CN109829407B - Intelligent lightning identification method based on convolutional neural network - Google Patents
Intelligent lightning identification method based on convolutional neural network Download PDFInfo
- Publication number
- CN109829407B CN109829407B CN201910063035.2A CN201910063035A CN109829407B CN 109829407 B CN109829407 B CN 109829407B CN 201910063035 A CN201910063035 A CN 201910063035A CN 109829407 B CN109829407 B CN 109829407B
- Authority
- CN
- China
- Prior art keywords
- lightning
- neural network
- waveform data
- convolutional neural
- intelligent
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
An intelligent lightning identification method based on a convolutional neural network, wherein the identification device comprises: the lightning waveform recognition system comprises an input end, a storage end and a recognition end, wherein the input end is used for inputting lightning waveform data to be recognized, and the lightning waveform data is a one-dimensional array sequence; the first-layer convolution layer is connected with the input end and is used for preliminarily extracting the characteristic vector of the lightning waveform data; the middle processing layer is connected with the first layer convolution layer and used for further extracting a characteristic vector of lightning waveform data and reducing the dimensionality of the characteristic vector; the global maximum pooling layer is connected with the intermediate processing layer and is used for reducing the feature vector of the lightning waveform data extracted by the intermediate processing layer to one dimension; the full connection layer is connected with the global maximum pooling layer and is used for integrating the feature vectors of the global maximum pooling layer reduced to one dimension; and the output end is connected with the full connection layer and used for outputting the one-dimensional characteristic vector representing the lightning type.
Description
Technical Field
The disclosure relates to the field of weather-lightning and machine learning and pattern recognition, in particular to an intelligent lightning recognition method based on a convolutional neural network.
Background
Thunder is a great and spectacular discharge phenomenon in nature, has the characteristics of high discharge intensity, wide range, long time and the like, and has very complex discharge mechanism and great research difficulty. For convenience of research, scholars mainly classify lightning into two types, cloud flash and ground flash, according to the location where the lightning occurs. Cloud flash is a discharge phenomenon inside a cloud layer and between clouds, and strong electromagnetic radiation of the cloud flash has strong killing power on microelectronic equipment and the like; the ground flash is discharge between a cloud layer and the ground, can directly convey charges to the ground, and has important influence on a power system, a building and personal safety. Therefore, how to distinguish the discharge type of the lightning plays an extremely important role in lightning protection and research on the physical process of lightning discharge.
The most intuitive way to distinguish the lightning type is optical observation, and at the end of the 19 th century, spectroscopy and photography are the main tools for studying lightning. Since optical observation does not allow a rich understanding of the in-cloud process of lightning, researchers have begun to attempt to use other means to study lightning. Stekolnikov in 1937 first acquired lightning current waveforms using tethered balloons. Wilson observed systematic changes in the polarity of the steady electric field produced by the thunderstorm cloud and the transient electric field produced by the lightning from ground measurements for the first time. In 1937, Scase measured the electric field in the cloud of thunderstorms and inferred the charge structure in the cloud. Workman in 1942 completed electromagnetic field multi-station measurement on the ground relatively close to lightning. Because the transient electrical activity of lightning is accompanied by wider electromagnetic radiation, students can obtain the waveform characteristics of different discharge types by combining optics with UHF, VHF and VLF frequency band antennas. In 1998, LASA preliminarily realizes the classification of different discharge types by using a low-frequency electric field antenna and combining the lightning positioning position according to the characteristics of rise time, fall time and the like of the radiation waveform of lightning discharge. The accuracy rate of automatic waveform identification and classification is always the bottleneck limiting the process of automatically analyzing thunderstorms, and since misjudgment may cause errors of final physical conclusions, it is an important task to try to improve the accuracy rate.
For physical process analysis of thunderstorms, effective waveform judgment is an important step. At present, the work for lightning waveform identification at home and abroad mainly focuses on means such as artificial experience judgment, time domain waveform feature identification and the like.
BRIEF SUMMARY OF THE PRESENT DISCLOSURE
Technical problem to be solved
Based on the above problems, the present disclosure provides an intelligent lightning identification method based on a convolutional neural network, so as to alleviate the technical problems in the prior art, such as the limitation of automatic analysis of a thunderstorm process, due to low accuracy of automatic identification and classification of lightning waveforms.
(II) technical scheme
In the present disclosure, an intelligent lightning identification method based on a convolutional neural network is provided, which identifies a lightning type by using an intelligent lightning identification apparatus based on a convolutional neural network, and the intelligent lightning identification method includes: step A: collecting original lightning waveform data, wherein the lightning waveform data is a one-dimensional array sequence; and B: preprocessing the raw lightning waveform data collected in the step A; and C: b, training and learning the intelligent lightning recognition device based on the convolutional neural network by using the lightning waveform data preprocessed in the step B; step D: and D, inputting lightning waveform data to be identified into the lightning identification device trained and learned in the step C to obtain a one-dimensional characteristic vector representing the lightning type, and completing lightning identification.
In the embodiment of the present disclosure, the lightning waveform data in step a includes a time domain signal of lightning, and the time domain signal is an electromagnetic radiation signal of low frequency and very low frequency bands generated in the lightning discharge process.
In the embodiment of the present disclosure, the step a classifies the collected raw lightning waveform data through manual experience, so as to perform tagging.
In an embodiment of the present disclosure, the preprocessing in step B includes: data length normalization and/or data amplitude normalization.
In the embodiment of the present disclosure, the data length normalization is to unify the data length to 1.6 ms.
In the embodiment of the present disclosure, the data amplitude is normalized to be a mean value of the data amplitude, and the calculation process is as in formula (3):
wherein S is max And S min Respectively representing the maximum and minimum values, S, of lightning waveform data mean Representing sequences of lightning waveform dataAverage value, n denotes the position of the lightning waveform data sequence, S n And S r Respectively representing the waveform data sequence after amplitude normalization and the original lightning waveform data sequence.
In an embodiment of the present disclosure, the step C includes:
step C1: setting the number of iterations of the lightning recognition device training;
step C2: b, reading the lightning waveform data and the lightning type data preprocessed in the step B into the lightning recognition device for training and learning; and
step C3: and storing the trained lightning recognition device, and finishing the training and learning of the intelligent lightning recognition device based on the convolutional neural network.
In this embodiment of the present disclosure, in the step C2, the lightning time domain signal is directly processed by using a one-dimensional convolution of a multilayer nonlinear activation function, specifically, a one-dimensional array sequence x is dot-product-multiplied by a one-dimensional vector k to obtain a new array sequence s, and the calculation is as shown in formula (2):
s j ＝k T x i-m+1:j (2)
in the formula, T represents the transposition of a matrix, m represents the length of a convolution kernel k, j represents the position of feature mapping, and finally the lightning type represented by the original one-dimensional array sequence is comprehensively judged through the features of a plurality of newly generated array sequences.
In an embodiment of the present disclosure, the lightning types include, according to their discharge types: negative ground flash, positive ground flash, NBE, initial breakdown, or normal cloud flash.
In the embodiment of the present disclosure, the training and learning of the intelligent lightning recognition device based on the convolutional neural network in step C is implemented on a Keras neural network framework, and the rear end is TensorFlow.
(III) advantageous effects
According to the technical scheme, the intelligent lightning identification method based on the convolutional neural network has at least one or part of the following beneficial effects:
(1) the deep convolutional neural network is utilized to process the low-frequency/very-low-frequency time domain waveform of the lightning detected, identified and positioned by the VLF/LF lightning detection network fixedly installed on the ground, so that the automatic identification of the lightning type in the large-scale wide-area thunderstorm is realized;
(2) the deep convolutional neural network is used for analyzing and discussing the low-frequency time domain waveform of lightning, so that the identification accuracy rate of various discharge types can reach more than 95%;
(3) the method further enriches the analysis method of the lightning signal, greatly reduces the workload of manual identification, and provides a good foundation for more conveniently researching the micro physical process of lightning.
Drawings
Fig. 1 is a schematic diagram of the architecture of an intelligent lightning recognition device based on a convolutional neural network.
Fig. 2 is a schematic flow chart of an intelligent lightning identification method based on a convolutional neural network according to an embodiment of the present disclosure.
Fig. 3 is a schematic diagram of a low-frequency electric field waveform of a lightning discharge type.
FIG. 4 is a comparison sample before and after normalization of lightning data according to an embodiment of the disclosure.
Fig. 5 is a schematic diagram of a one-dimensional convolution mapping process in a convolutional neural network according to an embodiment of the present disclosure.
Fig. 6 is a diagram illustrating accuracy in a training process according to an embodiment of the disclosure.
Detailed Description
The utility model provides an intelligence lightning identification method based on convolution neural network, it is based on the lightning information that the nature takes place, utilize the deep convolution neural network to the VLF/LF lightning detection network fixed mounting on ground to detect, discern, the lightning's that the location arrived low frequency/very low frequency time domain waveform carries out analysis processes, can reach more than 95% identification accuracy to all kinds of discharge types, has realized the automatic identification of lightning type in the large-scale wide area thunderstorm. The method further enriches the analysis and processing method of the lightning signals, greatly reduces the workload of manual identification, and provides a good basis for more conveniently researching the micro-physical process of lightning.
The waveform recognition method used in this disclosure is based on Convolutional Neural Networks (CNNs), whose development has its origin in the receptive field, whose concept was proposed by Hubel and Wiesel in 1962 through studies on the cat's cerebral visual cortex system (1962, Hubel). In 1980, Fukushima K was inspired by the concept of receptive fields to propose a neurocognitive machine (necognitiron) which receives two-dimensional pattern signals like the retina and then arranges the processing layers in a hierarchical fashion in the same way as the human visual cortex processes continuously, each layer consisting of planes, each plane consisting of cells. In 1989, Yann LeCun first proposed a device for recognizing handwritten numbers using a gradient back propagation algorithm and named LeNet, which is called the first convolutional neural network device. Subsequently, as the performance of computer hardware is continuously improved and data of various industries is continuously accumulated, the convolutional neural network starts to be widely used, and a new chapter of machine learning is opened. In recent years, convolutional neural networks have become one of the research hotspots in many scientific fields, especially in the fields of pattern classification, face recognition and speech recognition.
The convolutional neural network is a multi-layer feedforward artificial neural network, and each layer in the network is composed of a plurality of independent neurons. The convolutional neural network is mainly composed of convolutional layers and pooling layers. The convolution layer is mainly used for feature extraction, and extracts the local features of signals by utilizing the local highly-correlated characteristics of one-dimensional and two-dimensional signals through proper convolution kernel filtering. The form of the convolutional layer is shown as formula (1):
wherein f (-) represents the meaning of the nonlinear activation, x represents the feature mapping vector, i, j represent the position serial number of the feature mapping, l represents the number of layers, k is the convolution kernel, M j Representing a selection of input feature ranges. Compared with artificial feature extraction, the neural network can extract features with stronger distinguishability through continuous learning of a large number of parameters, and then the local features extend to the overall features, so that the optimal feature extraction is realized. ByAnother important component of convolutional neural networks is the pooling layer, since the large number of features extracted at the convolutional layer leads to a large computational overhead and may lead to overfitting situations. The pooling layer performs down-sampling processing through convolution characteristics, so that the purpose of reducing dimensionality and calculation amount is achieved, and the signal is expressed from a high level to a low level. And the pooling layer has good linear invariance, and if the signal is translated and rotated, the output value obtained after pooling also has significance. Pooling layers is a key design for convolutional neural networks to perform well, since it reduces dimensionality while retaining significant features.
In view of the excellent characteristics of the one-dimensional convolution network in sound recognition devices (such as SoundNet) and electrocardiosignal recognition, the lightning time-domain signal is directly processed by adopting the one-dimensional convolution of the multilayer nonlinear activation function. The practical idea of one-dimensional convolution is to perform dot product on the sequence x by using a one-dimensional vector k (convolution kernel) to obtain a new sequence s, as shown in formula (2):
s j ＝k T x j-m+1:j (2)
in the formula, T represents the transposition of the matrix, m represents the length of a convolution kernel k, and j represents the position of feature mapping;
and finally, comprehensively judging the type of the original sequence through the characteristics of a plurality of new sequences. The convolutional neural network is also well suited for lightning signals, mainly for two reasons: firstly, the convolutional neural network can automatically extract richer characteristic quantities from the time domain flash signal, so that more accurate identification is facilitated; and secondly, the convolutional neural network can be used for learning in parallel, so that the learning efficiency can be greatly improved. The convolutional neural network model structure of the lightning recognition device we used contains only convolutional layers and pooling layers, which contains the global maximum pooling. In order to improve the performance of the device, the training depth of the convolutional neural network is increased, and the used framework does not have an obvious overfitting phenomenon because a large number of lightning time domain original waveforms are used for training. Too many layers result in overfitting, as too few layers result in lower device accuracy. Therefore, we have tested the neural networks with different numbers of layers, and found that the neural network has the best performance when the convolutional layer is 7 layers.
In an embodiment of the present disclosure, there is provided an intelligent lightning identification device based on a convolutional neural network, which is used for identifying a lightning type, and fig. 1 is a schematic architecture diagram of the intelligent lightning identification device based on a convolutional neural network, as shown in fig. 1, the intelligent lightning identification device based on a convolutional neural network includes:
the lightning waveform recognition system comprises an input end, a data acquisition end and a data processing end, wherein the input end is used for inputting lightning waveform data to be recognized, and the lightning waveform data is a one-dimensional array sequence;
the first-layer convolution layer is connected with the input end and is used for preliminarily extracting a characteristic vector of lightning waveform data;
the intermediate processing layer is connected with the first-layer convolution layer and used for further extracting the feature vector of the lightning waveform data and reducing the dimension of the feature vector;
the global maximum pooling layer is connected with the intermediate processing layer and is used for reducing the feature vector of the lightning waveform data extracted by the intermediate processing layer to one dimension;
the full connection layer is connected with the global maximum pooling layer and is used for integrating the feature vectors of the global maximum pooling layer reduced to one dimension;
and the output end is connected with the full connection layer and used for outputting the one-dimensional characteristic vector representing the lightning type.
The dimension of the first layer of the convolutional layer is 8000;
the step length of the first layer of the convolution layer is 1-3;
the number of convolution kernels in the first convolution layer is 8-64; the size of the convolution kernel is 2-8.
The intermediate treatment layer comprises m (m is more than or equal to 4 and less than or equal to 9) intermediate convolution layers and n (n is more than or equal to 3 and less than or equal to 7) intermediate pooling layers;
the number of layers of the middle pooling layer is less than that of the middle convolution layer.
As an embodiment, as shown in fig. 1 and table1, the middle pooling layers of the middle treatment layers may be disposed adjacent to each other.
TABLE1 convolutional neural network parameter configuration reference
| Network layer | Conv1 | Pool1 | Conv2 | Pool2 | Conv3 | Conv4 | Pool4 | Conv5 | Pool5 | Conv6 | Conv7 |
| Dimension (d) of | 8000 | 3996 | 3993 | 1996 | 1993 | 1990 | 995 | 992 | 496 | 492 | 492 |
| Number of convolution kernels | 16 | 16 | 32 | 32 | 64 | 128 | 128 | 256 | 256 | 512 | 1024 |
| Convolution kernel size | 8 | 2 | 4 | 2 | 4 | 4 | 2 | 4 | 2 | 4 | 2 |
| Step size | 1 | 2 | 2 | 2 | 1 | 1 | 2 | 1 | 2 | 1 | 1 |
To enhance the performance of the network, a Dropout (a hidden layer for preventing overfitting) technique is introduced into the lightning recognition device to control the modification of the node weights. Dropout randomly sets the output of some hidden neurons to 0 in the network training stage to stop the operation of the weights of some hidden layer nodes, and some nodes at this time can be considered not to belong to the network and do not participate in any propagation in the network. When the sample is input into the network for weight updating each time, the hidden nodes randomly appear at a certain probability, so that the phenomenon that a plurality of hidden nodes appear at the same time each time is avoided. Therefore, the updating of the weight value does not depend on the simultaneous action of the fixed hidden nodes, so that the condition that some characteristics completely depend on other characteristics is effectively avoided, the generalization capability of the network is enhanced, and the generation of overfitting is avoided. The specific parameter information for each layer of the device is shown in table 1. The device is implemented on a Keras (Keras is a high-level neural network API supporting rapid modeling of prototype neural networks) neural network framework, and the rear end is TensorFlow (a symbolic mathematical system based on data flow programming).
First, we classify raw lightning waveform data according to human experience, thereby performing tagging. Because the amplitude difference of lightning signals with different intensities and different distances received by a site is large, the signals need to be normalized before being input. The time of a lightning discharge event is mostly within 2ms, and a segment with a fixed length of 1.6ms (the sampling rate is 5M, and the total sampling rate is 8000 points) is selected as the input of the lightning identification device. And training the deep convolutional neural network device to obtain a classifier, and storing the classifier to the local. The lightning waveform of unknown type is used as the input of the classifier, and the output result of the network is a 1 × 5 one-dimensional vector. Different vectors represent different discharge types, mainly according to the labeled values at the input, for example, NCG (negative ground flash) is [1, 0, 0, 0, 0], CC (cloud flash) is [0, 1, 0, 0, 0], etc.
In an embodiment of the present disclosure, there is provided an intelligent lightning identification method based on a convolutional neural network, as shown in fig. 2, the intelligent lightning identification method based on a convolutional neural network, which identifies lightning by using the above intelligent lightning identification device based on a convolutional neural network, including:
step A: collecting original lightning waveform data, wherein the lightning waveform data is a one-dimensional array sequence;
and B: preprocessing the raw lightning waveform data collected in the step A;
and C: b, training and learning the intelligent lightning recognition device based on the convolutional neural network by using the lightning waveform data preprocessed in the step B;
step D: and D, inputting lightning waveform data to be identified into the lightning identification device trained and learned in the step C to obtain a one-dimensional characteristic vector representing the lightning type, and completing lightning identification.
Step A, the lightning waveform data comprise a lightning time domain signal, and the time domain signal is an electromagnetic radiation signal of a low frequency band and a very low frequency band generated in the lightning discharge process;
the lightning types include, depending on their discharge type: negative ground flash, positive ground flash, NBE, initial breakdown, or normal cloud flash.
And B, acquiring original lightning waveform data in the step A from a Jianghuai sky electric array base station, wherein the time domain signals are electromagnetic radiation signals of low frequency and very low frequency bands generated in the lightning discharge process, the receiving frequency band of the fast electric field antenna is 800 HZ-400 kHZ, and the sampling rate of the system is 5M/s. The lightning discharge types are mainly classified into 5 types including: negative ground flash, positive ground flash, NBE (bipolar narrow pulse), initial breakdown, normal cloud flash. In order to show the diversity of the same lightning type time domain waveform as much as possible and ensure the reliability of the device training set, a sample diagram of various lightning discharge types is provided, as shown in fig. 3. In order to ensure the reliability of experimental training data, the sequences of the ground lightning, the NBE and the initial breakdown are selected as lightning events within 500KM of the positioning station network.
In the step B, the original lightning waveform data is preprocessed, and the step mainly plays a role in removing the influence of too large difference in amplitude of the original lightning signals on the identification effect and ensuring uniform sample length, as shown in fig. 4, because most of events of a lightning discharge event are within 2ms, the original lightning waveform is intercepted into fixed segments, the length of each fixed segment is 1.6ms, and the number of sampling points is 8000. We then classify the original lightning waveform segments according to human experience and thus label them. Because the amplitude difference of lightning signals with different intensities and different distances received by a site is large, in order to reduce the influence of the lightning signals on the identification effect, before the signals are used as input, mean value normalization processing needs to be carried out on the lightning signals, and the amplitude adjustment operation is represented by formula (3). Wherein S max And S min Respectively representing the maximum and minimum values, S, of each lightning signal segment mean Denotes the mean value of the sequence, n denotes the position of the sequence, S n And S r Respectively representing the normalized sequence and the original lightning signal sequence.
The step C comprises the following steps:
step C1: setting the number of iterations of the lightning recognition device training;
step C2: b, reading the lightning waveform data and the lightning type data preprocessed in the step B into the intelligent lightning recognition device based on the convolutional neural network for training and learning; and
step C3: and storing the trained intelligent lightning recognition device based on the convolutional neural network, and finishing the training and learning of the intelligent lightning recognition device based on the convolutional neural network.
The convolutional neural network can extract feature vectors with stronger distinguishability through continuous learning of a large number of parameters compared with artificial feature extraction, and then the local features extend to the overall features to realize the extraction of the optimal lightning feature vectors. In step C, the method directly processes the lightning time domain signal using a one-dimensional convolution of the multi-layered nonlinear activation function (ReLU). The practical idea of one-dimensional convolution is to perform dot product on a one-dimensional array sequence x by using a one-dimensional vector k (i.e. convolution kernel) to obtain a new array sequence s, and the specific mapping process is shown in fig. 5, and the calculation formula (2) thereof is shown as follows:
s j ＝k T x i-m+1:j (2)
in the formula, T represents the transpose of the matrix, m represents the length of the convolution kernel k, j represents the position of feature mapping, and finally the lightning type represented by the original one-dimensional array sequence is comprehensively judged through the features (i.e. feature vectors) of a plurality of newly generated array sequences.
For more preservation of texture information, the pooling selected by the lightning recognition means is maximum pooling, i.e. taking the maximum value for feature points in the neighborhood. The original data is processed by a convolution layer and a pooling layer to obtain multi-dimensional feature vectors, and the multi-dimensional features are subjected to one-dimensional processing through global maximum pooling due to too many parameters. Then, each feature is subjected to linear weighted summation through a full connection layer of the device, and the classification category of the lightning is obtained by adopting an activation function of softmax (). The output result of the lightning recognition device is a 1 × 5 one-dimensional vector, and different vectors represent different discharge types, mainly based on the tagged value during input, for example, [1, 0, 0, 0, 0] is ground flash, [0, 1, 0, 0, 0] is normal cloud flash, etc.
In order to enhance the performance of the convolutional neural network, a Dropout technology is introduced into the device to control the modification of the node weight. Dropout randomly sets the output of some hidden neurons to 0 in the network training stage to stop the operation of the weights of some hidden layer nodes, and some nodes at this time can be considered not to belong to the network and do not participate in any propagation in the network. When the sample is input into the network for weight updating each time, the hidden nodes can randomly appear with a certain probability, so that the phenomenon that a plurality of hidden nodes appear at the same time every time is avoided. Therefore, the updating of the weight value does not depend on the simultaneous action of the fixed hidden nodes, so that the condition that some characteristics completely depend on other characteristics is effectively avoided, the generalization capability of the network is enhanced, and the generation of overfitting is avoided.
In order to verify the rationality of the construction of the deep convolutional neural network-based device proposed herein, and the validity and adaptability of the device to automatically classify the lightning discharge types, labeled data are randomly divided into a training set and a test set and then used as system inputs for evaluation. The lightning discharge types are herein mainly classified into 5 types, including: negative ground flash, positive ground flash, NBE, initial breakdown, normal cloud flash. In order to show the diversity of the time domain waveforms of the same lightning type as much as possible and ensure the reliability of the training set of the device, a sample diagram of various lightning discharge types is provided, as shown in fig. 3. In order to ensure the reliability of experimental training data and the convenience of subsequent thunderstorm analysis, the selected ground lightning, NBE and initial breakdown sequences are lightning events within 500KM of the positioning station network. Data of a plurality of thunderstorm days in 2012 and 2017 are selected as training sets, wherein 15000 negative ground flashes, 5000 positive ground flashes, 10000 initial breakdown, 5000 NBE and 16000 common cloud flashes are counted.
Due to different training times and different depth networks, the final classification effect of the lightning recognition device is greatly different, the accuracy of the device with different training times is respectively given out under the condition that the convolutional layers are 7 layers of convolutional neural networks and 5 layers of convolutional neural networks, as shown in fig. 6, the 7-layer depth network has higher accuracy in data representation compared with the 5-layer depth network. Experiments show that when the network is too deep, overfitting of training data can be caused. The parameters selected herein are primarily those shown in Table1, which is a device used to analyze the physical process of a thunderstorm. The accuracy of the lightning identification device model on the identification of the training set is 96.7%, and in order to further verify the reliability of the classification used herein, the accuracy of the device identification is 97.2% by randomly selecting 500 ground lightning waveforms of different thunderstorm days at different distances (within 500 Km) as the verification set. Because the ground flash events screened by the lightning recognition device are partially not the thunderstorms needing to be researched, the ground flashes not belonging to the thunderstorms are removed through screening and matching of positioning results, then the ground flash waveforms obtained finally are subjected to manual secondary inspection, errors caused by some long-distance back shock are removed, and the accuracy of back shock recognition is obtained through statistics and is more than 98%.
The training kernel test of the convolutional neural network model is realized on a Keras neural network framework, and the rear end is TensorFlow. (Keras is a high-level neural network API that supports rapid modeling of prototypical neural networks). The specific experimental configuration is as follows: the display card is NVIDIA GeForce GTX 1060; the CPU is Intel Core i7-7700, 3.60Ghz, the memory is 16G, and the operating system is Windows 2010.
As the lightning waveform has no internationally public data set, in order to fully illustrate the effectiveness of the method, lightning data of a plurality of thunderstorm days in 2012 and 2017 of the laboratory station network are selected as training sets, wherein the lightning data comprise 15000 negative ground flashes, 5000 positive ground flashes, 10000 initial breakdowns, 5000 NBEs and 16000 common cloud flashes. And manually pick 2000 lightning cases of 2018 as test data. The average accuracy of the lightning recognition device for 5 types of recognition was 97.1%, and the same dataset was used herein to compare the accuracy of the neural network model to the traditional machine learning classifier. The traditional machine learning needs to extract feature quantities of a sample firstly, 6 feature quantities are selected according to parameter definitions in (Liu Heng Yi 2009 and MaDong2016) in the text, and the feature quantities are respectively as follows: rise time Tr, fall time Tf, pulse width Tw, overshoot ratio, main frequency (frequency corresponding to the maximum power spectral energy value), and signal-to-noise ratio (SNR). The selected classifiers are SVM and Logitics classifier, the accuracy of the SVM to the verification set is 91.4%, and the accuracy of the Logitics classifier is 87%. Since the logistic classifier is a linear classification, the accuracy may be low. Although the SVM also shows a high recognition rate for the data set, the research of the inventor on specific cases shows that the SVM classifier has a high missing recognition phenomenon compared with an intelligent lightning recognition device of a convolutional neural network.
So far, the embodiments of the present disclosure have been described in detail with reference to the accompanying drawings. It is to be understood that the implementations not shown or described in the drawings or in the text of this specification are in a form known to those skilled in the art and are not described in detail. Further, the above definitions of the various elements and methods are not limited to the various specific structures, shapes or arrangements of parts mentioned in the examples, which may be easily modified or substituted by those of ordinary skill in the art.
From the above description, those skilled in the art should clearly recognize that the intelligent lightning recognition apparatus and recognition method based on convolutional neural network of the present disclosure.
In summary, the present disclosure provides an intelligent lightning identification method based on a convolutional neural network, which analyzes and processes a low frequency/very low frequency time domain waveform of lightning detected, identified and located by a VLF/LF lightning detection network fixedly installed on the ground by using a deep convolutional neural network based on lightning information occurring in the nature, and can achieve an identification accuracy of over 95% for various types of discharge, thereby realizing automatic identification of lightning types in a large-scale wide-area thunderstorm. The method further enriches the analysis and processing method of the lightning signal, greatly reduces the workload of manual identification, and provides a good foundation for more conveniently researching the micro-physical process of lightning.
It should also be noted that the directional terms mentioned in the embodiments, such as "upper", "lower", "front", "back", "left", "right", etc., are only directions referring to the drawings, and are not intended to limit the protection scope of the present disclosure. Throughout the drawings, like elements are represented by like or similar reference numerals. Conventional structures or constructions will be omitted when they may obscure the understanding of the present disclosure.
And the shapes and sizes of the respective components in the drawings do not reflect actual sizes and proportions, but merely illustrate the contents of the embodiments of the present disclosure. Furthermore, in the claims, any reference signs placed between parentheses shall not be construed as limiting the claim.
Unless otherwise indicated, the numerical parameters set forth in the specification and attached claims are approximations that can vary depending upon the desired properties sought to be obtained by the present disclosure. In particular, all numbers expressing quantities of ingredients, reaction conditions, and so forth used in the specification and claims are to be understood as being modified in all instances by the term "about". Generally, the expression is meant to encompass variations of ± 10% in some embodiments, 5% in some embodiments, 1% in some embodiments, 0.5% in some embodiments by the specified amount.
Furthermore, the word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements.
The use of ordinal numbers such as "first," "second," "third," etc., in the specification and claims to modify a corresponding element does not by itself connote any ordinal number of the element or any ordering of one element from another or the order of manufacture, and the use of the ordinal numbers is only used to distinguish one element having a certain name from another element having a same name.
Further, unless steps are specifically described or must occur in sequence, the order of the steps is not limited to that listed above and may be changed or rearranged as desired by the desired design. The embodiments described above may be mixed and matched with each other or with other embodiments based on design and reliability considerations, i.e., technical features in different embodiments may be freely combined to form further embodiments.
Those skilled in the art will appreciate that the modules in the device in an embodiment may be adaptively changed and disposed in one or more devices different from the embodiment. The modules or units or components of the embodiments may be combined into one module or unit or component, and furthermore they may be divided into a plurality of sub-modules or sub-units or sub-components. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and all of the processes or elements of any method or apparatus so disclosed, may be combined in any combination, except combinations where at least some of such features and/or processes or elements are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise. Furthermore, in the unit claims enumerating several means, several of these means can be embodied by one and the same item of hardware.
Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the disclosure, various features of the disclosure are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the disclosure and aiding in the understanding of one or more of the various disclosed aspects. However, the disclosed method should not be construed to reflect the intent: rather, the present disclosure is directed to more features than are expressly recited in each claim. Rather, as the following claims reflect, disclosed aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of this disclosure.
The above-mentioned embodiments are intended to illustrate the objects, aspects and advantages of the present disclosure in further detail, and it should be understood that the above-mentioned embodiments are only illustrative of the present disclosure and are not intended to limit the present disclosure, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present disclosure should be included in the scope of the present disclosure.
Claims (6)
1. An intelligent lightning identification method based on a convolutional neural network is used for identifying the type of lightning by using an intelligent lightning identification device based on the convolutional neural network, and comprises the following steps:
step A: collecting original lightning waveform data, wherein the lightning waveform data is a one-dimensional array sequence;
and B: preprocessing the original lightning waveform data collected in the step A;
and C: b, training and learning the intelligent lightning recognition device based on the convolutional neural network by using the lightning waveform data preprocessed in the step B; and
step D: inputting lightning waveform data to be identified into the lightning identification device trained and learned in the step C to obtain a one-dimensional characteristic vector representing the lightning type, and completing lightning identification;
the lightning waveform data comprise a lightning time domain signal, and the time domain signal is an electromagnetic radiation signal of low frequency and very low frequency bands generated in the lightning discharge process;
the step C comprises the following steps:
step C1: setting the number of iterations of the lightning recognition device training;
step C2: b, reading the lightning waveform data and the lightning type data preprocessed in the step B into the lightning recognition device for training and learning; and
step C3: storing the trained lightning recognition device, and finishing the training and learning of the intelligent lightning recognition device based on the convolutional neural network;
in the step C2, the lightning time domain signal is directly processed by using one-dimensional convolution of a multilayer nonlinear activation function, specifically, a one-dimensional vector is used to perform dot product on a one-dimensional array sequence x to obtain a new array sequence s, and the calculation is as shown in formula (2):
s j ＝k T x i-m+1：j (2)；
in the formula, T represents the transposition of a matrix, m represents the length of a convolution kernel k, i and j represent the position serial numbers of feature mapping, and finally the features of a plurality of newly generated array sequences are used as the basis for comprehensively judging the lightning types represented by the original one-dimensional array sequences;
the lightning types include, depending on their discharge type: negative ground flash, positive ground flash, NBE, initial breakdown, or normal cloud flash.
2. An intelligent lightning recognition method based on convolutional neural network according to claim 1, wherein said step a classifies the collected raw lightning waveform data through artificial experience for tagging.
3. The intelligent convolutional neural network-based lightning recognition method of claim 1, wherein the preprocessing of step B comprises: data length normalization and/or data amplitude normalization.
4. An intelligent lightning recognition method based on a convolutional neural network according to claim 3, said data length normalization is to unify the data length to 1.6 ms.
5. The intelligent lightning recognition method based on the convolutional neural network of claim 3, wherein the data amplitude is normalized by performing mean normalization on the data amplitude, and the calculation process is as shown in formula (3):
wherein S is max And S min Respectively representing the maximum and minimum values, S, of lightning waveform data mean Representing the mean value of the lightning waveform data sequence, n representing the position of the lightning waveform data sequence, S n And S r Respectively representing the waveform data sequence after amplitude normalization and the original lightning waveform data sequence.
6. The intelligent lightning recognition method based on the convolutional neural network of claim 1, wherein the training and learning of the intelligent lightning recognition device based on the convolutional neural network of step C are implemented on a Keras neural network framework, and the backend is TensorFlow.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910063035.2A CN109829407B (en) | 2019-01-23 | 2019-01-23 | Intelligent lightning identification method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910063035.2A CN109829407B (en) | 2019-01-23 | 2019-01-23 | Intelligent lightning identification method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109829407A CN109829407A (en) | 2019-05-31 |
| CN109829407B true CN109829407B (en) | 2022-09-09 |
Family
ID=66861898
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910063035.2A Active CN109829407B (en) | 2019-01-23 | 2019-01-23 | Intelligent lightning identification method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109829407B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113358974A (en) * | 2021-06-08 | 2021-09-07 | 合肥工业大学 | Lightning overvoltage fault type identification method for power transmission line |
| CN113960375A (en) * | 2021-10-18 | 2022-01-21 | 特斯联科技集团有限公司 | Artificial intelligent lightning detection system and method for early warning of forest and grassland fire |
| CN117009784A (en) * | 2023-07-20 | 2023-11-07 | 南通大学 | Cloud-to-ground flash characteristic signal identification method for distributed cut space mapping |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102135560B (en) * | 2011-02-23 | 2013-02-27 | 山东大学 | Disturbance identification method used during intrusion of lightning waves in transformer substation |
| CN102186008B (en) * | 2011-03-18 | 2013-09-11 | 中国气象科学研究院 | All-in-view lightning event observation system and method |
| CN103679263B (en) * | 2012-08-30 | 2017-03-29 | 重庆邮电大学 | Forecasting Methodology is closed on based on the thunder and lightning of particle swarm support vector machine |
| CN105118511A (en) * | 2015-07-31 | 2015-12-02 | 国网电力科学研究院武汉南瑞有限责任公司 | Thunder identification method |
| CN107728234B (en) * | 2017-09-17 | 2020-06-16 | 北京工业大学 | Thunder and lightning strength value prediction method based on atmospheric electric field data |
| CN108052734A (en) * | 2017-12-12 | 2018-05-18 | 中国电力科学研究院有限公司 | A kind of method and system predicted based on meteorologic parameter amplitude of lightning current |
-
2019
- 2019-01-23 CN CN201910063035.2A patent/CN109829407B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN109829407A (en) | 2019-05-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109829407B (en) | Intelligent lightning identification method based on convolutional neural network | |
| CN103955702B (en) | SAR image terrain classification method based on depth RBF network | |
| CN102819745B (en) | Hyper-spectral remote sensing image classifying method based on AdaBoost | |
| CN107451614B (en) | Hyperspectral classification method based on fusion of space coordinates and space spectrum features | |
| CN107016405A (en) | A kind of insect image classification method based on classification prediction convolutional neural networks | |
| CN104732244B (en) | The Classifying Method in Remote Sensing Image integrated based on wavelet transformation, how tactful PSO and SVM | |
| CN105956612B (en) | Hyperspectral image classification method based on Active Learning and semi-supervised learning | |
| CN104794368A (en) | Rolling bearing fault classifying method based on FOA-MKSVM (fruit fly optimization algorithm-multiple kernel support vector machine) | |
| CN108830312B (en) | Integrated learning method based on sample adaptive expansion | |
| CN111913156A (en) | Radar radiation source individual identification method based on deep learning model and feature combination | |
| CN110197205A (en) | A kind of image-recognizing method of multiple features source residual error network | |
| CN106096660B (en) | Convolutional neural networks based on independent composition analysis algorithm | |
| CN112904299B (en) | Radar high-resolution range profile open set target identification method based on deep class segmentation | |
| CN109829408B (en) | Intelligent lightning recognition device based on convolutional neural network | |
| CN105512681A (en) | Method and system for acquiring target category picture | |
| CN107909109A (en) | SAR image sorting technique based on conspicuousness and multiple dimensioned depth network model | |
| CN107194423A (en) | The hyperspectral image classification method of the integrated learning machine that transfinites of feature based random sampling | |
| CN112861813B (en) | Method for identifying human behavior behind wall based on complex value convolution neural network | |
| CN111695407A (en) | Gender identification method, system, storage medium and terminal based on multispectral fusion | |
| CN116229380A (en) | Method for identifying bird species related to bird-related faults of transformer substation | |
| Puspitasari et al. | Partial discharge waveform identification using image with convolutional neural network | |
| CN104408472A (en) | Wishart and SVM (support vector machine)-based polarimetric SAR (synthetic aperture radar) image classification method | |
| CN101408945B (en) | Method for sorting radar two-dimension image based on multi-dimension geometric analysis | |
| CN111025100A (en) | Transformer ultrahigh frequency partial discharge signal mode identification method and device | |
| CN110414494A (en) | SAR image classification method with ASPP deconvolution network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |