CN113095215B - Solar radio filtering method and system based on improved LSTM network - Google Patents
Solar radio filtering method and system based on improved LSTM network Download PDFInfo
- Publication number
- CN113095215B CN113095215B CN202110383050.2A CN202110383050A CN113095215B CN 113095215 B CN113095215 B CN 113095215B CN 202110383050 A CN202110383050 A CN 202110383050A CN 113095215 B CN113095215 B CN 113095215B
- Authority
- CN
- China
- Prior art keywords
- data
- solar
- value
- channel
- neural network
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 89
- 238000001914 filtration Methods 0.000 title claims abstract description 25
- 238000013507 mapping Methods 0.000 claims abstract description 46
- 238000003062 neural network model Methods 0.000 claims abstract description 25
- 125000004122 cyclic group Chemical group 0.000 claims abstract description 20
- 238000007781 pre-processing Methods 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims description 30
- 238000012545 processing Methods 0.000 claims description 29
- 230000006870 function Effects 0.000 claims description 23
- 230000000306 recurrent effect Effects 0.000 claims description 20
- 230000015654 memory Effects 0.000 claims description 15
- 230000008569 process Effects 0.000 claims description 15
- 238000002372 labelling Methods 0.000 claims description 8
- 238000004422 calculation algorithm Methods 0.000 claims description 7
- 238000012216 screening Methods 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 3
- 238000012935 Averaging Methods 0.000 claims description 2
- 238000013528 artificial neural network Methods 0.000 abstract description 19
- 238000001228 spectrum Methods 0.000 abstract description 5
- 238000010586 diagram Methods 0.000 description 16
- 238000004880 explosion Methods 0.000 description 9
- 210000004027 cell Anatomy 0.000 description 7
- 238000004088 simulation Methods 0.000 description 7
- 230000002452 interceptive effect Effects 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000002474 experimental method Methods 0.000 description 5
- 230000002829 reductive effect Effects 0.000 description 5
- 238000011160 research Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000012546 transfer Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 239000003086 colorant Substances 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 230000002238 attenuated effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- 230000000977 initiatory effect Effects 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 230000006386 memory function Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 230000010287 polarization Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- 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
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Signal Processing (AREA)
- Monitoring And Testing Of Transmission In General (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention discloses a solar radio filtering method and a system based on an improved LSTM network, comprising the following steps: acquiring radio station channels needing to be processed, and selecting a set amount of channel data before a solar burst event for each radio station channel; preprocessing the channel data, and extracting data by using a segmented sliding window; inputting the extracted data into a trained cyclic neural network model based on a digital mapping method, and outputting a solar radio prediction value during solar burst; the numerical value appearing in the sequence is regarded as different classes by the cyclic neural network model based on the numerical mapping method, and the output predicted value is the numerical value with the maximum prediction probability in the classes. The invention provides a cyclic neural network structure based on a digital mapping method, which classifies the data of a single-frequency channel radio station before the data enters a network, establishes the mapping of time periods to time points, greatly improves the accuracy of a prediction result, and provides possibility for solar spectrum lossless filtering.
Description
Technical Field
The invention relates to the technical field of solar radio filtering, in particular to a solar radio filtering method and system based on an improved LSTM network.
Background
The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.
Solar activity is closely related to our daily life, and solar radio is therefore an important field in astronomical physics research. Especially, for the research on the solar radio explosion process, the carried important information is helpful to explain the physical process of the related plasma change, and can also find out the law of energy change and analyze important physical phenomena such as substance movement and the like.
In order to better research the fine structure of solar radio outbreak, project groups establish a solar radio frequency spectrograph with high time resolution and high frequency resolution in a honored combed landscape area. The mountain face sea of the combed mountain is kept away from towns, has avoided complicated communication environment, greatly reduced radio frequency interference, nevertheless at meter wave band, inevitable still has some radio station signals and other electromagnetic interference. The interference signals seriously affect the observation and analysis of solar explosion events, and especially, the frequency band occupied by some interference signals is wide, and the signal intensity is greater than the solar radio flux intensity so as to directly cover the explosion events, so that certain anti-interference measures are required.
Common interference filtering measures for the solar spectrogram are divided into hardware processing and software processing. In the prior art, various interferences from the ground are weakened by utilizing the spatial selectivity of an antenna in hardware processing, the interference introduced by a cable power supply is inhibited by adopting measures such as selecting an amplifier with good interference performance, shielding, grounding and the like, and out-of-band signals are inhibited by using a filter. However, the hardware processing cycle is long, the cost is high, the efficiency is low, and a large amount of interference still exists in the acquired data, so that the image enhancement and the interference removal are often performed by using software.
In astronomy in recent years, deep learning methods such as a multi-modal network, a long-and-short-term memory network, a deep belief network and a convolutional neural network have all been well represented in classification and archiving of solar radio frequency spectrograms and identification of outbreak events. However, the problem that the prediction value is not accurate in predicting the signal value of the interference radio station in the solar radio burst section is often solved.
Disclosure of Invention
In order to solve the problems, the invention provides a solar radio filtering method and a solar radio filtering system based on an improved LSTM network. The accuracy of the prediction result is greatly improved, and the possibility is provided for solar spectrum lossless filtering.
In some embodiments, the following technical scheme is adopted:
a solar radio filtering method based on an improved LSTM network comprises the following steps:
acquiring radio station channels needing to be processed, and selecting a set amount of channel data before a solar burst event for each radio station channel;
preprocessing the channel data, and extracting data by using a segmented sliding window;
inputting the extracted data into a trained cyclic neural network model based on a digital mapping method, and outputting a solar radio prediction value during solar burst;
the numerical value appearing in the sequence is regarded as different classes by the cyclic neural network model based on the numerical mapping method, and the output predicted value is the numerical value with the maximum prediction probability in the classes.
As a further scheme, selecting a set amount of channel data before a solar outbreak event specifically comprises:
screening out a target radio station by using the mean value and the standard deviation of all signal intensity values in the time period t;
averaging the intensity values of a plurality of channels where the target radio station is located, and converting the intensity values into a one-dimensional array which changes along with time;
selecting a time point with an average value obviously higher than that of a pure electric station as the starting position of a solar burst event;
and selecting a set amount of channel data before the starting time.
As a further scheme, the channel data is preprocessed, and a segmented sliding window is used to extract data, which specifically includes:
normalizing and normalizing the channel data;
segmenting the sequence, and searching the relation between adjacent time periods; respectively determining an input window, an output window and a segmented sliding window, wherein one segmented sliding window comprises a complete input window and a complete output window, and the three windows as a whole slide on a time sequence to take values; wherein, the output window only contains one frame data and is positioned next to the tail part of the input window;
and obtaining the mapping of the input window to the output window after the segmentation sliding value taking.
As a further scheme, the recurrent neural network model based on the digital mapping method sequentially comprises an input layer, a full connection layer, an LSTM layer, a full connection layer and an output layer;
a Dropout function is added in the cyclic neural network model based on the digital mapping method to avoid the overfitting phenomenon of the network, a loss function is a softmax cross entropy function, and learning and updating of weight values select an Adam algorithm in an adaptive learning rate optimization algorithm.
As a further scheme, the training process for the recurrent neural network model based on the numerical mapping method comprises the following steps:
selecting a set amount of data before a solar outbreak event to construct a data set;
preprocessing the data in the data set, and extracting the data by using a segmented sliding window; labeling the data in the section sliding window;
and dividing the labeled data into a training set and a prediction set, and training the cyclic neural network model based on the digital mapping method.
As a further scheme, a network search method is adopted to traverse all hyper-parameter combinations in an exhaustive mode within a certain range, and a group of optimal hyper-parameters is screened out to be used for learning, training and predicting of the model.
As a further scheme, the data in the segmented sliding window is subjected to labeling processing by adopting an One-Hot coding mode.
In other embodiments, the following technical solutions are adopted:
a solar-based rf filtering system based on an improved LSTM network, comprising:
the data acquisition module is used for acquiring radio station channels needing to be processed, and for each radio station channel, channel data of a set number before a solar burst event is selected;
the data processing module is used for preprocessing the channel data and extracting data by utilizing a segmented sliding window;
the model prediction module is used for inputting the extracted data into a trained recurrent neural network model based on a digital mapping method and outputting a solar radio prediction value during solar burst; the numerical value appearing in the sequence is regarded as different classes by the cyclic neural network model based on the numerical mapping method, and the output predicted value is the numerical value with the maximum prediction probability in the classes.
In other embodiments, the following technical solutions are adopted:
a terminal device comprising a processor and a memory, the processor being arranged to implement instructions; the memory is configured to store a plurality of instructions adapted to be loaded by the processor and to perform the above-described improved LSTM network based solar radiofiltering method.
In other embodiments, the following technical solutions are adopted:
a computer readable storage medium having stored therein a plurality of instructions adapted to be loaded by a processor of a terminal device and to execute the above-mentioned improved LSTM network based solar radiofiltering method.
Compared with the prior art, the invention has the beneficial effects that:
the invention adopts the sectional sliding window to extract data, and can avoid the problem that the prediction error is increased when the radio station value collected by the actual frequency spectrograph is changed at a certain moment under the influence of some factors.
The invention adopts the recurrent neural network model based on the digital mapping method to predict, and the error can be reduced and the accuracy of prediction can be improved by the digital mapping method.
When a model is trained, a group of optimal hyper-parameters is screened out and used for learning, training and predicting the model; the workload of model prediction is greatly reduced.
According to the signal characteristics of the radio station, mapping long time sequence signals according to a certain time step length, and searching the relation between time periods and segments; and meanwhile, a cyclic neural network structure based on a digital mapping method is provided, before data enters a network, data of the single-frequency channel radio station is classified, mapping of time periods to time points is established, and then corresponding labels are established for corresponding time periods according to the classification of output time points. The improved network greatly improves the accuracy of the prediction result and provides possibility for solar spectrum lossless filtering.
Additional features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
FIG. 1 is a diagram of a recurrent neural network architecture;
FIG. 2 is a diagram of a long-short term memory neural network architecture;
FIG. 3 is a graph of 2017-9-9 solar radio burst events (with international time on the abscissa and 150MHz-500MHz frequency points on the ordinate);
FIG. 4 is a diagram of the location of the initial position of a burst in the 360MHz-380MHz band;
FIG. 5 is a comparison graph of test data predictions for a recurrent neural network based on the LSTM model;
FIGS. 6(a) - (b) are original sequence diagrams of a channel of a 244MHz interference radio station, respectively;
FIG. 7 is a diagram illustrating data set creation based on numerical mapping in an embodiment of the present invention;
FIG. 8 is a graph of simulation event intensity in an embodiment of the present invention;
FIG. 9 is a diagram of the result of RNN prediction based on LSTM network;
FIG. 10 is a diagram showing the prediction results of the label classification method according to the present embodiment;
FIG. 11 is an unprocessed artwork in an embodiment of the present invention;
FIGS. 12(a) - (b) are diagrams of a device containing electrical circuitry near 244MHz and 253MHz, respectively;
FIGS. 13(a) - (b) are diagrams after treatment with stations near 244MHz and 253MHz, respectively;
FIGS. 14(a) - (b) are diagrams after burst section compensation around 244MHz and 253MHz, respectively;
FIG. 15 is a graph of the complete event intensity after processing by all stations in an embodiment of the present invention.
Detailed Description
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
Example one
According to the embodiment of the invention, the embodiment of the solar radio filtering method based on the improved LSTM network comprises the following steps:
(1) acquiring radio station channels needing to be processed, and selecting a set amount of channel data before a solar burst event for each radio station channel;
(2) preprocessing the channel data, and extracting data by using a segmented sliding window;
(3) inputting the extracted data into a trained cyclic neural network model based on a digital mapping method, and outputting a solar radio prediction value during solar burst;
the numerical value appearing in the sequence is regarded as different classes by the cyclic neural network model based on the numerical mapping method, and the output predicted value is the numerical value with the maximum prediction probability in the classes.
The method of this example is described in detail below.
1 theory of correlation
1.1 Recurrent Neural Network (RNN) principle
RNN can be kept on time axis by "loopingThereby realizing the function of memorizing information. FIG. 1 is a diagram of a recurrent neural network and its development in the time domain, where x t Representing the input at time t, A represents the model processing part of the neural network, (which is a repeating module with only one tanh layer), h t For the state of the hidden layer at time t, y t Then represents the output at the corresponding time, w 1 、w 2 、w 3 Is a matrix of weight coefficients between layers. When the network is running, unlike a conventional neural network, the RNN has the parameter w at the same location at different times 1 、w 2 、w 3 The shared network training system has the advantages that parameters required for training are greatly reduced, so that workload is reduced, and the network training speed is increased.
As can be seen from the figure, the hidden layer unit h at the current moment in the forward transfer process of the data t Will receive the hidden layer h of the previous moment at the same time t-1 Information of (2) and input layer information x of the current time t So that the output layer at each instant can contain the characteristics of the past instant. The propagation process of the RNN network can be represented as:
h t =f(w 1 ·x t +w 2 ·h t-1 +b) (1)
y t =f(w 3 ·h t +b) (2)
in the formula: h is a total of t-1 The state of the hidden layer at the previous moment; b represents a bias term; f is a non-linear mapping relationship, commonly referred to as an activation function.
Theoretically, an RNN network can process a sufficiently long time sequence, but since an activation function in the RNN often uses a tanh function and a Sigmoid function, a phenomenon that a gradient disappears easily occurs in an error reverse transfer process, so that the RNN is only suitable for processing a short time sequence signal in practical application. However, in the signal to be processed in this embodiment, if the time sequence characteristics are to be found sufficiently and the prediction is accurate, a long time sequence needs to be processed once, and two special RNN models are compared: both LSTM and GRU effectively alleviate the gradient vanishing problem, GRU convergence speed is fast because of fewer parameters. But LSTM expression performs better with larger data sets. The amount of data in this document is sufficient, and based on this status, LSTM (Long Short-Term Memory) is finally selected.
1.2 Long-short term memory model (LSTM) principle
Compared with the traditional RNN, the long-time memory model replaces the original hidden layer unit with an LSTM cell structure unit, namely, an information gate generated by a Sigmoid function is added into a repeatedly-linked module. The information gate comprises a forgetting gate, an input gate and an output gate. The functional properties are such that the information passing through each gate carries a parameter that controls the amount of information passed to the current neuron and the amount of information distributed to the next neuron.
Fig. 2 is a network structure diagram of LSTM, and the data transfer process in the diagram can be expressed by the following formula:
forget the door:
f t =σ(W xf ·x t +W hf ·h t-1 +b f ) (3)
an input gate:
i t =σ(W xi ·x t +W hi ·h t-1 +b i ) (4)
an output gate:
o t =σ(W xo ·x t +W ho ·h t-1 +b o ) (5)
LSTM cell status:
c t =tanh(W xc ·x t +W hc ·h t-1 +b c ) (6)
C t =f t *C t-1 +i t *c t (7)
hidden layer output:
h t =o t *tanh (C t ) (8)
wherein f, i and o respectively represent the related information of a forgetting gate, an input gate and an output gate, W generally refers to a weight coefficient matrix of the network, b is a bias term, x represents the input of the network, tAnd (C) represents the time, h is the hidden layer state, C is the state value (also called the current candidate memory state value) when the network updates the cell, and C is the state value of the current memory unit of the LSTM network. All three gates use the Sigmoid function as the activation function, while the memory of the network is updated by selecting the tanh function. The output information of the final hidden layer is related to the output value of the output gate and the current memory cell state value. In addition, in comparison with RNN, LSTM, the back propagation algorithm needs to calculate not only the hidden layer state h t Corresponding error gradient, and calculating cell state C t The error gradient of (2).
The application of three information gates in the LSTM network makes the memory cell in the network structure store history information for a longer time, and the formula (7) shows that the LSTM cell structure unit is formed by adding two parts, so that when the accumulated error is calculated, the condition that the result is 0 does not occur, the problem of gradient disappearance is relieved, and the long-time memory function is realized.
2 methods and uses
2.1 data selection
A large solar burst event is observed in 2017, 9 and 0 is observed in the background value of the event in a quiet solar state, so that the data of the date are selected for interference removal processing. The data come from a high-resolution solar radio receiver in the racking meter waveband, the frequency range is 150MHz-500MHz, the frequency resolution is 16kHz, and the time resolution is 10 ms. The data received by the data acquisition card is two paths of signals including a left-handed signal and a right-handed signal which are subjected to digital polarization synthesis operation, the two paths of signals exist in two data channels, and the values of the two paths of signals are similar, so that only one path of signal, namely the left-handed signal, is selected for drawing and processing.
FIG. 3 is an intensity diagram of a selected outbreak event, the abscissa represents international time, the ordinate is a frequency point of 150MHz-500MHz, the color scale axis represents different values corresponding to different colors, and different colors in the intensity diagram represent different solar energy current values after reaction. As can be seen from the figure, there are station interferences with different intensities in channels of some frequency points (32 channels correspond to 1MHz), and stations affecting event observation need to be filtered out by using the method herein. Such as single stations (215MHz, 245MHz, 262MHz, 400MHz, etc.) with signal strength greater than the burst event radio traffic strength and interference segments (360MHz-380MHz, approximately 640 channels) with a wide range of frequencies across the band.
For the above screening of the station signals to be processed, it can be located by formula (9).
Wherein I (t) represents the signal strength value over time in a single channel,and Var (i (t)) represents the mean and standard deviation, respectively, of all signal intensity values over a period of t. Thresh is a threshold value for screening the radio stations, and the size of the threshold value can be selected according to actual conditions.
After the target radio station is screened out, the time of the initiation of the burst event needs to be positioned so as to more accurately mark out a training set in the neural network. The method is characterized in that the intensity values of a plurality of channels where the target radio station is located are averaged and converted into a one-dimensional array which changes along with time. Taking an interfering radio station in a frequency band of 360MHz to 380MHz as an example, as shown in fig. 4, when the sun does not burst, the average value obtained is the average value of the radio station signals, and the variation is not large. After an outbreak event occurs, the average value is the average value after the station and the outbreak value are superposed and is obviously higher than the average value of a pure station, namely the time point pointed by an arrow in the figure, so that the starting position of the outbreak event is determined.
The initial time of the burst is determined and the channel data for the corresponding station (approximately 6000 frames) in the data packet prior to that time is divided into a training set of samples of the station signal. The training set herein consists of the first 6000 multiframe data for the 768 channels of stations, and it should be noted that each time data enters the training network structure, it is the single channel of station signal values.
2.2 station value prediction for Recurrent Neural Networks (RNNDM) based on number mapping
When the recurrent neural network based on the LSTM model predicts the radio station value, the model enters a prediction stage after training is completed, in order to check the network performance, firstly, the prediction is carried out by utilizing test data (the value of the known pure electric station signal position), and the network structure is adjusted through a comparison result with the original data of the corresponding position. Considering that the signal value of the station position to be predicted finally is unknown, the prediction needs to adopt a step prediction method, namely, each time a value is predicted, the value is added to the next input for the second prediction. Referring to fig. 5, in this prediction method, the predicted value and the original data have a generally consistent trend, but the prediction capability of the model gradually decreases with the passage of time and the error introduced by the step prediction method, which leads to the error between the predicted value and the original value gradually increasing.
Because the step prediction method can continuously introduce errors to influence the prediction result, and the characteristics of the interfering radio station signals are observed to find that the data of each channel only has fixed numerical values, the embodiment provides a method for adopting digital mapping to reduce the errors and improve the accuracy of prediction. For example, fig. 6(a) - (b) are the raw sequence diagram and histogram of 6000 frames of data for a certain channel of the interfering station at 244MHz, respectively, both of which show that the station signal values of the channel are composed of four types of numbers 2, 3, 4 and 5 in different arrangements. Therefore, numerical values appearing in the sequence can be regarded as different classes, and the prediction problem is firstly converted into a classification problem to be processed. The predicted value obtained in the way is the value with the maximum prediction probability in the classification, as long as the prediction trend is correct, the accuracy of the predicted value at the moment is 100%, and the problems that the trend is correct in the previous section and the predicted value has deviation can not occur.
The data set establishing process comprises the following steps:
in this embodiment, the training data selects about 6000 frames of data prior to the burst event. The data is also preprocessed before entering the training network. For example, the data is normalized to eliminate the dimensional influence among the data, and the data is normalized to be linearly transformed into the range of [0,1] so as to improve the convergence speed of the network training.
The method considers that the radio station values collected by the actual frequency spectrograph may change at a certain moment due to the influence of some factors, so that the relation between the whole sequence values is influenced, and the prediction error is increased. The embodiment selects to segment the sequence, finds the relation between adjacent time segments, and performs labeling processing on the data in the segment sliding window. The process of data set creation is shown in figure 7.
The drawing is divided into three parts: an input window, an output window, and a segment sliding window. A segmented sliding window comprises a complete input window and a complete output window, the three windows are taken as a whole to slide on a time sequence to obtain values, and the sliding unit is 1. The output window only contains one frame of data and is positioned next to the tail of the input window, namely, the sequence value at the time of (t +1) is predicted according to the sequence information from (t-time +1) to (t), and the sliding window still slides only one frame at a time. After the segmentation sliding value taking, firstly, the mapping of a time period (input window) to a time point (output window) is obtained, and then the classification corresponding to the time point data is converted into a label corresponding to the time period.
And for the data labeling processing in the sliding window, adopting an One-Hot coding mode. The encoding is to convert the category number into binary number, and encode the N category number by using an N-bit register, so as to ensure that only one effective position of each category is activated and marked as 1, and other positions are marked as 0. For example, if the station signal values have four types, N is 4. The numerical values are arranged from small to large and are 2, 3, 4 and 5 in sequence, wherein the first frame data is 3, the seventh frame data is 4, the first frame data is coded into 0100, the seventh frame data is coded into 0010, and the like.
The station value prediction process of the recurrent neural network based on the digital mapping method comprises the following steps:
the LSTM network built in this embodiment is totally divided into five layers, which are an input layer, a full connection layer, an LSTM layer, a full connection layer, and an output layer in sequence. After the input training set passes through a first full-connection layer, the data dimension is converted into the dimension required by the network, then the LSTM layer automatically extracts a series of time sequence features through internal cyclic operation, then the output feature vector enters a second full-connection layer, and finally the network outputs a predicted radio station value.
For better classification, a Dropout function is added in the model to avoid the over-fitting phenomenon of the network, a softmax cross entropy function more suitable for classification operation is selected as a loss function, and an Adam algorithm in an adaptive learning rate optimization algorithm is selected for learning and updating the weight.
Considering that the present embodiment needs to process station signals of 768 channels, which means that 768 network models are called for prediction, and the hyper-parameters involved in the network have time step size time _ step, number of hidden layer nodes unit, training number of single-time network entry batch _ size, network depth num _ layer, learning rate lr, and the like, if each channel adjusts the hyper-parameters to obtain an optimal model, the workload is huge, and therefore, a group of optimal hyper-parameters is screened out through a large number of experiments in the early stage for learning, training and prediction of 768 models.
The sample set for screening the optimal hyper-parameters is composed of 10 channel data randomly extracted from the radio signals of 768 channels, the data used for training of a single channel in the original sample set is about 6000 frames, in order to obtain the predicted accuracy in the experiment, the previous 5750 frames of data are selected as the training set, and the next 250 frames are selected as the test set. Initially, the hyper-parameters are set according to the experience of the predecessor and the random value taking of the signal characteristics of the radio station, and then a network search method is adopted to traverse all hyper-parameter combinations in an exhaustive mode within a certain range. Table 1 shows a part of the parameter tuning results, and the influence of the network depth, the hidden layer node, and the Dropout function on the network prediction result is studied in the case that the learning rate (lr is 0.001) and the time step (time _ step is 35) are fixed.
TABLE 1 number of layers, number of nodes, partial parameter adjustment results
It can be seen from the table that not the deeper the network depth the better. When the number of the nodes is 20 and the network has five layers, the prediction accuracy is only 44.8%, because the network structure is overfitting along with the depth of the network, and after the Dropout function is added, the prediction accuracy is improved to 78.4%, compared with other cases, the prediction accuracy when the Dropout function is added is higher than that when the Dropout function is not added, which is the meaning of adding the Dropout function in this section. A group of hyper-parameter-4-layer networks with the highest prediction accuracy and 25 nodes are selected from the table to continue the parameter adjusting experiment of the learning rate and the time step, and finally the learning rate in the optimal hyper-parameter combination is determined to be 0.001, and the time step is determined to be 20.
3 interference signal suppression results and analysis procedure
3.1 simulation test result analysis
In order to verify the effect of the method used in this embodiment and further verify the good performance of the digital mapping method in the filtering process, a simulation event is artificially created first, and the simulation event is processed and analyzed by using the existing station value prediction method of the recurrent neural network based on the LSTM model and the station value prediction method of the Recurrent Neural Network (RNNDM) based on the digital mapping method of this embodiment, respectively.
According to the linear additive principle of signals, simulation data is formed by superposing pure explosion data without interference and pure electric station data without an explosion section. The pure burst data is data of 471 channels and 351 frames in total, namely, 2019, 9, 6:52: 23.461.7-6: 52:27.131.7 (international time) and 225 MHz-240 MHz. The pure station data is data of 8064 frames of 6 channels including 6 channels of 248.85 MHz-249 MHz in 2017, 9, 6:49: 48.262-6: 51: 12.808.6. And then keeping the outburst frequency band unchanged and the radio station time sequence unchanged, superposing the pure electric station signal value to the position of 227.5MHZ, expanding the pure electric station signal value by 2 times, and superposing the pure electric station signal value to the position of 231MHZ, so as to obtain an outburst event containing electric station interference, wherein the frequency range is 244 MHz-247 MHZ, and the time range is 6:49: 48.262-6: 51: 12.808.6. The simulated event intensity plot is shown in fig. 8.
Firstly, the LSTM network is used to train the station data of the selected 6 channels and predict the station signal value of the burst position, and then the burst position data is used to subtract the predicted value, so as to obtain the processing result graph shown in fig. 9. It can be seen from the intensity map after treatment that some macroscopic interference signals remain in the outbreak event, and the treatment effect is not ideal.
And then processing the interference station signal by using a recurrent neural network based on a digital mapping method. Table 2 shows the accuracy of the prediction result after the 6 channels pass through the classification label, although the accuracy of some channels is as low as 0.5, the average accuracy of the overall data can reach 74.33%. The final processed intensity map is shown in fig. 10, and the interfering station signals in the burst area are basically removed, so that the observation of the event is not influenced. Comparing with fig. 9, the filtering effect of RNN network based on digit mapping method is obviously better than the processing result of LSTM network.
TABLE 2 prediction accuracy for categorised training stations
|
1 | 2 | 3 | 4 | 5 | 6 |
Number of |
3 | 4 | 4 | 3 | 3 | 2 |
Accuracy of classification prediction | 0.78 | 0.50 | 0.60 | 0.74 | 0.87 | 0.97 |
3.2 analysis of actual event processing results
The actual burst event processed is the aforementioned 768 channels of station data, the timing sequence of the processing is in the range of 6:54: 4.617.8-6: 54:46.854.4, the prediction time starts at 6:54:30.507.2, and the data is shown in fig. 11 before being processed. It can be seen from the figure that there are not only strong interference signals with a narrow frequency band but with a signal intensity much greater than the radio current intensity during solar burst, but also interference signals across a wide frequency band of tens of MHz, and these interference radio signals directly cover the burst event, which seriously affects the observation and further research of the event. Through comparison experiment result analysis of simulation events, a recurrent neural network RNNDM method based on digital mapping is used for processing a radio station at an explosion position, and interference signals irrelevant to an explosion section and weaker interference signals are quickly processed by a method of subtracting a flow average value of a corresponding channel.
FIGS. 12(a) - (b) and FIGS. 13(a) - (b) show the results of RNNDM processing for two stations located near 244MHz and 253 MHz. Fig. 12(a) and (b) are raw spectrum graphs without processing, and fig. 13(a) - (b) are graphs obtained by subtracting predicted network values from corresponding stations. Observing a processed result graph, finding that a certain gap exists between the actual effect of the processing and a theoretical result, and the situation of over-suppression of the radio station signal occurs, namely the actual radio station signal value of the burst section is smaller than the predicted value, and when the predicted value of the radio station is subtracted from the original value of the burst position, a part of burst information is also subtracted. This is because the system for acquiring data has a certain nonlinearity, when the solar radiation bursts, the signal intensity increases sharply, and when the signal is amplified, the amplifier may work in a nonlinear region, which may cause the acquired data to be attenuated. In addition, the radio signals are also influenced to a certain extent by the solar radiation outbreak. Therefore, the solar burst signal and the interfering station signal no longer satisfy the linear additivity of the signals, so that some compensation needs to be performed on the processing part.
In this embodiment, the system for acquiring data is a high frequency resolution system, and on average, 32 frequency channels correspond to 1 MHz. The solar burst event usually occupies a large bandwidth in a frequency domain, and is distributed in hundreds of frequency channels in the system, and the high frequency resolution enables the values in adjacent frequency channels to be similar, so that the burst data of adjacent interference radio station signal channels can be selected to compensate the processing result, and the compensation calculation formula is as follows:
wherein F (a, k) is the compensated value, F (a) is the value of the interference signal over-suppression area, g (b) is the value of the adjacent channel of the interference signal, k is the coefficient, and a and b represent different frequency channels. The coefficient k is selected according to actual conditions. Fig. 13(a) - (b) are respectively fig. 14(a) - (b) after compensation.
Although the true value of the radio station in the outbreak area cannot be known, that is, the accuracy of the network prediction result cannot be judged, the simulation experiment result is referred to, and when the predicted average accuracy is about 75%, a certain parameter is added for compensation, so that the experiment result can reach the expected target. The complete intensity map after all station channels have been processed is shown in fig. 15. Compared with the original spectrogram, the explosion event in the processed intensity map is more obvious, particularly in the frequency range of 360MHz-380MHz, the originally explosion region covered by the interference radio station signal in a large range becomes clear after filtering, which is beneficial to the observation and research of an observer on the complete event.
In the method for predicting the interfering radio station signal in the solar spectrogram based on the recurrent neural network with the improved LSTM structure, according to the characteristics of the radio station signal, a long time sequence signal is mapped according to a certain time step length, and the relation between a time period and a segment is searched. Meanwhile, the network is improved, a circular neural network structure based on a digital mapping method is provided, before data enters the network, the data of the single-frequency channel radio station is classified, mapping of time periods to time points is firstly established, and then corresponding labels are established for corresponding time periods according to the classification of the output time points. The improved network greatly improves the accuracy of the prediction result and provides possibility for solar spectrum lossless filtering.
Example two
According to an embodiment of the invention, a solar radio filtering system based on an improved LSTM network is disclosed, comprising:
the data acquisition module is used for acquiring radio station channels needing to be processed, and for each radio station channel, channel data of a set number before a solar burst event is selected;
the data processing module is used for preprocessing the channel data and extracting data by utilizing a segmented sliding window;
the model prediction module is used for inputting the extracted data into a trained cyclic neural network model based on a digital mapping method and outputting a solar radio prediction value during solar burst; the numerical value appearing in the sequence is regarded as different classes by the cyclic neural network model based on the numerical mapping method, and the output predicted value is the numerical value with the maximum prediction probability in the classes.
It should be noted that specific implementation manners of the modules are already described in detail in the first embodiment, and are not described again.
EXAMPLE III
According to an embodiment of the present invention, an embodiment of a terminal device is disclosed, which includes a processor and a memory, the processor being configured to implement instructions; the memory is used for storing a plurality of instructions which are suitable for being loaded by the processor and executing the solar radio filtering method based on the improved LSTM network in the first embodiment.
In other embodiments, a computer-readable storage medium is disclosed, having stored thereon a plurality of instructions adapted to be loaded by a processor of a terminal device and to perform the improved LSTM network based solar radiofiltering method described in the first embodiment.
Although the embodiments of the present invention have been described with reference to the accompanying drawings, it is not intended to limit the scope of the present invention, and it should be understood by those skilled in the art that various modifications and variations can be made without inventive efforts by those skilled in the art based on the technical solution of the present invention.
Claims (7)
1. A solar radio filtering method based on an improved LSTM network is characterized by comprising the following steps:
acquiring radio station channels needing to be processed, and selecting a set amount of channel data before a solar burst event for each radio station channel;
preprocessing the channel data, and extracting data by using a segmented sliding window;
inputting the extracted data into a trained recurrent neural network model based on a digital mapping method, namely RNNDM, and outputting a solar radio predicted value during solar burst;
the numerical value appearing in the sequence is regarded as different classes by the cyclic neural network model based on the digital mapping method, and the output predicted value is the numerical value with the maximum prediction probability in the classes;
the training process of the recurrent neural network model based on the numerical mapping method comprises the following steps:
selecting a set amount of data before a solar outbreak event to construct a data set;
preprocessing the data in the data set, and extracting the data by using a segmented sliding window; labeling the data in the segmented sliding window;
dividing the labeled data into a training set and a prediction set, and training the cyclic neural network model based on the digital mapping method;
traversing all hyper-parameter combinations in an exhaustive mode within a certain range by adopting a network search method, and screening out a group of optimal hyper-parameters for learning, training and predicting of the model;
performing labeling processing on the data in the segmented sliding window by adopting a One-Hot coding mode;
and selecting and utilizing burst data of a signal channel of an adjacent interference radio station to compensate the processing result, wherein a compensation calculation formula is as follows:
wherein F (a, k) is the compensated value, F (a) is the value of the interference signal over-suppression area, g (b) is the value of the adjacent channel of the interference signal, k is the coefficient, a and b represent different frequency channels, and the coefficient k is selected according to the actual situation.
2. The solar radiowave filtering method based on the improved LSTM network as claimed in claim 1, wherein selecting a set number of channel data before a solar burst event includes:
screening out a target radio station by using the mean value and the standard deviation of all signal intensity values in the time period t;
averaging the intensity values of a plurality of channels where the target radio station is located, and converting the intensity values into a one-dimensional array which changes along with time;
selecting a time point with an average value obviously higher than that of a pure electric station as the starting position of a solar burst event;
and selecting a set amount of channel data before the starting time.
3. The solar radio filtering method based on the improved LSTM network as claimed in claim 1, wherein the channel data is preprocessed and the data is extracted by using a segmented sliding window, specifically comprising:
normalizing and normalizing the channel data;
segmenting the sequence, and searching the relation between adjacent time periods; respectively determining an input window, an output window and a segmented sliding window, wherein one segmented sliding window comprises a complete input window and a complete output window, and the three windows as a whole slide on a time sequence to take values; wherein, the output window only contains one frame data and is positioned next to the tail part of the input window;
and obtaining the mapping of the input window to the output window after the segmentation sliding value taking.
4. The solar radiowave filtering method based on the improved LSTM network, according to claim 1, wherein the recurrent neural network model based on the digital mapping method comprises an input layer, a full-link layer, an LSTM layer, a full-link layer and an output layer in sequence;
a Dropout function is added in the cyclic neural network model based on the digital mapping method to avoid the overfitting phenomenon of the network, a loss function is a softmax cross entropy function, and learning and updating of weight values select an Adam algorithm in an adaptive learning rate optimization algorithm.
5. A solar-based radio filtering system based on an improved LSTM network, comprising:
the data acquisition module is used for acquiring radio station channels needing to be processed, and for each radio station channel, channel data of a set number before a solar burst event is selected;
the data processing module is used for preprocessing the channel data and extracting data by utilizing a segmented sliding window;
the model prediction module is used for inputting the extracted data into a trained cyclic neural network model based on a digital mapping method, namely RNNDM, and outputting a solar radio prediction value during solar burst; the numerical value appearing in the sequence is regarded as different classes by the cyclic neural network model based on the digital mapping method, and the output predicted value is the numerical value with the maximum prediction probability in the classes;
the training process of the recurrent neural network model based on the numerical mapping method comprises the following steps:
selecting a set amount of data before a solar outbreak event to construct a data set;
preprocessing the data in the data set, and extracting the data by using a segmented sliding window; labeling the data in the section sliding window;
dividing the labeled data into a training set and a prediction set, and training the cyclic neural network model based on the digital mapping method;
traversing all hyper-parameter combinations in an exhaustive mode within a certain range by adopting a network search method, and screening out a group of optimal hyper-parameters for learning, training and predicting of the model;
performing labeling processing on the data in the segmented sliding window by adopting a One-Hot coding mode;
and selecting and utilizing burst data of a signal channel of an adjacent interference radio station to compensate the processing result, wherein a compensation calculation formula is as follows:
wherein F (a, k) is the compensated value, F (a) is the value of the interference signal over-suppression area, g (b) is the value of the interference signal adjacent channel, k is the coefficient, a and b represent different frequency channels, and the coefficient k is selected according to the actual situation.
6. A terminal device comprising a processor and a memory, the processor being arranged to implement instructions; the memory is configured to store a plurality of instructions adapted to be loaded by the processor and to perform the improved LSTM network based solar radiofiltering method of any of claims 1-4.
7. A computer readable storage medium having stored therein a plurality of instructions, wherein the instructions are adapted to be loaded by a processor of a terminal device and to perform the solar radiofiltering method based on an LSTM network according to any one of claims 1 to 4.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110383050.2A CN113095215B (en) | 2021-04-09 | 2021-04-09 | Solar radio filtering method and system based on improved LSTM network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110383050.2A CN113095215B (en) | 2021-04-09 | 2021-04-09 | Solar radio filtering method and system based on improved LSTM network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113095215A CN113095215A (en) | 2021-07-09 |
CN113095215B true CN113095215B (en) | 2022-08-26 |
Family
ID=76675853
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110383050.2A Expired - Fee Related CN113095215B (en) | 2021-04-09 | 2021-04-09 | Solar radio filtering method and system based on improved LSTM network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113095215B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113804982B (en) * | 2021-09-18 | 2022-06-21 | 山东大学 | Solar radio burst real-time detection method and system based on digital filtering |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109284870A (en) * | 2018-10-08 | 2019-01-29 | 南昌大学 | Short-term method for forecasting photovoltaic power generation quantity based on shot and long term Memory Neural Networks |
CN110632386A (en) * | 2019-08-16 | 2019-12-31 | 山东大学 | Solar radio interference filtering method, readable storage medium and electronic equipment |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170091615A1 (en) * | 2015-09-28 | 2017-03-30 | Siemens Aktiengesellschaft | System and method for predicting power plant operational parameters utilizing artificial neural network deep learning methodologies |
-
2021
- 2021-04-09 CN CN202110383050.2A patent/CN113095215B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109284870A (en) * | 2018-10-08 | 2019-01-29 | 南昌大学 | Short-term method for forecasting photovoltaic power generation quantity based on shot and long term Memory Neural Networks |
CN110632386A (en) * | 2019-08-16 | 2019-12-31 | 山东大学 | Solar radio interference filtering method, readable storage medium and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
CN113095215A (en) | 2021-07-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109949278B (en) | Hyperspectral anomaly detection method based on antagonistic self-coding network | |
CN105427300B (en) | A kind of hyperspectral image abnormal detection method based on low-rank representation and study dictionary | |
CN110865357A (en) | Laser radar echo signal noise reduction method based on parameter optimization VMD | |
CN113627066B (en) | Displacement prediction method for reservoir bank landslide | |
CN112733725A (en) | Hyperspectral image change detection method based on multistage cyclic convolution self-coding network | |
CN109886464B (en) | Low-information-loss short-term wind speed prediction method based on optimized singular value decomposition generated feature set | |
CN113345463B (en) | Speech enhancement method, device, equipment and medium based on convolutional neural network | |
CN113298846A (en) | Intelligent interference detection method based on time-frequency semantic perception | |
CN113158553A (en) | Magnetotelluric signal noise suppression method and system based on CNN-LSTM | |
CN113095215B (en) | Solar radio filtering method and system based on improved LSTM network | |
CN112688746B (en) | Spectrum prediction method based on space-time data | |
CN114636975A (en) | LPI radar signal identification method based on spectrogram fusion and attention mechanism | |
CN114692947A (en) | VMD-Lasso-CNN-LSTM-based user electricity consumption prediction method | |
CN110632386B (en) | Solar radio interference filtering method, readable storage medium and electronic equipment | |
CN111008726A (en) | Class image conversion method in power load prediction | |
CN113126038A (en) | High-frequency ground wave radar working frequency optimization method, system, storage medium and application | |
CN115796351A (en) | Rainfall shorthand prediction method and device based on variational modal decomposition and microwave attenuation | |
CN113052099B (en) | SSVEP classification method based on convolutional neural network | |
CN112751633B (en) | Broadband spectrum detection method based on multi-scale window sliding | |
CN116471154A (en) | Modulation signal identification method based on multi-domain mixed attention | |
CN111104876A (en) | Infrared spectrum deconvolution method based on neural network | |
CN116343032A (en) | Classification method combining Gaussian regression mixed model and MRF hyperspectral function data | |
CN113433514B (en) | Parameter self-learning interference suppression method based on expanded deep network | |
CN114764577A (en) | Lightweight modulation recognition model based on deep neural network and method thereof | |
CN113344245A (en) | Hybrid deep learning short-term prediction model, method, storage medium, and computing device |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20220826 |
|
CF01 | Termination of patent right due to non-payment of annual fee |