CN116961805A - Short wave prediction method based on space-time diagram neural network - Google Patents
Short wave prediction method based on space-time diagram neural network Download PDFInfo
- Publication number
- CN116961805A CN116961805A CN202310843742.XA CN202310843742A CN116961805A CN 116961805 A CN116961805 A CN 116961805A CN 202310843742 A CN202310843742 A CN 202310843742A CN 116961805 A CN116961805 A CN 116961805A
- Authority
- CN
- China
- Prior art keywords
- space
- time
- graph
- data
- matrix
- 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.)
- Pending
Links
- 238000010586 diagram Methods 0.000 title claims abstract description 78
- 238000000034 method Methods 0.000 title claims abstract description 48
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 24
- 238000004891 communication Methods 0.000 claims abstract description 47
- 239000013598 vector Substances 0.000 claims abstract description 16
- 238000012545 processing Methods 0.000 claims abstract description 8
- 238000003062 neural network model Methods 0.000 claims abstract description 7
- 238000000547 structure data Methods 0.000 claims abstract description 6
- 238000012549 training Methods 0.000 claims abstract description 6
- 239000011159 matrix material Substances 0.000 claims description 61
- 238000001228 spectrum Methods 0.000 claims description 15
- 230000008569 process Effects 0.000 claims description 13
- 230000002776 aggregation Effects 0.000 claims description 9
- 238000004220 aggregation Methods 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 9
- 238000005096 rolling process Methods 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 6
- 230000000306 recurrent effect Effects 0.000 claims description 6
- 230000004807 localization Effects 0.000 claims description 5
- 238000005259 measurement Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 230000008878 coupling Effects 0.000 claims description 3
- 238000010168 coupling process Methods 0.000 claims description 3
- 238000005859 coupling reaction Methods 0.000 claims description 3
- 230000007774 longterm Effects 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 230000003595 spectral effect Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 2
- 239000011148 porous material Substances 0.000 claims description 2
- 230000001360 synchronised effect Effects 0.000 claims 2
- 230000004931 aggregating effect Effects 0.000 claims 1
- 230000001131 transforming effect Effects 0.000 claims 1
- 238000007619 statistical method Methods 0.000 description 6
- 238000010276 construction Methods 0.000 description 5
- 230000008901 benefit Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 238000007792 addition Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 238000005314 correlation function Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 230000000149 penetrating effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- 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
- G06N3/042—Knowledge-based neural networks; Logical representations of neural 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- 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
- G06N3/045—Combinations of 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/391—Modelling the propagation channel
- H04B17/3913—Predictive models, e.g. based on neural network models
Abstract
The invention belongs to the technical field of wireless communication, and provides a shortwave prediction method based on a space-time diagram neural network, which comprises the following steps: step 1: collecting, analyzing and processing short-wave communication resources of a plurality of receivers and transmitters within a period of time to obtain short-wave parameter information distributed in different frequency bands; step 2: based on the data processed in the step 1, carrying out graph structure modeling, and representing the relation between the sensor nodes by using a graph structure; step 3: performing embedded reconstruction on the graph structure modeled in the step 2, and converting the nodes into embedded representations in a vector space; step 4: and (3) constructing a graph neural network model locally, and training the graph structure data generated in the step (2) and the step (3) and carrying out local prediction. The method can better capture the relation between space-time data, thereby improving the prediction accuracy, having stronger self-adaptability and generalization capability and being capable of effectively predicting in different scenes.
Description
Technical Field
The invention belongs to the technical field of wireless communication, and particularly relates to a shortwave prediction method based on a space-time diagram neural network in the field.
Background
The short wave communication is a radio wave communication mode, the communication frequency of the short wave communication mode is usually between 3MHz and 30MHz, and the short wave communication mode has the characteristics of long broadcasting distance, strong penetrating power, strong anti-interference capability and the like, so the short wave communication mode is widely applied to the fields of military, aviation, maritime affairs, weather forecast, news broadcasting and the like. The technical principle of short wave communication is that an electric signal is modulated at a transmitting end, converted into electromagnetic waves which can propagate in the air, and then transmitted out through an antenna; the receiving end receives the electromagnetic wave through the antenna and then converts the electromagnetic wave into an original electric signal through a demodulation process. Because the frequency range of the short-wave communication is wide, the communication can be carried out on different frequency bands so as to avoid the waste of frequency spectrum resources; meanwhile, short-wave communication can be encrypted, and the safety of communication is guaranteed. Despite the more advanced communication technology, short-wave communication is still an important communication means, especially in remote areas or when disasters occur, which is a reliable communication means.
The short wave prediction method is to analyze the physical characteristics of the atmosphere, solar activity and other factors so as to predict the change of the short wave channel parameters in a future period of time, and the importance of the technology to short wave communication is self-evident. Because the propagation of the short wave signal is affected by the atmosphere, the parameters of the short wave channel change at any time, and if a communication person knows the parameters of the short wave channel in advance, corresponding measures such as adjusting frequency, power and the like can be taken, so that the success rate and quality of communication are improved. In addition, short wave prediction can provide important support for emergency communication, and conventional communication equipment can be damaged under the conditions of natural disasters, war and the like, and short wave communication can be used as a standby communication means. Therefore, the application of the short wave prediction technology can not only improve the efficiency and reliability of short wave communication, but also provide important guarantee for emergency communication.
The traditional statistical method works on short wave prediction mainly based on historical data and experience rules for analysis and prediction. The traditional statistical method can be used for establishing a mathematical model through statistical analysis of historical data and predicting transmission parameters of future short wave channels, such as frequency, field intensity, signal to noise ratio and the like. Although the traditional statistical method has a certain application value in the field of short wave prediction, the prediction accuracy and reliability of the short wave channel have certain limitations due to instability and complexity of the short wave channel. Therefore, modern shortwave prediction has begun to adopt advanced technologies such as machine learning, so as to improve prediction accuracy and reliability and better meet modern communication requirements.
Disclosure of Invention
The invention mainly aims to provide a short wave prediction method based on a space-time diagram neural network, which solves the problems of efficiency and reliability of short wave communication.
The invention adopts the following technical scheme:
a shortwave prediction method based on a space-time diagram neural network comprises the following steps:
step 1: collecting, analyzing and processing short-wave communication resources of a plurality of receivers and transmitters within a period of time to obtain short-wave parameter information distributed in different frequency bands;
step 2: based on the data processed in the step 1, carrying out graph structure modeling, and representing the relation between the sensor nodes by using a graph structure;
step 3: performing embedded reconstruction on the graph structure modeled in the step 2, and converting the nodes into embedded representations in a vector space;
step 4: and (3) constructing a graph neural network model locally, and training the graph structure data generated in the step (2) and the step (3) and carrying out local prediction.
Further, the step 1 specifically includes:
determining geographic positions of a short-wave communication signal transmitter and a signal receiver, constructing a short-wave communication service scene, and collecting spectrum parameters on the signal transmitter, wherein in general, in order to obtain spectrum data with finer granularity, a plurality of iterations of data in one day are needed to be used as current month data:
x i ＝x 1
wherein x is i For each day of the month, x is the collected month-granularity spectral data, and the month data is a set:
X＝{x 1 ，x 2 ，...，x d }
wherein X is a set of data of one month, and d is the total number of days of the current month;
the traditional spectrum information has the problems of non-smoothness, non-continuity and the like, and is difficult to be directly applied to actual application scenes such as frequency selection, channel parameter prediction and the like. According to the invention, a smooth fitting method is adopted for the original data, so that the spectrum evolution rule can be described more accurately, and the accuracy and reliability of frequency selection and channel parameter prediction are improved.
The smooth fitting mode is obtained by the following conversion:
wherein i is the ith day of one month, the value is an integer from 1 to 31, j is the j-th flow data of one day, the value is an integer from 1 to 48,jth traffic data representing the ith day of m months. />Data representing the j-th traffic of day i of month m+1.
Further, the step 2 specifically includes: and (3) carrying out graph structure modeling based on the data processed in the step (1), and representing the relation among the sensor nodes by using a graph structure. In the process of constructing the graph, the method mainly comprises the following steps: 1) Generating an adjacency matrix aiming at the space and time correlation of the short-wave space-time sequence data stream; 2) And extracting the node characteristics of the graph from the short-wave time-space sequence data stream.
The first point is that the state time sequence, the functional similarity degree and the geographic position similarity degree between every two nodes are used for making a cross-correlation measurement, and the communication strength is represented by the correlation. Connectivity between nodes serves as the value of the adjacency matrix to construct the graph structure.
And the second point is to couple the short wave time-space sequence data stream into a weighted undirected graph, and the acquired data is used as signal vectors on graph nodes to convert the original time-space data into a feature matrix.
Further, the step 3 specifically includes: using the graph structure in step 2, feature vector generation node embedding of neighborhood nodes is recursively aggregated and transformed, mapping each node into a low-dimensional vector representation, while preserving some key information in the original graph.
Embedding operation based on the graph rolling network (Graph Convolution Networks, GCN) is performed by using the graph adjacent matrix and the graph signal matrix generated in the step 2. The spectrum-based GCN is expressed as:
wherein At represents graph adjacency matrix data input At time t, sigma represents activation functions of all layers except the output layer,representing node feature matrix->The initial node characteristic matrix is the graph signal matrix constructed in the step 2. Weight matrix->Updating node feature matrix to +.>And takes it as output. In->Expressed as:
in the method, in the process of the invention,for the degree matrix of the graph, I is the identity matrix, which is an updated formula for a typical graph rolling network.
Generating node embeddings using graph rolling networks is centered on weight matricesIs implemented using a recursive structure based on current and historical information. Will W t (l) As hidden state of the dynamic system, a gating loop unit (Gated Recurrent Unit, GRU) is used to update hidden layer state according to input at system t time, expressed as:
where GRU is a variant of a recurrent neural network (Recurrent Neural Network, RNN) that solves the long-term dependence problem in traditional RNNs by using gating mechanisms. The method constructs a multi-layer GCN update layer, and takes the finally generated matrix as an embedded vector representation.
Further, the step 4 specifically includes: and (3) constructing a locally trained graph neural network model, which is used for training the graph structure data generated in the step (2) and the step (3) and carrying out local prediction.
Using the adjacency matrix generated in step 2To further construct a localized space-time diagram, expressed as:
where 3 represents the time-space diagram of the present invention using 3 adjacent time steps to construct a localization time-space diagram. On the basis of constructing the time-space diagram in the step 2, the same nodes of adjacent time steps are connected, the weight between the nodes is set to be 1, and the time sequence influence on the same node is captured, so that a new localization time-space diagram is constructed.
In order to capture the space-time information in the short wave data, position embedding is added, so that the model can give consideration to the space-time information, and the capacity of modeling the space-time correlation is enhanced. For space-time network series X G ∈R N×C×T Creating a learnable time embedding matrix T emb ∈R N×C×T And a learnable spatial embedding matrix S emb ∈R N×C×T . Adding the two embedding matrices to a spatio-temporal network sequence with broadcast operation, resulting in a new representation of the network sequence:
the space-time diagram synchronization network model constructs a space-time diagram synchronization module to capture localized space-time correlations. The space-time diagram synchronization module consists of a set of diagram convolution operations, which can aggregate the characteristics of each node with its neighboring nodes, the input of which is a diagram signal matrix of the local space-time diagram, each node aggregates the characteristics of itself and neighbors in adjacent time steps, the diagram convolution operations (Graph Convolution, GCON) are defined as:
in the method, in the process of the invention,adjacency matrix representing local space-time diagram, +.>Is the input to the first layer of the picture scroll laminate, W and b are the learnable parameters, σ represents the activation function.
The output of the l-layer convolution operation is finally fed into an Aggregation (AGG) layer, from which the final output is aggregated and cropped, first at an element level max-pore, which can be expressed as:
the output after this aggregation is then clipped, leaving only the middle part, which can be expressed as:
the above operation is performed on the local space-time diagram matrix of every three time steps, and the outputs of all the diagram convolution operations are added together to form an output of a space-time diagram synchronization layer, which can be expressed as:
the final predicted output of the model can be obtained by adding an output layer, and can be expressed as:
wherein X is T For the matrix after the output of the last space-time diagram synchronization layer passes reshape, W and b are the learnable parameters,is the predicted value of the ith time step.
The invention has the advantages in design principle:
compared with the traditional statistical method and the traditional machine learning method, the space-time diagram neural network is a deep learning method based on a diagram structure, can effectively process space-time data, can utilize the correlation of short-wave signals in space-time, can represent the space-time data into the diagram structure, and can learn space-time characteristics by utilizing the diagram neural network. The method can better capture the relation between the space-time data, thereby improving the prediction accuracy. In addition, the space-time diagram neural network can also process data inconsistency between different spatial positions and time steps, so that the space-time diagram neural network has advantages when processing space-time short wave data compared with a traditional statistical method and a traditional machine learning method. And thirdly, the space-time diagram neural network has stronger adaptivity and generalization capability, and can be used for effectively predicting in different scenes. Although the traditional machine learning method still has application value on some problems, the space-time diagram neural network has obvious advantages when processing complex short wave prediction problems.
Further, the invention has the beneficial effects that:
in the acquisition of data, the original data is fully and properly processed, so that the original data shows a smooth fluctuation effect which accords with reality. This step is a critical step in short wave prediction, since only fully processed data can be utilized by subsequent processing methods.
On the basis of the collected short-wave communication data, the nodes are converted into a form of a graph data structure. The graph data structure can fully represent the coupling relation between parameters and the spatial relation between different transmitters and receivers, which lays a foundation for subsequent prediction.
And performing embedded reconstruction on the modeled graph structure, and further converting the graph structure into an embedded representation in a vector space. The graph embedding can be used for describing node characteristics by utilizing the similarity among the nodes, so that the information quantity is increased during prediction, and the model learns to richer characteristic representation.
The graph neural network is used for processing the graph structure formed by the short wave data. The processing method can better capture the relation between the features and has more accurate and better effect on the final short wave prediction. This step is critical to the overall method, since efficient management and utilization of the short-wave channel can only be achieved by accurate prediction.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a schematic flow chart of the present invention;
FIG. 3 is a schematic diagram of the steps for generating node embedment for a graph structure in accordance with the present invention;
FIG. 4 is a schematic diagram of a locally trained neural network model constructed in accordance with the present invention;
FIG. 5 is a graph of adjacency matrix construction results involved in one embodiment of the invention;
fig. 6 is a predicted result of frequency values in one embodiment of the invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Example 1
The embodiment discloses a shortwave prediction method based on a space-time diagram neural network, which comprises the following steps:
step 1: collecting, analyzing and processing short-wave communication resources of a plurality of receivers and transmitters within a period of time to obtain short-wave parameter information distributed in different frequency bands;
step 2: based on the data processed in the step 1, carrying out graph structure modeling, and representing the relation between the sensor nodes by using a graph structure;
step 3: performing embedded reconstruction on the graph structure modeled in the step 2, and converting the nodes into embedded representations in a vector space;
step 4: and (3) constructing a graph neural network model locally, and training the graph structure data generated in the step (2) and the step (3) and carrying out local prediction.
The following details the individual steps
The step 1 specifically comprises the following steps:
data acquisition was performed by the american sound coverage analysis program (Voice of America Coverage Analysis Program, VOACAP), setting a signal emission point as japan Wu Xiongxian (131.13 e,32.56 n), and six signal reception points were set in china as follows: the antennas of the signal transmitting point and the receiving point are set to be three element antennas and eight mesh antennas in a short distance transmission mode, the modulation mode adopted is a CW mode (continuous wave signal), and the frequency, the time delay, the signal to noise ratio and four parameters of a signal received from six preset receiving points transmitted from the Wu Xue county in Japan to China are acquired for the sky wave communication signal of the short wave communication, the frequency, the time delay, the signal to noise ratio and the four parameters are acquired in one half hour, ten month data of six places in 2022 month to 10 month are acquired in one month, and the short wave communication data are generated in month units.
Typically, a plurality of iterations of data of one day are needed as current month data to obtain spectrum data with finer granularity:
x i ＝x 1
wherein x is i For each day of the month, x is the collected month-granularity spectral data, and the month data is a set:
X＝{x 1 ,x 2 ,…,x d }
wherein X is a set of data of one month, and d is the total number of days of the current month;
the traditional spectrum information has the problems of non-smoothness, non-continuity and the like, and is difficult to be directly applied to actual application scenes such as frequency selection, channel parameter prediction and the like. According to the invention, a smooth fitting method is adopted for the original data, so that the spectrum evolution rule can be described more accurately, and the accuracy and reliability of frequency selection and channel parameter prediction are improved.
The smooth fitting mode is obtained by the following conversion:
wherein i is the ith day of one month, the value is an integer from 1 to 31, j is the j-th flow data of one day, the value is an integer from 1 to 48,jth traffic data representing the ith day of m months. />Data representing the j-th traffic of day i of month m+1. The smoothing processing can enable the collected coarse-grained information to be more suitable for actual application scenes, so that the accuracy and reliability of prediction are improved.
The step 2 specifically comprises the following steps:
and (3) carrying out graph structure modeling based on the data processed in the step (1), and representing the relation among the sensor nodes by using a graph structure. In the process of constructing the graph, the method mainly comprises the following steps: 1) Generating an adjacency matrix aiming at the space and time correlation of the short-wave space-time sequence data stream; 2) And extracting the node characteristics of the graph from the short-wave time-space sequence data stream.
In the first place, correlation analysis is often used as a tool to process and analyze time series by virtue of its simple and easy-to-operate nature. The response between nodes is delayed, so that the correlation can be used to represent the communication strength by making a cross-correlation measure for the time series of states between every two nodes.
Consider a time series of equally spaced samplesAnd->To determine the similarity between the two, a cross-correlation function may be defined as
In the formula, tau is equal to or greater than 0, and in the data set, the acquired time delay is tau, and T is the sequence length. Further, the normalized cross-correlation coefficient may be defined as:
the strength of the communication relationship between nodes can be calculated by considering the hysteresis value as τBy usingIndicating that the cross-correlation metric value is>And 0.7, regarding the nodes where the two time sequences are located as connected nodes, calculating the connectivity among different sensor nodes in the mode, and setting the weight among the nodes as 1 when constructing the graph if the nodes are connected, otherwise setting the weight as 0. The adjacency matrix between sensor nodes derived from the state-time correlation metric is expressed as:
the cross-correlation measurement of the degree of functional similarity is obtained by the following method: if the sensors represented by the two nodes are similar in function in an actual short-wave communication scene, the two nodes are regarded as connected nodes, and the weight between the nodes is set to be 1 when the graph is constructed, and otherwise, the weight between the nodes is set to be 0. In the data set of the specific embodiment, the functions of the three sensor nodes are basically similar, so that the cross correlation of the degree of similarity of functions is defined as 1, namely, the three nodes are all regarded as nodes with strong connectivity.
The geographic position similarity cross-correlation measurement is obtained by the following method: if the geographical environments of the sensors represented by the two nodes are similar in an actual short-wave communication scene, the two nodes are regarded as connected nodes, and the weight between the nodes is set to be 1 when the graph is constructed, and otherwise, the weight between the nodes is set to be 0. In the data set of the specific embodiment, the geographical environments of the three sensor nodes are basically similar, so that the cross correlation of the degree of functional similarity is defined as 1, namely, the three nodes are regarded as nodes with strong connectivity.
The adjacency matrix between sensor nodes derived from the above functional similarity degree cross-correlation measure and geographic position similarity degree cross-correlation measure is expressed as:
finally, for the spatial and temporal correlation of short-wave time-space sequence data streams, the generation of the adjacency matrix is expressed as:
as shown in fig. 5, the present example performs the construction of the adjacency matrix on the acquired short-wave data set. Fig. 5 shows the result of the construction of the adjacency matrix for three parameters of frequency, field strength, signal to noise ratio at six receivers.
Second, the present example represents the coupling relationship between electromagnetic spectrum data as a weighted undirected graph g= (V, E, a), where V is a set of n= |v|nodes in space, E is a set of edges representing the association relationship between nodes,is a weighted adjacency matrix of the graph representing the correlation between nodes. the electromagnetic environment condition at time t is denoted as the map signal +.> Where C is the number of feature species studied. Each electromagnetic monitoring device is set as a node (namely the sensor node), the type of the channel parameter collected and monitored on the node is C, and each node on the transformed graph signal matrix is a C-dimensional vector. The goal of this approach can be seen as a sequence of observations for N nodes given P historical time instants
The prediction task can be abstracted as: predicting electromagnetic environment parameters of all N nodes at Q moments in future, namely
The step 3 specifically comprises the following steps:
and (3) performing embedded reconstruction on the graph structure modeled in the step (2), and converting the nodes into embedded representations in a vector space. First, the spectrum-based GCN is expressed as:
wherein A is t Graph adjacency matrix data representing the input at time t,representing node feature matrix->The initial node feature matrix is the graph signal matrix constructed in step 2, and σ represents the activation functions of all layers except the output layer. Weight matrix W using GCN t (l) Updating the node characteristic matrix to +.>And takes it as output. In->Expressed as:
in the method, in the process of the invention,for the degree matrix of the graph, I is the identity matrix, which is an updated formula for a typical graph rolling network.
Generating node embeddings using graph rolling networks is centered on weight matricesIs implemented using a recursive structure based on current and historical information. Will->As a hidden state of the dynamic system, a GRU is used to update a hidden layer state according to an input at a system t time, expressed as:
in the formula, the GRU is a gating circulating unit, and solves the problem of long-term dependence in the traditional RNN by using a gating mechanism. And constructing a plurality of GCN updating layers, and taking the finally generated matrix as an embedded vector representation. Fig. 3 is a schematic structural diagram of this step.
And 4, constructing a locally trained graph neural network model, and training the graph structure data generated in the step 2 and the step 3 and carrying out local prediction.
Using the adjacency matrix generated in step 2To further construct a localized space-time diagram, expressed as:
where 3 represents the time-space diagram of the present invention using 3 adjacent time steps to construct a localization time-space diagram. On the basis of constructing the time-space diagram in the step 2, the same nodes of adjacent time steps are connected, the weight between the nodes is set to be 1, and the time sequence influence on the same node is captured, so that a new localization time-space diagram is constructed.
In order to capture the space-time information in the short wave data, position embedding is added, so that the model can give consideration to the space-time information, and the capacity of modeling the space-time correlation is enhanced. For space-time network series X G ∈R N×C×T Creating a learnable time embedding matrix T emb ∈R N×C×T And a learnable spatial embedding matrix S emb ∈R N×C×T . Adding the two embedding matrices to a spatio-temporal network sequence with broadcast operation, resulting in a new representation of the network sequence:
the space-time diagram synchronization network model constructs a space-time diagram synchronization module to capture localized space-time correlations. The space-time diagram synchronization module is composed of a group of diagram convolution operation, the characteristics of each node and adjacent nodes can be aggregated together by the diagram convolution operation, the input of the diagram convolution operation is a diagram signal matrix of a local space-time diagram, each node aggregates the characteristics of the node and the adjacent nodes in adjacent time steps, and the diagram convolution operation is as follows:
in the method, in the process of the invention,adjacency matrix representing local space-time diagram, +.>Is the input to the first layer of the picture scroll laminate, W and b are the learnable parameters, σ represents the activation function.
The output of all the l-layer convolution operations is finally fed into the AGG aggregation layer, from which the final output is aggregated and cropped, first at an element level max-pooling, which can be expressed as:
the output after this aggregation is then clipped, leaving only the middle part, which can be expressed as:
the above operation is performed on the local space-time diagram matrix of every three time steps, and the outputs of all the diagram convolution operations are added together to form an output of a space-time diagram synchronization layer, which can be expressed as:
the final predicted output of the model may add an output layer, and the final result may be expressed as:
wherein X is T For the matrix after the output of the last space-time diagram synchronization layer passes reshape, W and b are the learnable parameters,is the predicted value of the ith time step. Fig. 4 is a schematic diagram of the network structure at this step.
And (3) further effect verification:
as shown in fig. 6, the present example performs graph construction and embedding work on a short-wave communication data set, and verifies the validity thereof by adopting the graph neural network prediction method described in the present invention. Fig. 6 shows the mean absolute percentage error trend of the frequency parameter predictions under this dataset, fig. 6 using a graph as a predictive performance presentation.
Furthermore, the particular embodiments described herein may vary from one embodiment to another, and the above description is merely illustrative of the structure of the present invention. All such small variations and simple variations in construction, features and principles of the inventive concept are intended to be included within the scope of the present invention. Various modifications or additions to the described embodiments or similar methods may be made by those skilled in the art without departing from the structure of the invention or exceeding the scope of the invention as defined in the accompanying claims.
Claims (5)
1. A shortwave prediction method based on a space-time diagram neural network is characterized by comprising the following steps:
step 1: collecting, analyzing and processing short-wave communication resources of a plurality of receivers and transmitters within a period of time to obtain short-wave parameter information distributed in different frequency bands;
step 2: based on the data processed in the step 1, carrying out graph structure modeling, and representing the relation between the sensor nodes by using a graph structure;
step 3: performing embedded reconstruction on the graph structure modeled in the step 2, and converting the nodes into embedded representations in a vector space;
step 4: and (3) constructing a graph neural network model locally, and training the graph structure data generated in the step (2) and the step (3) and carrying out local prediction.
2. The short wave prediction method based on space-time diagram neural network as set forth in claim 1, wherein the step 1 specifically includes:
determining geographic positions of a short-wave communication signal transmitter and a signal receiver, constructing a short-wave communication service scene, and collecting spectrum parameters on the signal transmitter, wherein in order to obtain spectrum data with finer granularity, a plurality of iterations of data in one day are needed to be used as current month data:
x i ＝x 1
wherein x is i For each day of the month, x is the collected month-granularity spectral data, and the month data is a set:
X＝{x 1 ,x 2 ,…,x d }
wherein X is a set of data of one month, and d is the total number of days of the current month;
a smooth fitting method is adopted for the original data, and a spectrum evolution rule is described;
the smooth fitting mode is obtained by the following conversion:
wherein i is the ith day of one month, the value is an integer from 1 to 31, j is the j-th flow data of one day, the value is an integer from 1 to 48,jth traffic data representing ith day of m months,/day of m months>Data representing the j-th traffic of day i of month m+1.
3. The short wave prediction method based on space-time diagram neural network according to claim 1, wherein the step 2:
in the process of constructing the graph, the method mainly comprises the following steps:
1) Generating an adjacency matrix aiming at the space and time correlation of the short-wave space-time sequence data stream;
the communication strength is represented by the correlation by performing cross-correlation measurement on the state time sequence between every two nodes and the functional similarity and the geographic position similarity; connectivity between nodes serves as values of an adjacency matrix to construct a graph structure;
2) Extracting graph node characteristics from the short-wave time-space sequence data stream;
and coupling the short-wave space-time sequence data stream into a weighted undirected graph, taking the acquired data as signal vectors on graph nodes, and converting the original space-time data into a feature matrix.
4. The short wave prediction method based on space-time diagram neural network according to claim 1, wherein the step 3 is specifically: using the graph structure in the step 2, recursively aggregating and transforming feature vector generation node embedding of neighborhood nodes, mapping each node into a low-dimensional vector representation, and simultaneously retaining certain key information in the original graph;
performing a graph rolling network (Graph Convolution Networks, GCN) based embedding operation using the graph adjacency matrix and the graph signal matrix generated in step 2, the spectrum based GCN expressed as:
wherein A is t Graph adjacency matrix data input at time t is represented, sigma represents the activation functions of all layers except the output layer,representing node feature matrix->The initial node characteristic matrix is the graph signal matrix constructed in the step 2; weight matrix->Updating node feature matrix to +.>And takes it as output; in->Expressed as:
in the method, in the process of the invention,the degree matrix of the graph is represented by I, which is an identity matrixAn updated formula for a typical graph rolling network;
generating node embeddings using graph rolling networks is centered on weight matricesIs realized by using a recursive structure according to the current and the history information; will->As hidden state of the dynamic system, a gating loop unit (Gated Recurrent Unit, GRU) is used to update hidden layer state according to input at system t time, expressed as:
wherein the GRU is a variant of a recurrent neural network (Recurrent Neural Network, RNN) that solves the long-term dependence problem in conventional RNNs by using gating mechanisms; and constructing a plurality of GCN updating layers, and taking the finally generated matrix as an embedded vector representation.
5. The short wave prediction method based on space-time diagram neural network according to claim 1, wherein the step 4 is specifically:
using the adjacency matrix generated in step 2To further construct a localized space-time diagram, expressed as:
where 3 represents a space-time diagram using 3 adjacent time steps to construct a localized space-time diagram; on the basis of constructing a time-space diagram in the step 2, the same nodes of adjacent time steps are connected, the weight between the nodes is set to be 1, and the time sequence influence on the same node is captured, so that a new localization time-space diagram is constructed;
position embedding is added for capturing space-time information in the short wave data, so that the model can give consideration to the space-time information, and the modeling capacity of the space-time correlation is enhanced; for space-time network series X G ∈R N×C×T Creating a learnable time embedding matrix T emb ∈R N×C×T And a learnable spatial embedding matrix S emb ∈R N×C×T The method comprises the steps of carrying out a first treatment on the surface of the Adding the two embedding matrices to a spatio-temporal network sequence with broadcast operation, resulting in a new representation of the network sequence:
the space-time diagram synchronous network model constructs a space-time diagram synchronous module to capture localized space-time correlation; the space-time diagram synchronization module consists of a set of diagram convolution operations, which aggregate the characteristics of each node with its neighbors, the input of which is a diagram signal matrix of the local space-time diagram, each node aggregates its own and neighbor characteristics in adjacent time steps, the diagram convolution operations (Graph Convolution, GCON) being defined as:
in the method, in the process of the invention,adjacency matrix representing local space-time diagram, +.>Is the input of the first picture volume layer, W and b are the learnable parameters, sigma represents the activation function;
the output of the l-layer convolution operation is finally fed into an Aggregation (AGG) layer, from which the final output is aggregated and cropped, first at an element level max-pore, denoted as:
the output after this aggregation is then clipped, leaving only the middle part, denoted:
the above operation is performed on the local space-time diagram matrix of every three time steps, and then the outputs of all the diagram convolution operations are added together to form an output of a space-time diagram synchronization layer, which is expressed as:
the final predicted output of the model is finally obtained by adding an output layer, and is expressed as:
wherein X is T For the matrix after the output of the last space-time diagram synchronization layer passes reshape, W and b are the learnable parameters,is the predicted value of the ith time step.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310843742.XA CN116961805A (en) | 2023-07-10 | 2023-07-10 | Short wave prediction method based on space-time diagram neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310843742.XA CN116961805A (en) | 2023-07-10 | 2023-07-10 | Short wave prediction method based on space-time diagram neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116961805A true CN116961805A (en) | 2023-10-27 |
Family
ID=88455854
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310843742.XA Pending CN116961805A (en) | 2023-07-10 | 2023-07-10 | Short wave prediction method based on space-time diagram neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116961805A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117808128A (en) * | 2024-02-29 | 2024-04-02 | 浪潮电子信息产业股份有限公司 | Image processing method, federal learning method and device under heterogeneous data condition |
-
2023
- 2023-07-10 CN CN202310843742.XA patent/CN116961805A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117808128A (en) * | 2024-02-29 | 2024-04-02 | 浪潮电子信息产业股份有限公司 | Image processing method, federal learning method and device under heterogeneous data condition |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112152948B (en) | Wireless communication processing method and device | |
| CN103347268B (en) | Based on the self-adapting compressing reconstructing method of energy efficiency observation in cognitive sensor network | |
| CN116961805A (en) | Short wave prediction method based on space-time diagram neural network | |
| CN104703216A (en) | Energy-efficient-based detecting method of multi-task Bayes compressed sensing broadband spectrum | |
| CN114240000A (en) | Air quality prediction method based on space-time graph convolution network | |
| CN113673769A (en) | Graph neural network traffic flow prediction method based on multivariate time sequence interpolation | |
| CN112819236B (en) | Low-orbit satellite Internet of things active terminal prediction method based on neural network | |
| CN113446988A (en) | Airport runway pavement state monitoring system and method based on cloud edge fusion framework | |
| CN111242276A (en) | One-dimensional convolution neural network construction method for load current signal identification | |
| Elawaad et al. | DeepCReg: Improving cellular-based outdoor localization using CNN-based regressors | |
| CN115065992A (en) | Edge calculation-based cooperative sensing and resource allocation method | |
| Jacoby et al. | Short-term prediction of the attenuation in a commercial microwave link using LSTM-based RNN | |
| CN113610329B (en) | Short-time rainfall approaching forecasting method of double-current convolution long-short term memory network | |
| CN116720156A (en) | Weather element forecasting method based on graph neural network multi-mode weather data fusion | |
| CN115314925A (en) | Intelligent power grid fusion network flow prediction method based on GNN and LSTM | |
| CN115908547A (en) | Wireless positioning method based on deep learning | |
| CN115379478A (en) | Robust energy consumption optimization method based on RIS (RIS) assisted data and energy simultaneous transmission network | |
| CN111935766B (en) | Wireless network flow prediction method based on global spatial dependency | |
| CN114706144A (en) | Marine environment forecasting method, device, system and storage medium | |
| CN113596757A (en) | Rapid high-precision indoor fingerprint positioning method based on integrated width learning | |
| CN114764610A (en) | Channel estimation method based on neural network and communication device | |
| CN114915526B (en) | Communication signal modulation identification method, device and system | |
| CN115549823B (en) | Radio environment map prediction method | |
| CN113406583B (en) | Approximate calculation method for cloud MIMO radar target detection probability | |
| Tegos et al. | Nonlinear energy harvesting evaluation through the logit Pearson distribution |
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 |