CN112700007A - Training method, forecasting method and device of ionosphere electron content forecasting model - Google Patents
Training method, forecasting method and device of ionosphere electron content forecasting model Download PDFInfo
- Publication number
- CN112700007A CN112700007A CN202011629360.XA CN202011629360A CN112700007A CN 112700007 A CN112700007 A CN 112700007A CN 202011629360 A CN202011629360 A CN 202011629360A CN 112700007 A CN112700007 A CN 112700007A
- Authority
- CN
- China
- Prior art keywords
- module
- neural network
- neuron
- electron content
- training
- 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.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S19/00—Satellite radio beacon positioning systems; Determining position, velocity or attitude using signals transmitted by such systems
- G01S19/38—Determining a navigation solution using signals transmitted by a satellite radio beacon positioning system
- G01S19/39—Determining a navigation solution using signals transmitted by a satellite radio beacon positioning system the satellite radio beacon positioning system transmitting time-stamped messages, e.g. GPS [Global Positioning System], GLONASS [Global Orbiting Navigation Satellite System] or GALILEO
- G01S19/42—Determining position
-
- 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
- Engineering & Computer Science (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computer Networks & Wireless Communication (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The disclosed embodiment provides a training method, a forecasting method and a device of an ionosphere electron content forecasting model, which can adopt ionosphere electron content data with relevance in space-time dimension, obtain a forecasting model of the ionosphere electron content based on preset attention residual convolution cyclic neural network training, obtain variation characteristics in time and space dimensions and reduce the influence of the deep neural network degradation problem based on the preset attention residual convolution cyclic neural network in the training process, and obtain local interest characteristics by utilizing a multi-head attention module. High-precision forecast data can be obtained.
Description
Technical Field
The disclosure belongs to the technical field of ionosphere detection, and particularly relates to a training method, a forecasting method and a device of an ionosphere electron content forecasting model.
Background
The ionosphere affects the propagation of electromagnetic wave signals, so in the Navigation positioning technology, positioning errors caused by the ionosphere influence are one of the important error sources of Global Navigation Satellite System (GNSS). In order to reduce the influence of the ionosphere on navigation positioning, the GNSS continuous observation station carries out real-time continuous detection on the electron content of the ionosphere and forecasts the change of the electron content.
At present, electron content observation data are obtained based on GNSS technology monitoring, and a global or regional forecasting method for the electron content of an ionized layer mainly utilizes a mathematical experience model to forecast. The forecasting methods which are applied more are processing methods based on time series, the methods carry out modeling forecasting on the electron content time series of a certain geographic position, but the relevance of an ionized layer on the space-time dimension is lacked, so that the forecasting result is not ideal. In the prior art, deep learning is applied to the electron content prediction, although the relevance of an ionized layer in a space-time dimension is considered, the deep learning network has certain defects, and as the network depth is deepened, the gradient change of the network is in a white noise state, so that the deep learning network has a degradation problem, and the accuracy of the electron content prediction is influenced finally.
Disclosure of Invention
The embodiment of the disclosure provides a training method, a forecasting method and a device for an ionosphere electron content forecasting model, which can be applied to ionosphere electron content forecasting and are beneficial to obtaining high-precision forecasting data.
In a first aspect, an embodiment of the present disclosure provides a method for training an ionospheric electron content prediction model, where the method includes:
acquiring a training sample set; the training sample set comprises a plurality of training samples, and each training sample is t + p ionosphere electron content data sequences acquired according to a preset time interval;
training a preset attention residual error convolution cyclic neural network by taking ionosphere electron content data of the first t sequences in a training sample as an input sample; the attention residual error convolution cyclic neural network comprises t neuron modules which are sequentially connected, wherein each neuron module comprises a residual error convolution cyclic neural network module and a multi-head attention module; the residual convolution cyclic neural network module of the ith neuron module is used for correspondingly acquiring the ith ionosphere electron content data in the input sample and the output value of the multi-head attention module of the (i-1) th neuron module, extracting sequence characteristics and obtaining a characteristic value and a first characteristic matrix; the multi-head attention module of the ith neuron module is used for obtaining a second feature matrix according to the first feature matrix of the ith neuron module, and the second feature matrix is used as an output value of the multi-head attention module of the ith neuron module; i is more than or equal to 1 and less than or equal to t;
acquiring characteristic values of a residual convolution cyclic neural network module as forecast data through the last p neuron modules in the t neuron modules;
and taking the last p ionized layer electron content data in the training samples as output samples, and training a preset attention residual error convolution cyclic neural network according to a loss function between the forecast data and the output samples to obtain an ionized layer electron content forecast model.
In some embodiments, the residual convolutional recurrent neural network module comprises n layers of sequentially connected residual network modules; the residual error network module comprises a convolution neural network, a convolution cyclic neural network and an inverse convolution neural network which are sequentially connected, wherein n is a positive integer;
inputting a first output value of a j-1 layer residual error network module in an ith neuron module into a convolutional neural network of the j layer residual error network module in the ith neuron module, and inputting a j value in a second feature matrix output by a multi-head attention module in the i-1 layer neuron module into a convolutional cyclic neural network of the j layer residual error network module in the ith neuron module;
and adding the output value of the inverse convolutional neural network of the jth layer residual error network module in the ith neuron module and the first output value of the jth-1 layer residual error network module in the ith neuron module to obtain a first output value of the jth layer residual error network module in the ith neuron module.
In some embodiments of the present invention, the,
taking the hidden state of the convolution cycle neural network output of the jth layer residual error network module of the ith neuron module as a second output value of the jth layer residual error network module of the ith neuron module;
the first feature matrix comprises n values, and the second output value of the jth layer residual error network module of the ith neuron module is used as the jth value in the first feature matrix.
In some embodiments, the residual convolution cyclic neural network module of the ith neuron module is configured to correspondingly obtain the ith ionospheric electron content data in the input sample and the output value of the multi-head attention module of the (i-1) th neuron module, and specifically includes:
a residual convolution cyclic neural network module of the 1 st neuron module, which is used for correspondingly acquiring the 1 st ionized layer electron content data in the input sample and initializing a zero matrix
And will initialize the zero matrix
Convolution cyclic neural network corresponding to each residual error network module inputted to the 1 st neuron module;
and correspondingly inputting the ith ionized layer electron content data in the input sample into a convolutional neural network of a layer 1 residual network module of a residual convolutional cyclic neural network module of the ith neuron module.
In some embodiments, obtaining, by the last p neuron modules of the t neuron modules, a feature value of the residual convolutional recurrent neural network module as prediction data specifically includes:
and taking the first output value of the nth layer residual error network module in the last p neuron modules in the t neuron modules as a characteristic value and taking the first output value as forecast data.
In some embodiments, the ionospheric electron content data is grid data formed at predetermined latitude and longitude intervals.
In a second aspect, an embodiment of the present disclosure provides a training apparatus for an ionospheric electron content prediction model, the apparatus including:
the first acquisition module is used for acquiring a training sample set; the training sample set comprises a plurality of training samples, and each training sample is t + p ionosphere electron content data sequences acquired according to a preset time interval;
the training module is used for training a preset attention residual convolution cyclic neural network by taking ionosphere electron content data of the first t sequences in a training sample as an input sample; the attention residual error convolution cyclic neural network comprises t neuron modules which are sequentially connected, wherein each neuron module comprises a residual error convolution cyclic neural network module and a multi-head attention module; the residual convolution cyclic neural network module of the ith neuron module is used for correspondingly acquiring the ith ionosphere electron content data in the input sample and the output value of the multi-head attention module of the (i-1) th neuron module, extracting sequence characteristics and obtaining a characteristic value and a first characteristic matrix; the multi-head attention module of the ith neuron module is used for obtaining a second feature matrix according to the first feature matrix of the ith neuron module, and the second feature matrix is used as an output value of the multi-head attention module of the ith neuron module; i is more than or equal to 1 and less than or equal to t;
the second acquisition module is used for acquiring the characteristic value of the residual convolution cyclic neural network module as forecast data through the last p neuron modules in the t neuron modules;
and the adjusting module is used for training a preset attention residual error convolution cyclic neural network by taking the last p ionized layer electron content data in the training samples as output samples according to a loss function between the forecast data and the output samples so as to obtain an ionized layer electron content forecast model.
In a third aspect, an embodiment of the present disclosure provides a training device for an ionospheric electron content prediction model, where the device includes: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the method for training the ionospheric electron content prediction model according to any one of the above embodiments.
In a fourth aspect, an embodiment of the present disclosure provides a method for predicting an ionospheric electron content, which is implemented by using a prediction model obtained by using a training method of an ionospheric electron content prediction model according to any one of the embodiments; the forecasting method comprises the following steps:
obtaining a forecast sample; forecasting samples into t ionized layer electron content data sequences collected according to preset time intervals;
taking ionospheric electron content data of t sequences of forecast samples as input samples, and inputting the input samples into a forecast model;
and acquiring characteristic values of the corresponding residual convolution cyclic neural network module as prediction data of the electron content of the ionized layer through p neuron modules in the last t neuron modules of the prediction model.
In a fifth aspect, an embodiment of the present disclosure provides an ionospheric electron content predictor, including:
the third acquisition module is used for acquiring a forecast sample; forecasting samples into t ionized layer electron content data sequences collected according to preset time intervals;
the input module is used for inputting the forecasting model by taking ionosphere electron content data of t sequences of the forecasting samples as input samples, and the forecasting model comprises t neuron modules;
and the forecasting module is used for acquiring the characteristic value of the corresponding residual convolution cyclic neural network module as the forecasting data of the ionosphere electron content through the last p neuron modules in the t neuron modules of the forecasting model.
In some embodiments, the forecasting model is obtained by training through the training method of the ionospheric electron content forecasting model in any one of the above embodiments.
The ionosphere electron content forecasting model training method, the ionosphere electron content forecasting method and the ionosphere electron content forecasting device can adopt ionosphere electron content data with relevance in space-time dimension, obtain the ionosphere electron content forecasting model based on preset attention residual convolution cyclic neural network training, obtain the change characteristics in time and space dimension and reduce the influence of the deep neural network degradation problem based on the preset attention residual convolution cyclic neural network in the training process, and obtain the local interest characteristics by utilizing a multi-head attention module, so that the ionosphere electron content forecasting model can improve the gradient degradation problem of a deep learning network while giving consideration to the whole and local transformation of ionosphere data, is favorable for obtaining a more accurate forecasting model and is applied to ionosphere electron content forecasting, high-precision forecast data can be obtained.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the embodiments of the present disclosure will be briefly described below, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
Fig. 1 is a schematic flowchart of a method for training an ionospheric electron content prediction model according to an embodiment of the present disclosure;
FIG. 2 is a schematic flow chart of the training method shown in FIG. 1 for obtaining training samples, wherein FIG. 2a is a schematic flow chart of step S102 shown in FIG. 1; FIG. 2b is a schematic diagram of acquired ionospheric electron content data in a training sample;
fig. 3 is a schematic structural diagram of an attention residual convolution cyclic neural network based on a training method of an ionosphere electron content prediction model according to an embodiment of the present disclosure;
FIG. 4 is a schematic diagram of a structure of an attention residual convolution cyclic neural network in one embodiment; wherein FIG. 4a is a schematic diagram of the structure of a neuron module; FIG. 4b is a schematic diagram of a residual network module;
FIG. 5 is a schematic diagram of a training process of a sample in an attention residual convolution cyclic neural network;
fig. 6 is a schematic structural diagram of a training device for an ionospheric electron content prediction model according to an embodiment of the present disclosure;
FIG. 7 is a flowchart illustrating a method for training an ionospheric electron content prediction model according to an embodiment of the present disclosure;
fig. 8 is a schematic flowchart of a training apparatus for an ionospheric electron content prediction model according to an embodiment of the present disclosure
Fig. 9 is a schematic structural diagram of a training device of an ionospheric electron content prediction model according to an embodiment of the present disclosure.
Detailed Description
Features and exemplary embodiments of various aspects of the present disclosure will be described in detail below, and in order to make objects, technical solutions and advantages of the present disclosure more apparent, the present disclosure will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are intended to be illustrative only and are not intended to be limiting of the disclosure. It will be apparent to one skilled in the art that the present disclosure may be practiced without some of these specific details. The following description of the embodiments is merely intended to provide a better understanding of the present disclosure by illustrating examples of the present disclosure.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.
The ionosphere affects the propagation of electromagnetic wave signals, which causes carrier phase advance modulated by GNSS (Global Navigation Satellite System) signals and pseudo range code group delay, and finally affects the accuracy of Navigation positioning, especially in the active ionosphere region. At present, a global or regional forecasting method for the electron content of the ionized layer is mainly to forecast by using a mathematical experience model, wherein more traditional mathematical methods based on time series are applied, and the change characteristics of the ionized layer in the time dimension are processed through a traditional machine learning or deep learning network without considering the relevance of the electron content in the space; in addition, the self defects of the traditional machine learning or deep learning network cause the deep learning network to have the degradation problem along with the deepening of the network depth, so that the network performance is poor, and the accuracy of an output value is reduced; in the traditional technology, an encoder or a decoder is constructed, and a scheme that an output value is similar to an input value as much as possible in the mean square error sense is adopted through an encoding-decoding mode of a neural network, but because an input sample can only be encoded into integral information, the data effect far away from a decoding layer is reduced, meanwhile, the problem that a local interest region cannot be obtained exists, and the training effect is poor.
In order to solve the problems in the prior art, the embodiments of the present disclosure provide a training method, a forecasting method, and a device for an ionospheric electron content forecasting model.
First, a method for training an ionospheric electron content prediction model provided in an embodiment of the present disclosure is described below.
Fig. 1 shows a flowchart of a training method of an ionospheric electron content prediction model according to an embodiment of the present disclosure. As shown in fig. 1, the method may include the steps of:
s101, acquiring a training sample set; the training sample set comprises a plurality of training samples, and each training sample is t + p ionosphere electron content data sequences acquired according to a preset time interval;
s102, taking ionosphere electron content data of the first t sequences in a training sample as an input sample, and training a preset attention residual convolution cyclic neural network; the attention residual error convolution cyclic neural network comprises t neuron modules which are sequentially connected, wherein each neuron module comprises a residual error convolution cyclic neural network module and a multi-head attention module; the residual convolution cyclic neural network module of the ith neuron module is used for correspondingly acquiring the ith ionosphere electron content data in the input sample and the output value of the multi-head attention module of the (i-1) th neuron module, extracting sequence characteristics and obtaining a characteristic value and a first characteristic matrix; the multi-head attention module of the ith neuron module is used for obtaining a second feature matrix according to the first feature matrix of the ith neuron module, and the second feature matrix is used as an output value of the multi-head attention module of the ith neuron module; i is more than or equal to 1 and less than or equal to t;
s103, acquiring characteristic values of a residual convolution cyclic neural network module as forecast data through the last p neuron modules in the t neuron modules;
s104, taking the last p ionized layer electron content data in the training samples as output samples, and training a preset attention residual error convolution cyclic neural network according to a loss function between the forecast data and the output samples to obtain an ionized layer electron content forecast model.
In this embodiment, training is performed based on a preset attention residual convolution cyclic neural network to obtain a prediction model of the ionosphere electron content. The single neuron module in the attention residual convolution cyclic neural network adopts the structures of the residual convolution cyclic neural network and the multi-head attention module, so that the problem that training is difficult to perform due to degradation of the traditional deep neural network can be avoided while extracting a plurality of relevance characteristics of an input sample; on the basis of feature extraction, local interest features are obtained by using a multi-head attention module, so that feature representation with higher resolution is obtained, training can be completed more effectively, and the performance of a model is improved.
Illustratively, in the training sample set obtained in step S101, the training samples include a plurality of ionospheric electron content data sequences collected at preset time intervals, and grid data is formed at preset longitude and latitude intervals, so that the training samples have time characteristics and spatial characteristics, and a subsequent attention residual convolution cyclic neural network model is trained. Specifically, in this example, as shown in fig. 2a in fig. 2, in the process of acquiring the training sample set, step S101 may include:
s201, collecting ionized layer electron content data;
acquiring global or regional ionospheric electron content data acquired according to a certain time interval, and forming grid data by the ionospheric electron content data of each time point according to a certain longitude and latitude interval so as to form a training sample set. Because the acquired ionospheric electron content data has time and space longitude and latitude characteristics, the forecasting model obtained by subsequent model training can perform relevance extraction and analysis on the time and space change characteristics of the ionospheric electron content data, and then improves the accuracy of the forecasting model in forecasting the ionospheric electron content.
By utilizing a global navigation positioning system GNSS continuous observation station in a global range or a regional range, the ionosphere can be continuously monitored in a large range in real time, and massive ionosphere monitoring data in a space-time dimension is obtained and is used as a training sample in the embodiment; for example, a Global Ionospheric grid Map (GIM) issued by an IGS (International GNSS Service) every day is used, the GIM Map is Global Ionospheric electron content data, the spatial longitude and latitude resolution is 2.5 ° × 5 °, the coverage longitude range is-180 ° to 180 °, the coverage latitude range is-87.5 ° to 87.5 °, and the time resolution is 2 h; 12 GIM plots were obtained 24 hours a day.
Step S101 in this example further includes S202. collecting ionosphere electron content data, and presetting a training sample;
in this step, training samples formed by the collected ionospheric electron content data include t + p ionospheric electron content data sequences with a certain time interval, and a training sample set formed by a plurality of training samples is used for subsequent model training. As shown in fig. 2b, if the obtained grid data (e.g., the GIM map) of the ionospheric electron content at each time point is regarded as one image, the entire training sample set may be regarded as time-series data of the image at a certain time interval. Illustratively, a training sample in the set of training samples is[x1 x2 … xi … xt … xt+p]Wherein, with x1For example, x1The electronic content map at the 1 st moment of the sample is a two-dimensional array (namely grid data) with certain spatial resolution, which can be expressed as an array of n multiplied by m, wherein n and m respectively represent the number of longitude and latitude pixel points divided according to the spatial resolution; x is the number ofiAnd the electronic content graph represents the ith ionized layer electronic content in the sample, the first t graph sequences of the sample are input sample data, the last p graph sequences of the sample are output sample data, and t is more than or equal to p.
Based on a preset training sample set, the preset attention residual convolution cyclic neural network can be trained through steps S102 to S104 to obtain a prediction model of ionosphere electron content. Illustratively, as shown in fig. 3, the Attention residual convolution cyclic neural network preset in the present disclosure includes a plurality of neuron modules (Cell _ Block)310 connected in sequence, and each of the neuron modules 310 includes a residual convolution cyclic neural network module (Res _ CRNN)311 and a Multi-Head Attention module (Multi-Head Attention) 312. In this example, the attention residual convolution cyclic neural network includes t neuron modules 310; the residual convolution cyclic neural network module 311 of the ith neuron module 310 may be configured to correspondingly obtain the ith ionospheric electron content data in the input sample and the output value of the multi-head attention module 312 of the (i-1) th neuron module 310, extract the sequence feature, and obtain a feature value and a first feature matrix; the multi-head attention module 312 of the ith neuron module 310 may be configured to obtain a second feature matrix according to the first feature matrix of the ith neuron module 310, and use the second feature matrix as an output value of the multi-head attention module 312 of the ith neuron module 310; wherein i is more than or equal to 1 and less than or equal to t; i and t are positive integers.
In a specific example, as shown in fig. 4a of fig. 4, a residual convolutional cyclic neural network module (Res _ CRNN)411 in the neuron module (Cell _ Block)410 includes N layers of sequentially connected residual network modules (Res _ Block)4110, where N is a positive integer; and each layer of residual network module 4110 comprises a convolutional neural network (Conv)4111, a Convolutional Recurrent Neural Network (CRNN)4112, and an inverse convolutional neural network (Convd)4113 connected in sequence.
For example, as shown in fig. 4b, the connection between the residual error network modules 4110 in each layer of the residual error convolutional neural network module 411 includes the normalization and activation function operation processes, wherein the convolutional neural networks 4111 in the residual error network modules 4110 may be arranged in multiple layers and connected in sequence; the network types selectable by the convolution cyclic neural network 4112 include a convolution long-term memory network LSTM, a convolution gate control cyclic unit neural network GRU, and the like, but are not limited thereto; the inverse convolutional neural network 4113 may also be arranged in multiple layers and connected in sequence.
In this example, based on a preset attention residual convolution cyclic neural network, when data is transmitted between neuron modules (Cell _ Block), a residual convolution cyclic neural network module (Res _ CRNN) of an ith neuron module (Cell _ Block) correspondingly obtains the ith ionospheric electron content data in the input sample and the output value of a multi-head attention module of the (i-1) th neuron module, and transmits the data in the forward direction; for example, initially, the residual convolution cyclic neural network module (Res _ CRNN) of the 1 st neuron module (Cell _ Block) correspondingly acquires the 1 st ionospheric electron content data in the input sample and initializes the zero matrix
Wherein, the 1 st ionospheric electron content data is input into the convolutional neural network of the 1 st layer residual error convolutional recurrent neural network module in the 1 st layer residual error convolutional recurrent neural network module of the 1 st neuron module (Cell _ Block), and a zero matrix is initialized
And correspondingly inputting the data output by each residual convolution cyclic neural network module (Res _ CRNN) in the single neuron module into an internal multi-head attention module to obtain a final output value, and entering the next neuron module (Cell _ Block). The electron content of the ith ionized layer in the input sampleAnd inputting the data to the convolution neural network of the residual error network module at the layer 1 of the residual error convolution cyclic neural network module of the ith neuron module correspondingly.
When a traditional deep Neural Network (such as a Convolutional Neural Network, CNN) performs weight and bias parameter updating based on back propagation, the gradient change of the Network appears in a white noise state as the Network depth deepens due to self defects, the Network gradually degrades, and the accuracy on a training set reaches saturation or even declines. The neuron module in the preset attention residual convolution cyclic neural network utilizes the performance of the residual network module, can directly transmit the input value to the output, protects the integrity of information, and reduces the degradation influence of the traditional neural network; a feature matrix formed by output values of the multilayer residual error network module can dynamically concern a local interest region of the ionosphere electron content data through the multi-head attention module, the influence of irrelevant feature information in the ionosphere electron content data on model forecasting training is reduced, and more effective model training is facilitated.
Specifically, based on the attention residual convolution cyclic neural network, a training sample set can be utilized to perform multiple rounds of training; before multiple rounds of training, setting a loss function and initialized weight and bias in a prediction model based on the attention residual convolution cyclic neural network, and then training the model one by one through each training sample; and obtaining a forecasting model after multiple rounds of training. Then we use any one of the training samples x below1 x2 … xi … xt … xt+p]The training process in a certain round is taken as an example for explanation; s102, taking ionospheric electron content data of the first t sequences in the training sample as an input sample x1 … xi … xt]The output sample is [ x ]t+1 xt+2 … xt+p]Referring to fig. 3 and 4, when training, the following training process may be specifically included:
first of all, initializing
Is a zero matrix; the 1 st neuron module Cell _ Block1The input data comprises grid data x of ionospheric electron content at the 1 st moment1And zero matrix
The 2 nd neuron module Cell _ Block2The input data comprises grid data x of ionospheric electron content at the 2 nd moment2And the 1 st neuron module Cell _ Block1Output value of
And so on subsequently, the ith neuron module Cell _ BlockiThe input data comprise grid data x of ionospheric electron content at the ith momentiAnd the i-1 th Cell _ Blocki-1The input lug
Using the ith neuron module Cell _ BlockiFor example, as shown in FIG. 5, the mesh data xiAnd the i-1 th neuron module Cell _ Blocki-1Output of (2)
Are input into the ith neuron module Cell _ Block togetheriFirstly, the signal passes through a neuron module Cell _ BlockiResidual convolution cyclic neural network module Res _ CRNN ofiAfter that, the residual convolution cyclic neural network module Res _ CRNNiIs a hidden state matrix HiAnd a characteristic value
Hidden state matrix HiNamely a first feature matrix; next, the hidden state matrix H is divided intoiInputting the ith neuron module Cell _ BlockiAfter the multi-head attention module, the output value of the multi-head attention module is a second feature matrix of
Finally, when i is less than or equal to t-p, the neuron module Cell _ BlockiIs output as
When i > t-q, the neuron module Cell _ BlockiIs output as
And predicted value at the i + q th time
The forecast value is also the forecast data.
If the i-th neuron module Cell _ BlockiResidual convolution cyclic neural network module Res _ CRNN ofiIf n layers of residual network modules Res _ Block are included, the ith neuron module Cell _ Block is usediMiddle j-1 layer residual network module Res _ Blocki，j-1First output value x ofi，j-1(j＝0，1，…，n，xi，0＝xi) Is inputted to the i-th neuron module Cell _ BlockiMiddle j layer residual error network module Res _ Blocki，jAnd the i-1 th neuron module Cell _ Blocki-1Second feature matrix of multi-head attention module output
The j-th value is input to the i-th neuron module Cell _ BlockiMiddle j layer residual error network module Res _ Blocki，jThe convolutional recurrent neural network CRNN of (1).
If the residual network module Res _ Block is ini，jThe convolutional layers Conv (i.e., the convolutional neural network, the same applies hereinafter), 1 convolutional recurrent neural network CRNN, and a plurality of deconvolution layers Convd (i.e., the deconvolution neural network, the same applies hereinafter). Residual network module Res _ Blocki，jIs xi，j-1Firstly, convolution operation is carried out through the convolution layer Conv, image characteristic information is extracted, and after extraction, the output characteristic value is xi，j，conv(ii) a Then the characteristic value xi，j，convAnd
inputting the jth value into a convolution cyclic neural network (CRNN) to obtain a hidden state hi，jAnd xi，j，CRNN(ii) a Then x is puti，j，CRNNRestoring the image pixel to x by inputting to the deconvolution layer Convdi，j，convd(ii) a Finally, the output data x of the deconvolution layer is processedi，j，convdAfter the activation function f is f (x)i，j，convd) Then the current residual network module Res _ Blocki，jIs xi，j＝xi，j-1+f(xi，j，convd) And hidden state hi，j，xi，jCell _ Block as a neuron moduleiMiddle j layer residual error network module Res _ Blocki，jFirst output value of (d), hidden state hi，jCell _ Block as a neuron moduleiMiddle j layer residual error network module Res _ Blocki，jThe second output value of (1).
Then the hidden states output by all the layer residual error network modules form a residual error convolution cyclic neural network module Res _ CRNNiThe output value of (i.e. the first feature matrix) is the hidden state matrix Hi＝[hi，1 … hi，j … hi，n](ii) a In which the hidden state hi，jIs the neuron module Cell _ BlockiMiddle j layer residual error network module Res _ Blocki，jIs also the jth value in the first feature matrix. Residual convolution cyclic neural network module Res _ CRNNiLayer n residual error network module Res _ Blocki，nFirst output value x ofi，nAs residual convolution cyclic neural network module Res _ CRNNiThe characteristic value of (2).
If neuron module Cell _ BlockiResidual convolution cyclic neural network module Res _ CRNN ofiHidden state matrix Hi＝[hi，1 … hi，j … hi，n](j ═ 1, 2, …, n), hidden state matrix HiIs the neuron module Cell _ BlockiThe input and the output of the multi-head attention module
Is represented by the following formula (1):
wherein softmax denotes the softmax function, Wi，QRepresents the ith neuron module Cell _ BlockiQuery vector weight, W, for the Mead-Bull attention Modulei，KRepresents the ith neuron module Cell _ BlockiMiddle-to-multi-head attention module key vector weight, Wi，VRepresents the ith neuron module Cell _ BlockiThe vector weight of the attention value of the middle-multiple head; dk represents the dimensions of the query vector and key vector described above.
After the model is trained in step S102 by using each training sample, in step S103, the eigenvalues of the residual convolution cyclic neural network module of the model are obtained as prediction data by using the last p neuron modules of the t neuron modules; that is, in the specific example, the predicted values respectively output by the p neuron modules are taken
As forecast data for the current round of training.
Through step S104, prediction data obtained based on the attention residual convolution cyclic neural network
And output sample [ xt+1 xt+2 … xt+p]Calculating the loss value of the loss function, then transferring the accumulated gradient backwards, adjusting the weight and bias set in the forecasting model. And (3) performing multiple rounds of training on the prediction model based on the attention residual convolution cyclic neural network through the training sample set to finally obtain an ionosphere electron content prediction model which can be used for predicting the subsequent ionosphere electron content.
The training method of the ionospheric electron content prediction model of the embodiment of the disclosure adopts ionospheric electron content data having relevance in space-time dimension, obtains the ionospheric electron content prediction model based on the preset attention residual convolution cyclic neural network training, based on the preset attention residual convolution cyclic neural network in the training process, the residual convolution cyclic neural network module can acquire the variation characteristics in time and space dimensions and reduce the influence of the degradation problem of the deep neural network, the multi-head attention module is utilized to acquire the local interest characteristics, therefore, the method and the device can improve the gradient degradation problem existing in the deep learning network while giving consideration to the whole and local transformation of the ionosphere data, are favorable for obtaining a more accurate prediction model, are applied to the prediction of the electron content of the ionosphere, and can obtain high-precision prediction data.
Fig. 6 shows a schematic structural diagram of a training device of an ionospheric electron content prediction model according to an embodiment of the present disclosure. As shown in fig. 6, the training apparatus includes:
a first obtaining module 601, configured to obtain a training sample set; the training sample set comprises a plurality of training samples, and each training sample is t + p ionosphere electron content data sequences acquired according to a preset time interval;
a training module 602, configured to train a preset attention residual convolution cycle neural network with ionospheric electron content data of the first t sequences in the training samples as input samples; the attention residual convolution cyclic neural network comprises t neuron modules which are sequentially connected, wherein each neuron module comprises a residual convolution cyclic neural network module and a multi-head attention module; the residual convolution cyclic neural network module of the ith neuron module is used for correspondingly acquiring the ith ionospheric electron content data in the input sample and the output value of the multi-head attention module of the (i-1) th neuron module, extracting sequence characteristics and obtaining a characteristic value and a first characteristic matrix; the multi-head attention module of the ith neuron module is used for obtaining a second feature matrix according to the first feature matrix of the ith neuron module, and the second feature matrix is used as an output value of the multi-head attention module of the ith neuron module; i is more than or equal to 1 and less than or equal to t; i and t are positive integers;
a second obtaining module 603, configured to obtain, through last p neuron modules of the t neuron modules, a feature value of the residual convolutional recurrent neural network module as prediction data;
the adjusting module 604 is configured to train a preset attention residual convolution cyclic neural network according to a loss function between the prediction data and the output sample by using the last p ionospheric electron content data in the training samples as the output sample, so as to obtain an ionospheric electron content prediction model.
In this embodiment, the preset attention residual convolution cyclic neural network has a network structure shown in fig. 3 and fig. 4, and based on the attention residual convolution cyclic neural network, for example, the first obtaining module 601 may perform the step S101 shown in fig. 1, the training module 602 may perform the step S102 shown in fig. 1, the second obtaining module 603 may perform the step S103 shown in fig. 1, and the adjusting module 604 may perform the step S104 shown in fig. 1.
It should be noted that all relevant contents of each step related to the above method embodiment may be referred to the functional description of the corresponding functional module, and the corresponding technical effect can be achieved, and for brevity, no further description is provided herein.
Illustratively, as shown in fig. 3, the Attention residual convolution cyclic neural network preset in the present embodiment includes a plurality of neuron modules (Cell _ Block)310 connected in sequence, and each neuron module 310 includes a residual convolution cyclic neural network module (Res _ CRNN)311 and a Multi-Head Attention module (Multi-Head Attention) 312. In this example, the attention residual convolution cyclic neural network includes t neuron modules 310; the residual convolution cyclic neural network module 311 of the ith neuron module 310 may be configured to correspondingly obtain the ith ionospheric electron content data in the input sample and the output value of the multi-head attention module 312 of the (i-1) th neuron module 310, extract the sequence feature, and obtain a feature value and a first feature matrix; the multi-head attention module 312 of the ith neuron module 310 may be configured to obtain a second feature matrix according to the first feature matrix of the ith neuron module 310, and use the second feature matrix as an output value of the multi-head attention module 312 of the ith neuron module 310; wherein i is more than or equal to 1 and less than or equal to t; i and t are positive integers.
In a specific example, as shown in fig. 4a of fig. 4, a residual convolutional cyclic neural network module (Res _ CRNN)411 in the neuron module (Cell _ Block)410 includes n layers of sequentially connected residual network modules (Res _ Block)4110, where n is a positive integer; and each layer of residual network module 4110 comprises a convolutional neural network (Conv)4111, a Convolutional Recurrent Neural Network (CRNN)4112, and an inverse convolutional neural network (Convd)4113 connected in sequence.
For example, as shown in fig. 4b, the connection between the residual error network modules 4110 in each layer of the residual error convolutional neural network module 411 includes the normalization and activation function operation processes, wherein the convolutional neural networks 4111 in the residual error network modules 4110 may be arranged in multiple layers and connected in sequence; the network types selectable by the convolution cyclic neural network 4112 include a convolution long-term memory network LSTM, a convolution gate control cyclic unit neural network GRU, and the like, but are not limited thereto; the inverse convolutional neural network 4113 may also be arranged in multiple layers and connected in sequence.
In this example, based on a preset attention residual convolution cyclic neural network, when data is transmitted between neuron modules (Cell _ Block), a residual convolution cyclic neural network module (Res _ CRNN) of an ith neuron module (Cell _ Block) correspondingly obtains the ith ionospheric electron content data in the input sample and the output value of a multi-head attention module of the (i-1) th neuron module, and transmits the data in the forward direction; for example, initially, the residual convolution cyclic neural network module (Res _ CRNN) of the 1 st neuron module (Cell _ Block) correspondingly acquires the 1 st ionospheric electron content data in the input sample and initializes the zero matrix
Wherein, the 1 st ionospheric electron content data is input into the convolutional neural network of the 1 st layer residual error convolutional recurrent neural network module in the 1 st layer residual error convolutional recurrent neural network module of the 1 st neuron module (Cell _ Block), and a zero matrix is initialized
And correspondingly inputting the data output by each residual convolution cyclic neural network module (Res _ CRNN) in the single neuron module into an internal multi-head attention module to obtain a final output value, and entering the next neuron module (Cell _ Block). And inputting the other ith ionospheric electron content data in the input sample into the convolutional neural network of the residual network module of the layer 1 of the residual convolutional recurrent neural network module of the ith neuron module correspondingly.
Fig. 7 is a schematic flow chart of an ionospheric electron content prediction method according to an embodiment of the present disclosure, and is implemented by a prediction model obtained by the ionospheric electron content prediction model training method according to any of the embodiments; as shown in fig. 7, the forecasting method includes:
s701, acquiring a forecast sample; forecasting samples into t ionized layer electron content data sequences collected according to preset time intervals;
s702, taking ionosphere electron content data of t sequences of forecast samples as input samples, and inputting the input samples into a forecast model;
and S703, acquiring characteristic values of the corresponding residual convolution cyclic neural network module as prediction data of the electron content of the ionized layer through the last p neuron modules in the t neuron modules of the prediction model.
In an exemplary case, in the prediction samples, t ionospheric electron content data sequences acquired according to a preset time interval form grid data at a preset longitude and latitude interval, the grid data is used as an input sample and input into the prediction model, then corresponding processing is performed through t neuron modules in the prediction model, and finally, a characteristic value output by a residual error convolution cyclic neural network module in the last p neuron modules is used as prediction data of ionospheric electron content at a corresponding prediction time, so that the method can be applied to positioning calculation of a global navigation positioning system, and based on the obtained prediction value, the influence of an ionospheric layer on a high-precision positioning result is weakened.
Fig. 8 is a schematic structural diagram of an ionospheric electron content prediction apparatus according to an embodiment of the present disclosure, and is implemented by a prediction model obtained by the method for training an ionospheric electron content prediction model according to any of the embodiments; as shown, the forecasting apparatus comprises:
a third obtaining module 801, configured to obtain a forecast sample; forecasting samples into t ionized layer electron content data sequences collected according to preset time intervals;
an input module 802, configured to input the prediction model by using ionospheric electron content data of t sequences of the prediction samples as input samples; the forecasting model comprises t neuron modules;
and the forecasting module 803 is configured to obtain, through the last p neuron modules in the t neuron modules of the forecasting model, a feature value of the corresponding residual convolutional recurrent neural network module as forecasting data of the ionospheric electron content.
In this embodiment, the preset attention residual convolution cyclic neural network has the network structure shown in fig. 3 and 4, and based on the attention residual convolution cyclic neural network, for example, the third obtaining module 801 may perform the step S701 shown in fig. 7, the input module 802 may perform the step S702 shown in fig. 1, and the forecasting module 803 may perform the step S703 shown in fig. 7.
It should be noted that all relevant contents of each step related to the above method embodiment may be referred to the functional description of the corresponding functional module, and the corresponding technical effect can be achieved, and for brevity, no further description is provided herein.
Fig. 9 shows a hardware structure diagram of a training device for an ionospheric electron content prediction model provided by an embodiment of the present disclosure.
The training apparatus for the ionospheric electron content prediction model may comprise a processor 901 and a memory 902 in which computer program instructions are stored.
Specifically, the processor 901 may include a Central Processing Unit (CPU), an Application Specific Integrated Circuit (ASIC), or one or more Integrated circuits configured to implement the embodiments of the present disclosure.
The processor 901 reads and executes the computer program instructions stored in the memory 902 to implement the methods/steps S101 to S104 in the embodiment shown in fig. 1, and achieve the corresponding technical effects achieved by the embodiment shown in fig. 1 executing the methods/steps thereof, which are not described herein again for brevity.
In one example, the training device for the ionospheric electron content prediction model may also include a communication interface 903 and a bus 910. As shown in fig. 9, the processor 901, the memory 902, and the communication interface 903 are connected via a bus 910 to complete communication with each other.
The communication interface 903 is mainly used for implementing communication between modules, apparatuses, units and/or devices in the embodiments of the present disclosure.
The bus 910 includes hardware, software, or both to couple the components of the training equipment of the ionospheric electron content prediction model to each other. By way of example, and not limitation, a Bus may include an Accelerated Graphics Port (AGP) or other Graphics Bus, an Enhanced Industry Standard Architecture (EISA) Bus, a Front-Side Bus (Front Side Bus, FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) Bus, an infiniband interconnect, a Low Pin Count (LPC) Bus, a memory Bus, a Micro Channel Architecture (MCA) Bus, a Peripheral Component Interconnect (PCI) Bus, a PCI-Express (PCI-X) Bus, a Serial Advanced Technology Attachment (SATA) Bus, a video electronics standards association local (VLB) Bus, or other suitable Bus or a combination of two or more of these. Bus 910 can include one or more buses, where appropriate. Although this disclosed embodiment describes and illustrates a particular bus, this disclosure contemplates any suitable bus or interconnect.
In addition, in combination with the training method of the ionospheric electron content prediction model in the foregoing embodiment, the embodiments of the present disclosure may provide a computer storage medium to implement. The computer storage medium having computer program instructions stored thereon; the computer program instructions, when executed by a processor, implement a method for training an ionospheric electron content prediction model according to any of the above embodiments.
It is to be understood that this disclosure is not limited to the particular configurations and processes described above and shown in the drawings. A detailed description of known methods is omitted herein for the sake of brevity. In the above embodiments, several specific steps are described and shown as examples. However, the method processes of the present disclosure are not limited to the specific steps described and illustrated, and those skilled in the art may make various changes, modifications, and additions or change the order between the steps after comprehending the spirit of the present disclosure.
The functional blocks shown in the above-described structural block diagrams may be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, it may be, for example, an electronic Circuit, an Application Specific Integrated Circuit (ASIC), suitable firmware, plug-in, function card, or the like. When implemented in software, the elements of the present disclosure are the programs or code segments used to perform the required tasks. The program or code segments may be stored in a machine-readable medium or transmitted by a data signal carried in a carrier wave over a transmission medium or a communication link. A "machine-readable medium" may include any medium that can store or transfer information. Examples of a machine-readable medium include electronic circuits, semiconductor memory devices, ROM, flash memory, Erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, Radio Frequency (RF) links, and so forth. The code segments may be downloaded via computer networks such as the internet, intranet, etc.
It should also be noted that the exemplary embodiments mentioned in this disclosure describe some methods or systems based on a series of steps or devices. However, the present disclosure is not limited to the order of the above-described steps, that is, the steps may be performed in the order mentioned in the embodiments, may be performed in an order different from the order in the embodiments, or may be performed several steps at the same time.
Aspects of the present disclosure are described above with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions/acts specified in the flowchart and/or block diagram block or blocks. Such a processor may be, but is not limited to, a general purpose processor, a special purpose processor, an application specific processor, or a field programmable logic circuit. It will also be understood that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware for performing the specified functions or acts, or combinations of special purpose hardware and computer instructions.
As described above, only the specific embodiments of the present disclosure are provided, and it can be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working processes of the system, the module and the unit described above may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again. It should be understood that the scope of the present disclosure is not limited thereto, and any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope of the present disclosure, and these modifications or substitutions should be covered within the scope of the present disclosure.
Claims (11)
1. A training method of an ionospheric electron content prediction model is characterized by comprising the following steps:
acquiring a training sample set; the training sample set comprises a plurality of training samples, and each training sample is t + p ionosphere electron content data sequences collected according to a preset time interval;
training a preset attention residual convolution cyclic neural network by taking ionosphere electron content data of the first t sequences in the training samples as input samples; the attention residual convolution cyclic neural network comprises t neuron modules which are sequentially connected, wherein each neuron module comprises a residual convolution cyclic neural network module and a multi-head attention module; the residual convolution cyclic neural network module of the ith neuron module is used for correspondingly acquiring the ith ionosphere electron content data in the input sample and the output value of the multi-head attention module of the (i-1) th neuron module, extracting sequence characteristics and obtaining a characteristic value and a first characteristic matrix; the multi-head attention module of the ith neuron module is used for obtaining a second feature matrix according to the first feature matrix of the ith neuron module, and taking the second feature matrix as an output value of the multi-head attention module of the ith neuron module; i is more than or equal to 1 and less than or equal to t;
acquiring characteristic values of the residual convolution cyclic neural network module as forecast data through the last p neuron modules in the t neuron modules;
and taking the last p ionized layer electron content data in the training samples as output samples, and training the preset attention residual error convolution cyclic neural network according to a loss function between the forecast data and the output samples to obtain an ionized layer electron content forecast model.
2. The method for training the ionospheric electron content prediction model according to claim 1, wherein the residual convolutional neural network module comprises n layers of sequentially connected residual network modules; the residual error network module comprises a convolution neural network, a convolution cyclic neural network and an inverse convolution neural network which are sequentially connected, and n is a positive integer;
inputting a first output value of the residual network module at the j-1 th layer in the ith neuron module into a convolutional neural network of the residual network module at the j-1 th layer in the ith neuron module, and inputting a j-th value in a second feature matrix output by a multi-head attention module in the i-1 th neuron module into a convolutional circular neural network of the residual network module at the j-th layer in the ith neuron module;
and adding the output value of the inverse convolutional neural network of the residual error network module at the jth layer in the ith neuron module and the first output value of the residual error network module at the jth-1 layer in the ith neuron module to obtain a first output value of the residual error network module at the jth layer in the ith neuron module.
3. The method of claim 2, wherein the ionospheric electron content prediction model is trained on a model of ionospheric electron content,
taking the hidden state of the convolution cycle neural network output of the residual error network module at the jth layer of the ith neuron module as a second output value of the residual error network module at the jth layer of the ith neuron module;
wherein the first feature matrix comprises n values, and a second output value of the residual error network module at a j-th layer of the i-th neuron module is used as a j-th value in the first feature matrix.
4. The ionospheric electron content neural network training method of claim 2, wherein the residual convolution cycle neural network module of the ith neuron module is configured to correspondingly obtain the ith ionospheric electron content data in the input sample and an output value of a multi-head attention module of the (i-1) th neuron module, and specifically includes:
a residual convolution cyclic neural network module of the 1 st neuron module, which is used for correspondingly acquiring the 1 st ionized layer electron content data in the input sample and initializing a zero matrix
And initializing the zero matrix
A convolution cyclic neural network corresponding to each of the residual network modules inputted to the 1 st neuron module;
and correspondingly inputting the ith ionospheric electron content data in the input sample to the convolutional neural network of the residual network module at the layer 1 of the residual convolutional cyclic neural network module of the ith neuron module.
5. The method for training the ionospheric electron content prediction model according to claim 2, wherein the obtaining, by the last p neuron modules among the t neuron modules, the eigenvalues of the residual convolutional recurrent neural network module as prediction data specifically comprises:
and taking the first output value of the residual error network module of the nth layer in the last p neuron modules in the t neuron modules as a characteristic value and forecast data.
6. The training method of the ionospheric electron content prediction model according to any one of claims 1 to 5, wherein the ionospheric electron content data is grid data formed at a predetermined longitude and latitude interval.
7. A training device for an ionospheric electron content prediction model, comprising:
the first acquisition module is used for acquiring a training sample set; the training sample set comprises a plurality of training samples, and each training sample is t + p ionosphere electron content data sequences collected according to a preset time interval;
the training module is used for training a preset attention residual convolution cyclic neural network by taking ionosphere electron content data of the first t sequences in the training samples as input samples; the attention residual convolution cyclic neural network comprises t neuron modules which are sequentially connected, wherein each neuron module comprises a residual convolution cyclic neural network module and a multi-head attention module; the residual convolution cyclic neural network module of the ith neuron module is used for correspondingly acquiring the ith ionospheric electron content data in the input sample and the output value of the multi-head attention module of the (i-1) th neuron module, extracting sequence characteristics and obtaining a characteristic value and a first characteristic matrix; the multi-head attention module of the ith neuron module is used for obtaining a second feature matrix according to the first feature matrix of the ith neuron module, and the second feature matrix is used as an output value of the multi-head attention module of the ith neuron module; i is more than or equal to 1 and less than or equal to t;
a second obtaining module, configured to obtain, by last p neuron modules in the t neuron modules, a feature value of the residual convolutional recurrent neural network module as prediction data;
and the adjusting module is used for training the preset attention residual convolution cyclic neural network by taking the last p ionized layer electron content data in the training samples as output samples according to a loss function between the forecast data and the output samples so as to obtain an ionized layer electron content forecast model.
8. Training equipment for an ionospheric electron content prediction model, comprising: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the method of training an ionospheric electron content prediction model according to any one of claims 1 to 6.
9. A method for forecasting the electron content in the ionosphere, which is implemented by a forecasting model obtained by the training method of the ionosphere electron content forecasting model according to any one of claims 1 to 6; the forecasting method comprises the following steps:
obtaining a forecast sample; the forecast samples are t ionospheric electron content data sequences collected according to a preset time interval;
taking ionospheric electron content data of the t sequences of the forecast samples as input samples, and inputting the input samples into the forecast model;
and acquiring characteristic values of the corresponding residual convolution cyclic neural network module as prediction data of the electron content of the ionized layer through the last p neuron modules in the t neuron modules of the prediction model.
10. An ionospheric electron content predictor, the predictor comprising:
the third acquisition module is used for acquiring a forecast sample; the forecast samples are t ionospheric electron content data sequences collected according to a preset time interval;
the input module is used for inputting a forecasting model by taking ionosphere electron content data of t sequences of the forecasting samples as input samples, and the forecasting model comprises t neuron modules;
and the forecasting module is used for acquiring the characteristic value of the corresponding residual convolution cyclic neural network module as the forecasting data of the electron content of the ionized layer through the last p neuron modules in the t neuron modules of the forecasting model.
11. The ionospheric electron content predictor according to claim 10, wherein the predictor model is trained by the method for training an ionospheric electron content predictor model according to any one of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011629360.XA CN112700007B (en) | 2020-12-30 | 2020-12-30 | Training method, forecasting method and device of ionosphere electronic content forecasting model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011629360.XA CN112700007B (en) | 2020-12-30 | 2020-12-30 | Training method, forecasting method and device of ionosphere electronic content forecasting model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112700007A true CN112700007A (en) | 2021-04-23 |
| CN112700007B CN112700007B (en) | 2023-06-20 |
Family
ID=75513410
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011629360.XA Active CN112700007B (en) | 2020-12-30 | 2020-12-30 | Training method, forecasting method and device of ionosphere electronic content forecasting model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112700007B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115062526A (en) * | 2022-02-22 | 2022-09-16 | 中国科学院自动化研究所 | Deep learning-based three-dimensional ionosphere electron concentration distribution model training method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111539453A (en) * | 2020-03-30 | 2020-08-14 | 东南大学 | Global ionized layer electron total content prediction method based on deep cycle neural network |
| CN111814855A (en) * | 2020-06-28 | 2020-10-23 | 东南大学 | Global ionospheric total electron content prediction method based on residual seq2seq neural network |
| CN111932017A (en) * | 2020-08-13 | 2020-11-13 | 江苏师范大学 | Short-term forecasting method suitable for single-station ionized layer TEC |
-
2020
- 2020-12-30 CN CN202011629360.XA patent/CN112700007B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111539453A (en) * | 2020-03-30 | 2020-08-14 | 东南大学 | Global ionized layer electron total content prediction method based on deep cycle neural network |
| CN111814855A (en) * | 2020-06-28 | 2020-10-23 | 东南大学 | Global ionospheric total electron content prediction method based on residual seq2seq neural network |
| CN111932017A (en) * | 2020-08-13 | 2020-11-13 | 江苏师范大学 | Short-term forecasting method suitable for single-station ionized layer TEC |
Non-Patent Citations (3)
| Title |
|---|
| ALEXANDRE BOULCH ET AL.: "Ionospheric activity prediction using convolutional recurrent neural networks", 《ARXIV: 1810:13273V2》 * |
| 罗欣: "基于LSTM的区域电离层总电子含量预测建模", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * |
| 袁天娇 等: "基于深度学习递归神经网络的电离层总电子含量经验预报模型", 《空间科学学报》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115062526A (en) * | 2022-02-22 | 2022-09-16 | 中国科学院自动化研究所 | Deep learning-based three-dimensional ionosphere electron concentration distribution model training method |
| CN115062526B (en) * | 2022-02-22 | 2024-05-28 | 中国科学院自动化研究所 | Three-dimensional ionosphere electron concentration distribution model training method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112700007B (en) | 2023-06-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111784041B (en) | Wind power prediction method and system based on graph convolution neural network | |
| WO2023134666A1 (en) | Terminal positioning method and apparatus, and device and medium | |
| CN110781266B (en) | Urban perception data processing method based on time-space causal relationship | |
| CN112669594B (en) | Method, device, equipment and storage medium for predicting traffic road conditions | |
| CN112700007A (en) | Training method, forecasting method and device of ionosphere electron content forecasting model | |
| CN116643293A (en) | GNSS positioning method and device, equipment and storage medium | |
| CN114330120A (en) | 24-hour PM prediction based on deep neural network2.5Method of concentration | |
| CN116861262B (en) | Perception model training method and device, electronic equipment and storage medium | |
| Rumapea et al. | Improving Convective Cloud Classification with Deep Learning: The CC-Unet Model. | |
| CN112598590A (en) | Optical remote sensing time series image reconstruction method and system based on deep learning | |
| CN115563888B (en) | Spacecraft residual life prediction method, system, electronic equipment and medium | |
| CN111709438A (en) | Heterogeneous sensor information fusion method | |
| CN115062526B (en) | Three-dimensional ionosphere electron concentration distribution model training method based on deep learning | |
| CN115880580A (en) | Intelligent extraction method for optical remote sensing image road information under influence of cloud layer | |
| CN116189008A (en) | Remote sensing image change detection method based on fixed point number quantification | |
| CN116306790A (en) | Offshore ship track real-time prediction method, system, equipment and medium based on CNN-GRU and attention mechanism | |
| CN111814855B (en) | Global ionospheric total electron content prediction method based on residual seq2seq neural network | |
| Li et al. | MP mitigation in GNSS positioning by GRU NNs and adaptive wavelet filtering | |
| CN111353441B (en) | Road extraction method and system based on position data fusion | |
| CN117271959B (en) | Uncertainty evaluation method and equipment for PM2.5 concentration prediction result | |
| CN111369795B (en) | Traffic flow statistical method, device, equipment and storage medium | |
| CN117491987B (en) | Ship track splicing method based on LSTM neural network and space-time motion distance algorithm | |
| CN116828397B (en) | Track information acquisition method and device, electronic equipment and storage medium | |
| CN116972837B (en) | Self-adaptive vehicle-mounted combined navigation positioning method and related equipment | |
| Li et al. | Spatial–Temporal Graph-Enabled Convolutional Neural Network–Based Approach for Traffic Networkwide Travel Time |
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 |