the decoding and reconstruction module.

[0094] Step S40: Calculate the loss value between the compressed data and the reconstructed data, and update the parameters of the lightning data compression model through the random gradient descent optimization function as the current lightning data compression model.

[0095] Step S50: Repeat steps S30 - S40 for model iterative training until the set training end condition is reached, and obtain the trained lightning data compression model.

[0096] During the training process of the model, the selection of the loss function and the optimizer directly determines the final training effect of the model. The loss function judges the quality of the model parameters by comparing the gap between the model output and the true output, and the optimizer is used to update the network parameters to make the model parameters optimal. After debugging and training, the present invention uses the random gradient descent optimization function to update the model parameters. The calculation method of the loss function of random gradient descent is shown in Equation (8):

[0097]

[0098] where N is the number of data samples included in the training batch, D n is the nth data in a batch, and F(D n ) represents the reconstructed signal output after D n goes through the encoding and decoding processes, and L is the calculated loss value.

[0099] Let represent the matrix or vector of the trainable parameters in the model, such as the weights and biases in each layer of the model. The update rule of the parameters during the model training process can be expressed as Equation (9):

[0100]

[0101] where and respectively represent the matrix or vector of the trainable parameters before and after the update, η represents the learning rate, is the gradient operator, and t represents the number of training times. In an embodiment of the present invention, the learning rate η is set to 0.01, the number of training times t is 100 times, and the batch size for each iteration is 64.

[0102] The present invention uses the mean square error (MSE) loss function to evaluate the model training process. The smaller the MSE, the closer the predicted value is to the true value, that is, the better the fitting effect of the reconstructed waveform. The calculation method of the mean square error is shown in Equation (10):

[0103]

[0104] Among them, y i represents the true value, and represents the predicted value. In special cases, when the MSE is 0, it means that the reconstructed waveform is exactly the same as the original waveform.

[0105] In an embodiment of the present invention, using the above stacked autoencoder model and training rules, the data uses lightning electromagnetic pulse data, and the model is built based on the Keras deep learning tool on the Python language platform. Due to the large amount of data, the running speed on the CPU is slow. The present invention runs the data with the help of a GPU, and the NVIDIA Quadro RTX 4000 is used.

[0106] Step S60, through the trained lightning data compression model, compress the real-time collected lightning electromagnetic pulse waveform data.

[0107] The present invention conducts experiments on various types of lightning waveform data sets. The lightning electromagnetic pulse waveform compression and reconstruction model and the corresponding training rules are used to compress and train the lightning electromagnetic pulse waveform data set. The compression ratio of the model is set to 0.1%, and the mean square error MSE is used as the evaluation index for the final result of the model training compression. The number of rounds (epochs) of the model training is set to 100, and the data set is divided into a training set and a validation set, and the ratio of the two is 9:1. Finally, after training the model with the training set, the validation set is used to verify the data. The mean square error results of the experiment are shown in Table 1:

[0108] Table 1 Mean Square Error MSE Results

[0109]

[0110]

[0111] In addition to evaluating the experimental results using the above indicators, the present invention also visually compares the waveform reconstructed from the compressed data with the original waveform, and intuitively shows the final effect by checking the fitting degree between the reconstructed waveform and the original waveform. As Figure 6 shown, it is a comparison diagram of the original waveforms and reconstructed waveforms of various categories in an embodiment of the lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention. The upper waveform on the left side of (a), (b), (c), (d), and (e) in the figure is a piece of original waveform data in the lightning data, and the lower waveform on the left side is the waveform reconstructed from the feature data after model compression. The right figure is the effect diagram of superimposing the original waveform and the reconstructed waveform. In Figure 6In [Figure 0], it can be seen that the similarity between the original waveform and the reconstructed waveform is extremely high. This similarity also proves that the lightning electromagnetic pulse waveform compression and reconstruction model (i.e., the lightning data compression model) of the present invention can reconstruct waveform data from the internal layer with a reduced number of compression nodes. In addition, from Figure 6 As can be seen from [Figure 1], after compressing each piece of data from 1000 floating-point numbers to 1 floating-point number, the peak points of the waveform reconstructed from the compressed data are clear, the rising and falling edges are steep, and the waveform shape is complete without fragment loss, indicating that the lightning electromagnetic pulse waveform compression and reconstruction model of the present invention can achieve the expected effect for the compression of these types of waveforms.

[0112] When the model of the present invention is applied to unclassified data, the mean square error MSE between its original waveform and the reconstructed waveform is 0.0925. As Figure 7 shown, it is the MSE waveform diagram of the model of an embodiment of the lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention applied to unclassified data.

[0113] As Figure 8 shown, it is the comparison diagram of the original waveform and the reconstructed waveform of the model of an embodiment of the lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention applied to unclassified data. In the figure, the upper waveform on the left side of (a) and (b) is an original waveform data in the lightning data, the lower waveform on the left side is the waveform reconstructed from the feature data after model compression, and the right-side figure is the effect diagram of superimposing the original waveform and the reconstructed waveform together.

[0114] In summary, the lightning electromagnetic pulse waveform compression method based on artificial intelligence of the present invention is based on the lightning electromagnetic pulse data collected by the three-dimensional lightning detection network of the Institute of Electrical Engineering, Chinese Academy of Sciences, as the training data for the lightning electromagnetic pulse waveform compression and reconstruction model, and uses the reconstructed waveform to evaluate the compression effect. It can compress waveform data, and the optimal compression ratio can reach 0.1%. Finally, the fitting effect between the original waveform and the reconstructed waveform is evaluated. The minimum value of the evaluation index MSE for single-category data training is 0.0188, and the evaluation index MSE for mixed multi-category data training is 0.0925. And from the comparison diagram of the original waveform and the reconstructed waveform, it can be seen that the shape of the reconstructed waveform is complete without the loss of important features. The present invention provides an artificial intelligence method for the compression of lightning data. Compared with the previous methods, it can compress data in a larger proportion and has achieved good results for different categories of data, greatly improving the efficiency of data transmission. In addition, the compression method of the present invention can also adjust the compression ratio, achieving the compatibility between reducing the compression calculation complexity and increasing the compression ratio.

[0115] Although the steps in the above embodiments are described in the above order, those skilled in the art can understand that in order to achieve the effects of this embodiment, different steps do not have to be executed in such an order. They can be executed simultaneously (in parallel) or in a reversed order, and these simple changes are all within the protection scope of the present invention.

[0116] The lightning electromagnetic pulse waveform compression system based on artificial intelligence according to the second embodiment of the present invention includes the following modules:

[0117] A data acquisition module configured to collect a set number of lightning electromagnetic pulse waveform data through a lightning radiation field signal acquisition device as a model training data set; and to collect lightning electromagnetic pulse waveform data in real time as data to be compressed;

[0118] A model construction module configured to construct a lightning data compression model based on a stacked autoencoder; the lightning data compression model includes an input layer, an encoding compression module, a variable ratio adjustment module, a decoding reconstruction module, and an output layer;

[0119] A model training module configured to select a training data from the model training data set, input the data into the current lightning data compression model, obtain the compressed data output by the encoding compression module and the reconstructed data output by the decoding reconstruction module, calculate the loss value between the compressed data and the reconstructed data, and update the parameters of the lightning data compression model through a stochastic gradient descent method optimization function as the current lightning data compression model, and perform model iterative training until a set training end condition is reached to obtain a trained lightning data compression model;

[0120] A data compression module configured to compress the data to be compressed collected in real time through the trained lightning data compression model.

[0121] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working process and related explanations of the above-described system can refer to the corresponding process in the foregoing method embodiments and will not be repeated here.

[0122] It should be noted that the lightning electromagnetic pulse waveform compression system based on artificial intelligence provided in the above embodiments is only illustrated by the division of the above function modules. In practical applications, the above functions can be allocated to different function modules according to needs, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be combined into one module, or can be further split into multiple sub-modules to complete all or part of the functions described above. For the names of the modules and steps involved in the embodiments of the present invention, they are only used to distinguish each module or step and are not regarded as an improper limitation of the present invention.

[0123] An electronic device according to a third embodiment of the present invention includes:

[0124] At least one processor; and

[0125] A memory communicatively connected to the at least one processor; wherein,

[0126] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned artificial intelligence-based lightning electromagnetic pulse waveform compression method.

[0127] A computer-readable storage medium according to a fourth embodiment of the present invention, the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned artificial intelligence-based lightning electromagnetic pulse waveform compression method.

[0128] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes and related descriptions of the above-mentioned storage device and processing device can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0129] Those skilled in the art should be able to realize that the modules and method steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. The programs corresponding to the software modules and method steps can be placed in a random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in the form of electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0130] Terms such as "first" and "second" are used to distinguish similar objects, rather than to describe or represent a specific order or sequence.

[0131] The term "including" or any other similar term is intended to cover non-exclusive inclusion, so that a process, method, article, or device / equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to these processes, methods, articles, or devices / equipment.

[0132] So far, the technical solution of the present invention has been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based lightning electromagnetic pulse waveform compression method, characterized in that, the method includes: Step S10, collecting a set number of lightning electromagnetic pulse waveform data through a lightning radiation field signal acquisition device as a model training data set; Step S20, constructing a lightning data compression model based on a stacked autoencoder; the lightning data compression model includes an input layer, an encoding compression module, a variable ratio adjustment module, a decoding reconstruction module, and an output layer; Among them, the encoding compression module includes a multi-layer one-dimensional convolutional neural network and a max pooling layer, and the decoding reconstruction module includes a one-dimensional transposed convolutional neural network corresponding to the multi-layer one-dimensional convolutional neural network and a max upsampling layer corresponding to the max pooling layer; the variable ratio adjustment module includes a set number of residual blocks connected in sequence, a fully connected layer, and a linear activation layer, and the residual block includes a set number of fully connected layers and a Tanh activation layer; The max upsampling layer is expressed as: Among them, U j (n×L+r) is the (n×L+r)-th index value of the j-th output feature vector of the output feature U of the maximum upsampling layer, S i (n) is the n-th index value of the i-th input feature vector of the input feature S; Step S30, selecting a training data from the model training data set, inputting the data into the current lightning data compression model, and obtaining the compressed data output by the encoding compression module and the reconstructed data output by the decoding reconstruction module; Step S40, calculating the loss value between the compressed data and the reconstructed data, and updating the parameters of the lightning data compression model through a stochastic gradient descent optimization function as the current lightning data compression model; Step S50, repeating Step S30 - Step S40 for model iterative training until the set training end condition is reached to obtain a trained lightning data compression model; Step S60, compressing the lightning electromagnetic pulse waveform data collected in real time through the trained lightning data compression model.

2. The artificial intelligence-based lightning electromagnetic pulse waveform compression method according to claim 1, characterized in that, the one-dimensional convolutional neural network is expressed as: Among them, X j is the j-th output feature vector of the output feature X of the one-dimensional convolutional neural network, and S i is the i-th input feature vector of the input feature S, and K ij is the convolutional kernel vector in the convolutional kernel K of the one-dimensional convolutional neural network that connects the i-th input feature vector and the j-th output feature vector, and b j is the j-th bias vector of the bias b of the one-dimensional convolutional neural network, f(·) represents the non-linear activation calculation, and M j is the number of mappings connecting the input feature and the output feature.

3. The artificial intelligence-based lightning electromagnetic pulse waveform compression method according to claim 2, characterized in that, the max pooling layer is expressed as: where, P j (n) is the n-th index value of the j-th output feature vector of the output feature P of the maximum pooling layer, R represents the size of the pooling window, r represents the element index of the pooling window, L represents the stride of the pooling, and max[·] represents the maximum value calculation.

4. The artificial intelligence-based lightning electromagnetic pulse waveform compression method according to claim 3, characterized in that, the one-dimensional transposed convolutional neural network is expressed as: Among them, Y is the Jth output feature vector of the output feature Y of the one-dimensional transposed convolutional neural network, and T represents the transpose calculation.

5. The artificial intelligence-based lightning electromagnetic pulse waveform compression method according to claim 1, characterized in that, the linear activation layer is expressed as: A j = S i × (K ij ) T + b j Among them, A j is the j-th output feature vector of the output feature A of the linear activation layer.

6. The artificial intelligence-based lightning electromagnetic pulse waveform compression method according to claim 5, characterized in that, the Tanh activation layer is expressed as: Among them, T j (n) is the n-th index value of the j-th output feature vector of the output feature T of the Tanh non-linear activation layer.

7. An artificial intelligence-based lightning electromagnetic pulse waveform compression system, characterized in that, the system includes the following modules: A data acquisition module configured to collect a set number of lightning electromagnetic pulse waveform data through a lightning radiation field signal acquisition device as a model training data set; and collect lightning electromagnetic pulse waveform data in real time as data to be compressed; A model construction module, configured to construct a lightning data compression model based on a stacked autoencoder; the lightning data compression model includes an input layer, an encoding and compression module, a variable ratio adjustment module, a decoding and reconstruction module, and an output layer; Among them, the encoding and compression module includes a multi-layer one-dimensional convolutional neural network and a max pooling layer, and the decoding and reconstruction module includes a one-dimensional transposed convolutional neural network corresponding one-to-one to the multi-layer one-dimensional convolutional neural network and a max upsampling layer corresponding to the max pooling layer; the variable ratio adjustment module includes a set number of residual blocks connected in sequence, a fully connected layer, and a linear activation layer, and the residual block includes a set number of fully connected layers and a Tanh activation layer; The max upsampling layer is expressed as: Among them, U j (n×L + r) is the (n×L + r)-th index value of the j-th output feature vector of the output feature U of the maximum upsampling layer, S i (n) is the n-th index value of the i-th input feature vector of the input feature S; A model training module, configured to select a training data from the model training dataset, input the data into the current lightning data compression model, obtain the compressed data output by the encoding and compression module and the reconstructed data output by the decoding and reconstruction module, calculate the loss value between the compressed data and the reconstructed data, and update the parameters of the lightning data compression model by optimizing the function with the stochastic gradient descent method as the current lightning data compression model, perform model iterative training until the set training end condition is reached, and obtain a trained lightning data compression model; A data compression module, configured to compress the data to be compressed collected in real time through the trained lightning data compression model.

Images