CN114422059B - Channel prediction method, device, electronic equipment and storage medium - Google Patents

Channel prediction method, device, electronic equipment and storage medium Download PDF

Info

Publication number
CN114422059B
CN114422059B CN202210082374.7A CN202210082374A CN114422059B CN 114422059 B CN114422059 B CN 114422059B CN 202210082374 A CN202210082374 A CN 202210082374A CN 114422059 B CN114422059 B CN 114422059B
Authority
CN
China
Prior art keywords
channel
historical
matrix
time slot
historical time
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202210082374.7A
Other languages
Chinese (zh)
Other versions
CN114422059A (en
Inventor
戴凌龙
蒋浩
崔铭尧
张涵
王飞
杜洋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tsinghua University
Huawei Technologies Co Ltd
Original Assignee
Tsinghua University
Huawei Technologies Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Tsinghua University, Huawei Technologies Co Ltd filed Critical Tsinghua University
Priority to CN202210082374.7A priority Critical patent/CN114422059B/en
Publication of CN114422059A publication Critical patent/CN114422059A/en
Application granted granted Critical
Publication of CN114422059B publication Critical patent/CN114422059B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • Electromagnetism (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

The invention provides a channel prediction method, a device, an electronic device and a storage medium, wherein the channel prediction method is applied to an MIMO base station, and the method comprises the following steps: creating a channel prediction model; training the channel prediction model based on a channel training data set to obtain a trained channel prediction model, wherein the trained channel prediction model comprises an encoder and a decoder; predicting a future time slot channel in real time based on the encoder, the decoder, and a historical time slot channel. The channel prediction method can effectively solve the problems of rate performance loss caused by channel outdating.

Description

Channel prediction method, device, electronic equipment and storage medium
Technical Field
The present invention relates to the field of wireless mobile communications technologies, and in particular, to a channel prediction method and apparatus, an electronic device, and a storage medium.
Background
In order to meet the increasing business requirements, the mobile communication is carried out by using the extremely high bandwidth provided by high frequency bands such as millimeter wave (adopted by the standard of 30GHz-300GHz, 5G), terahertz (0.1 THz-10 THz) and the like, and the method becomes an important technical means of a future mobile communication network. However, in the frequency bands such as millimeter waves and terahertz waves with rich spectrum resources, serious path loss exists in wireless propagation. Massive multiple-input multiple-output (MIMO) technology is recognized as one of the key technologies to overcome this challenge.
However, in the conventional all-digital MIMO architecture, each antenna needs a dedicated rf link (including mixer, digital-to-analog converter, etc.) for supporting, and the power consumption is often large and the price is not very high. In order to reduce the number of radio frequencies of the system and alleviate the bottleneck problems of high power consumption and high cost, the hybrid precoding structure is considered as an almost only feasible solution for the practical application of massive MIMO.
However, as the frequency band is increased and the moving speed of the user is increased, the channel coherence time is shortened, so that the transmission period of the reference signal is likely to be shorter than the channel coherence time, which causes mismatching between the channel estimated from the received reference signal and the actual channel and causes the channel outdating problem.
Disclosure of Invention
The invention provides a channel prediction method, a channel prediction device, electronic equipment and a storage medium, which are used for solving the defects of rate loss caused by channel outdating in the prior art, realizing the real-time prediction of a channel and effectively solving the problem of rate performance loss caused by channel outdating.
The invention provides a channel prediction method, which is applied to an MIMO base station and comprises the following steps: creating a channel prediction model; training the channel prediction model based on a channel training data set to obtain a trained channel prediction model, wherein the trained channel prediction model comprises an encoder and a decoder; predicting a future time slot channel in real time based on the encoder, the decoder, and a historical time slot channel.
According to a channel prediction method provided by the present invention, the historical time slot channels include a first historical time slot channel and a second historical time slot channel, wherein the first historical time slot channel includes the second historical time slot channel, and the predicting future time slot channels in real time based on the encoder, the decoder and the historical time slot channels comprises: generating beams on a plurality of strongest channel angle lattice points based on energy for a channel angle domain of the MIMO base station; collecting a first number of the first historical time-slot channels based on the beams and feeding the first historical time-slot channels to the encoder in time-slot order, resulting in first historical channel characteristics for the first historical time-slot channels; feeding a second number of the second historical time-slot channels to the decoder, resulting in second historical channel characteristics for the second historical time-slot channels; inputting a first number of the first historical channel characteristics to the decoder, and predicting the future time slot channel in real time by the decoder in combination with a second number of the second historical channel characteristics, wherein the first number is greater than the second number.
According to a channel prediction method provided by the present invention, the encoder includes a self-attention mechanism processing layer, the feeding the first historical time-slot channel to the encoder, obtaining a first historical channel characteristic related to the first historical time-slot channel, including: determining a sequence of channel state information for the first historical time-slotted channel; vectoring the channel state information sequence of the first historical time slot channel to obtain a first vector matrix related to the first historical time slot channel; and performing feature extraction on the first historical time slot channel based on the self-attention mechanism processing layer and the first vector matrix to obtain first historical channel features of the first historical time slot channel.
According to a channel prediction method provided by the present invention, the encoder further includes a normalization processing layer, and the feature extraction for the first historical timeslot channel based on the self-attention mechanism processing layer and the first vector matrix includes: determining a key matrix, a query matrix and a value matrix for the first historical timeslot channel based on the first vector matrix, a linear transformation matrix corresponding to the key matrix, a linear transformation matrix corresponding to the query matrix and a linear transformation matrix corresponding to the value matrix; obtaining an attention matrix about the first historical time slot channel through the key matrix and the query matrix of the first historical time slot channel based on the self-attention mechanism processing layer; and extracting the characteristics of the first historical time slot channel based on the normalization processing layer, the attention matrix of the first historical time slot channel and the value matrix.
According to a channel prediction method provided by the present invention, the encoder further includes a full connection layer, and the feature extraction of the first historical timeslot channel based on the normalization processing layer, the attention matrix of the first historical timeslot channel, and the value matrix includes: obtaining an implicit variable related to the first historical time slot channel through the attention matrix and the value matrix of the first historical time slot channel based on the normalization processing layer; and extracting the characteristics of the first historical time slot channel based on the fully-connected layer and the hidden variables of the first historical time slot channel.
According to a channel prediction method provided by the present invention, the decoder comprises a masked self-attention mechanism processing layer, the feeding of a second number of the second historical time slot channels to the decoder results in a second historical channel characteristic for the second historical time slot channel, comprising: determining a sequence of channel state information for the second historical time-slotted channel; vectorizing the channel state information sequence of the second historical time slot channel to obtain a second vector matrix related to the second historical time slot channel; and performing feature extraction on the second historical time slot channel based on the mask self-attention mechanism processing layer and the second vector matrix to obtain second historical channel features related to the second historical time slot channel.
According to the channel prediction method provided by the present invention, the decoder further includes a normalization processing layer, and the feature extraction for the second historical timeslot channel based on the mask attention mechanism processing layer and the second vector matrix includes: determining a key matrix, a query matrix, and a value matrix for the second historical timeslot channel based on the second vector matrix, a linear transformation matrix corresponding to the key matrix, a linear transformation matrix corresponding to the query matrix, and a linear transformation matrix corresponding to the value matrix; obtaining a masked attention matrix related to the second historical time slot channel through the key matrix and the query matrix of the second historical time slot channel based on the masked self-attention mechanism processing layer; and extracting the characteristics of the second historical time slot channel based on the normalization processing layer, the masked attention matrix of the second historical time slot channel and the value matrix.
According to a channel prediction method provided by the present invention, the decoder further includes a full attention mechanism processing layer, and the predicting the future time slot channel in real time by the decoder includes: obtaining a key matrix, a query matrix and a value matrix related to the future time slot channel based on the first historical channel characteristics, the second historical channel characteristics, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix and the linear transformation matrix corresponding to the value matrix; obtaining hidden variables related to the future time slot channel through a key matrix, a query matrix and a value matrix of the future time slot channel based on the full attention mechanism processing layer and the normalization processing layer; predicting the future time slot channel based on the implicit variable of the future time slot channel.
According to a channel prediction method provided by the present invention, obtaining a key matrix, a query matrix and a value matrix for the future time slot channel based on the first historical channel characteristic, the second historical channel characteristic, a linear transformation matrix corresponding to a key matrix, a linear transformation matrix corresponding to a query matrix and a linear transformation matrix corresponding to a value matrix is implemented by the following formulas:
Figure BDA0003486510050000041
Q＝W q Z (d)
Figure BDA0003486510050000042
wherein K represents a key matrix of the future time slot channel; w k A linear transformation matrix corresponding to the key matrix is represented;
Figure BDA0003486510050000051
representing the first historical channel characteristics(ii) a Q represents a look-up matrix for the future time-slot channel; w q Representing a linear transformation matrix corresponding to the query matrix; z (d) Representing the second historical channel characteristics; v represents a matrix of values for the future time slot channel; w v And the value matrix corresponds to a linear transformation matrix.
According to the channel prediction method provided by the invention, the training of the channel prediction model based on the channel training data set to obtain the trained channel prediction model comprises the following steps:
dividing the channel training data set into channel training samples and channel identification verification samples;
calculating a loss function of a normalized mean square error based on the prediction results for the channel training samples and the channel identification verification samples;
adjusting parameters of the channel prediction model based on a gradient postback algorithm and the loss function to obtain the trained channel prediction model, wherein the loss function is determined by the following formula:
Figure BDA0003486510050000052
wherein L (Θ) represents the loss function,
Figure BDA0003486510050000053
representing the channel identification verification sample,
Figure BDA0003486510050000054
representing a prediction result with respect to the channel training samples.
The present invention also provides a channel prediction apparatus applied to a MIMO base station, the apparatus comprising: a creation module for creating a channel prediction model; the processing module is used for training the channel prediction model based on a channel training data set to obtain a trained channel prediction model, wherein the trained channel prediction model comprises an encoder and a decoder; and the prediction module is used for predicting a future time slot channel in real time based on the encoder, the decoder and the historical time slot channel.
The present invention also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the channel prediction method as described in any of the above when executing the program.
The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the channel prediction method as described in any of the above.
The invention also provides a computer program product comprising a computer program which, when executed by a processor, performs the steps of the channel prediction method as described in any one of the above.
According to the channel prediction method, the device, the electronic equipment and the storage medium, the channel prediction model is trained offline, and the future time slot channel is predicted online in real time based on the encoder and the decoder in the trained channel prediction model, so that accurate channel prediction can be realized, and the problems of rate performance loss caused by channel outdating are effectively solved.
Drawings
In order to more clearly illustrate the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
FIG. 1 is a flow chart of a channel prediction method according to the present invention;
fig. 2 is one of application scenarios of the channel prediction method provided by the present invention;
FIG. 3 is a second exemplary view of the channel prediction method according to the present invention;
FIG. 4 is a schematic diagram of a process for predicting a future time slot channel in real time based on an encoder, a decoder and a historical time slot channel according to the present invention;
FIG. 5 is a schematic flow chart of feeding a first historical time-slot channel to an encoder to obtain a first historical channel characteristic related to the first historical time-slot channel, according to the present invention;
FIG. 6 is a schematic flow chart of feature extraction performed on a first historical timeslot channel based on a self-attention mechanism processing layer and a first vector matrix according to the present invention;
FIG. 7 is a schematic diagram of a flow structure of a self-attention mechanism provided by the present invention;
FIG. 8 is a schematic flow chart of feature extraction performed on a first historical time slot channel based on a normalization processing layer and an attention matrix of the first historical time slot channel according to the present invention;
FIG. 9 is a schematic flow chart of predicting a future time slot channel in real time by a decoder according to the present invention;
FIG. 10 is a schematic diagram of a channel prediction model of a transformer according to the present invention;
FIG. 11 is a graph illustrating the performance of the application channel prediction method provided by the present invention;
fig. 12 is a schematic structural diagram of a channel prediction apparatus provided in the present invention;
fig. 13 is a schematic structural diagram of an electronic device provided by the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
In order to greatly increase channel capacity, a multiple-input multiple-output (MIMO) technology uses multiple antennas at both transmitting and receiving ends, and forms an antenna system with multiple channels between the transmitting and receiving ends.
In order to meet the increasing business requirements, the mobile communication is carried out by using the extremely high bandwidth provided by high frequency bands such as millimeter wave (adopted by the standard of 30GHz-300GHz, 5G), terahertz (0.1 THz-10 THz) and the like, and the method becomes an important technical means of a future mobile communication network. However, in the frequency bands such as millimeter waves and terahertz with rich spectrum resources, serious path loss exists in wireless transmission. Taking terahertz signals in the 0.16THz frequency band as an example, the propagation process of the terahertz signals will experience serious path loss as high as 80 dB/km. Large-scale multiple-input multiple-output (MIMO) technology is recognized as one of the key technologies to overcome this challenge. By configuring a super-large-scale antenna array (for example, 256 antennas), the massive MIMO technology forms a directional beam with extremely high array gain, which can compensate the path loss of a high frequency band and improve the spectral efficiency of the system.
However, in the conventional all-digital MIMO structure, each antenna needs a dedicated rf link (including a mixer, a digital-to-analog converter, etc.) for supporting, and the power consumption is often very large and the price is very low. If the conventional structure is directly applied to a massive MIMO system configured with hundreds of antennas, a huge radio frequency network will be required, and the power consumption and cost thereof will be unacceptable.
In order to reduce the number of radio frequencies of the system and alleviate the bottleneck problems of high power consumption and high cost, the hybrid precoding structure is considered as an almost only feasible solution for the practical application of massive MIMO. The hybrid precoding structure decomposes the traditional high-dimensional all-digital precoding into two steps, namely firstly carrying out high-dimensional analog beamforming (realized by a phase shifting network) to obtain array gain, and then carrying out low-dimensional digital precoding on a baseband after a small amount of radio frequency sampling to eliminate interference among data streams.
However, as the frequency band is increased and the moving speed of the user is increased, the channel coherence time is shortened, so that the transmission period of the reference signal is likely to be shorter than the channel coherence time, which causes mismatching between the channel estimated from the received reference signal and the actual channel and causes the channel outdating problem. For example, when the frequency is 28GHz and the user moving speed is 60km/h, the channel coherence time is 0.32ms, and the reference signal transmission period is 0.625ms at minimum. The channel coherence time is less than the reference signal transmission period and the actual channel is likely to change significantly, resulting in loss of system and rate performance. At present, although some schemes based on an autoregressive model or a deep learning model exist, the schemes cannot extract a long-period rule of channel change so as to solve the problem.
The invention designs a new channel prediction frame which can extract the long-period characteristic of channel change so as to accurately predict a channel, thereby solving the performance loss caused by the outdated problem of the channel.
The present invention will be described with reference to the following embodiments.
Fig. 1 is a flow chart of a channel prediction method according to the present invention.
In an exemplary embodiment of the present invention, the channel prediction method may be applied to a MIMO base station, where the MIMO base station may be a centralized MIMO base station or a distributed MIMO base station, the connection manner of the radio frequency link and the antenna unit includes, but is not limited to, full connection, sub connection or dynamic connection, the network is not limited to single carrier and multiple carriers, and the frequency band is not limited to Sub-6G, millimeter wave or terahertz. As shown in fig. 1, the channel prediction method may include steps 110 to 130, which will be described separately below.
In step 110, a channel prediction model is created.
In one embodiment, the channel prediction model may be a transform-based channel prediction model, wherein the transform-based channel prediction model may include a transform encoder and a transform decoder.
In step 120, the channel prediction model is trained based on the channel training data set to obtain a trained channel prediction model, where the trained channel prediction model may include an encoder and a decoder.
In one embodiment, the channel prediction model may be trained off-line from the base station. The training mode is to use the marked data (corresponding to the channel training data set) to perform supervised training. Wherein, the marked data can adopt a channel generated by a standard model or an actual channel collected by a base station.
In another embodiment, the channel prediction model is trained based on the channel training data set, and the trained channel prediction model may be obtained by:
dividing a channel training data set into a channel training sample and a channel identification verification sample;
calculating a loss function of the normalized mean square error based on the prediction result of the channel training sample and the channel identification verification sample;
based on a gradient back-transmission algorithm and a loss function, adjusting parameters of the channel prediction model to obtain a trained channel prediction model, wherein the loss function can be determined by the following formula:
Figure BDA0003486510050000091
wherein L (Θ) represents a loss function,
Figure BDA0003486510050000101
representing a sample of the channel identification verification,
Figure BDA0003486510050000102
representing the prediction results for the channel training samples.
In step 130, a prediction is made in real time for a future time slot channel based on the encoder, decoder, and historical time slot channels.
In one embodiment, the online prediction phase may make real-time predictions of future channels (corresponding to future time-slot channels). In the application process, according to the historical channel (corresponding to the historical time slot channel), the future channel is predicted through an encoder and a decoder of the trained channel prediction model. The historical channel acquisition method includes, but is not limited to, channel estimation methods such as a minimum mean square error method, a least square method, and the like, and even the historical pilot received by the base station may be used as an input instead of the historical channel.
The channel prediction method provided by the invention can realize accurate channel prediction by training the channel prediction model offline and predicting the future time slot channel on the basis of the encoder and the decoder in the trained channel prediction model online in real time, and effectively solves the problems of rate performance loss caused by channel outdating.
To further describe the channel prediction method provided by the present invention, the following description will be made with reference to fig. 2 to 3.
Fig. 2 is a schematic view of an application scenario of the channel prediction method provided in the present invention; fig. 3 is a schematic diagram of a second application scenario of the channel prediction method provided by the present invention.
As can be seen from fig. 2, in an example, it can be assumed that the number of antennas is N and the number of rf links is N RF The number of users is K, and the number of antennas of each user is N k . The channel prediction framework provided by the invention can comprise two stages of off-line training and on-line prediction. In the off-line training stage, the channel prediction model can be trained in a supervised learning manner. Specifically, a channel generated according to a 3GPP standard channel model or a continuous channel actually acquired by the base station may be used as the training data (corresponding to the channel training data set). Training samples N for each iteration s And (3) a group of P + L channels of continuous time slots, wherein the former P channels are used as historical channels, and the latter L channels are used as future predicted channels. The strongest N of the channel is due to the negligible distance the user moves over several reference signal transmission periods RF The angular grid points can also be regarded as being constant, and only the channel gain at each grid point changes. Therefore, the high-dimensional channel prediction problem of the massive MIMO can be converted into the low-dimensional equivalent channel prediction problem. Predicting the output of the future L channels according to the historical P channels as input, which can be expressed as:
Figure BDA0003486510050000111
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003486510050000112
representing the equivalent channel predicted by the model at the T + T time, f Θ For the prediction model function representation, Θ is a parameter of the prediction model. The loss function of the Normalized Mean Square Error (NMSE) calculated by the output L channel prediction results and the corresponding training data labels can be expressed as
Figure BDA0003486510050000113
Wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003486510050000114
the label representing the equivalent channel at time T + T is then trained using the gradient back-propagation algorithm. And when the NMSE does not descend any more, ending the off-line training stage to obtain the trained channel prediction model.
As can be seen from fig. 3, in another example, the prediction model obtained by offline training may be configured at the base station to perform online real-time channel prediction. Similar to the off-line training process, the high-dimensional antenna frequency domain channel is first converted into the low-dimensional angle delay domain channel. And predicting the change of the future equivalent channel gain at each angle lattice point. Secondly, the change of the equivalent channel of L time slots in the future can be predicted by using the equivalent channel estimated by historical P time slots. In the application process, the predicted channel can be obtained according to the model output, so that the predicted channel can be used for precoding instead of the outdated channel. Thus, the channel prediction framework provided solves the problems of channel outdating and rate loss.
The present invention will be described with reference to the following embodiments, which are used to describe a process for predicting a future time slot channel in real time based on an encoder, a decoder and a historical time slot channel.
Fig. 4 is a schematic flow chart of predicting a future time slot channel in real time based on an encoder, a decoder and a historical time slot channel according to the present invention.
In an exemplary embodiment of the present invention, the historical time slot channels may include a first historical time slot channel and a second historical time slot channel, wherein the first historical time slot channel includes the second historical time slot channel. It will be appreciated that the second historical channel of time slots is part of the first historical channel of time slots. As shown in fig. 4, predicting the future time slot channel in real time based on the encoder, the decoder and the historical time slot channel may include steps 410 to 440, which will be described separately below.
In step 410, beams on the strongest channel angle lattice points are generated based on the energy of the channel angle domain for the MIMO base station.
In one embodiment, N may be generated sequentially in order of the channel angle domain energy magnitude RF (number of radio frequency links) the beams at the strongest channel angle lattice points.
In step 420, a first number of first historical time slot channels is collected based on the beams and fed to the encoder in time slot order, resulting in a first historical channel characteristic for the first historical time slot channel.
In one embodiment, at the selected angular grid point, P channels of historical time slots (corresponding to a first number of channels of a first historical time slot, where the first number is P) may be collected based on the generated beams. And feeding the channels of the P historical time slots into an encoder according to the time sequence, and obtaining the historical channel characteristics (corresponding to the first historical channel characteristics) of the P time slots after processing.
In step 430, a second number of second historical time slot channels is fed to the decoder, resulting in second historical channel characteristics for the second historical time slot channels.
In step 440, a first number of first historical channel characteristics is input to the decoder, and a future time slot channel is predicted by the decoder in real time in combination with a second number of second historical channel characteristics, wherein the first number is greater than the second number.
In one embodiment, G second historical time slot channels may be fed to the decoder and a second historical channel for the second historical time slot channel is obtainedAnd (5) characterizing. And G second historical time slot channels are part of the P first historical time slot channels, and G is smaller than P. Further, the decoder combines G (G) input to the decoder according to the P first historical channel characteristics about the first historical time slot channel obtained by the encoder<P) second historical channel characteristics, generating predictions of channel gains for L slots in the future. Repeating the above steps until the selected N is predicted to be completed RF The channels at the strongest channel angle lattice points. In the embodiment, based on a transformer channel prediction model, a long-term channel is mined to mine a long-period change rule of the channel, so that accurate prediction of a future channel is realized.
The present invention will be described with reference to the following embodiments, wherein the first historical timeslot channel is fed to the encoder, and the first historical channel characteristic of the first historical timeslot channel is obtained.
Fig. 5 is a schematic flow chart of feeding a first historical time slot channel to an encoder to obtain a first historical channel characteristic related to the first historical time slot channel according to the present invention.
In an exemplary embodiment of the invention, the encoder may include a self-attention mechanism processing layer. Feeding the first historical time-slotted channel to the encoder, obtaining the first historical channel characteristics for the first historical time-slotted channel may include steps 510 through 530, each of which will be described separately below.
In step 510, a sequence of channel state information for a first historical time-slotted channel is determined.
In one embodiment, a sequence of channel state information for a first number (e.g., P) of first historical time slot channels may be determined. During the application process, the input of the coder is the channel state information sequence of the first historical time slot channel, and the output of the coder is the extracted internal correlation characteristic of the channel state information sequence of the first historical time slot channel
Figure BDA0003486510050000131
Wherein, the channel state information sequence of the first historical timeslot channel can be represented as:
Figure BDA0003486510050000132
in step 520, a vectoring operation is performed on the channel state information sequence of the first historical timeslot channel to obtain a first vector matrix for the first historical timeslot channel.
In one embodiment, for the encoder, each historical channel information (the sequence of channel state information corresponding to the first historical timeslot channel) may be first processed
Figure BDA0003486510050000133
Vectorization operations are performed to obtain a simple vector representation, which can be expressed as:
Figure BDA0003486510050000141
and deriving a first vector matrix for the first historical timeslot channel based on the vectoring operated channel state information sequence, which may be expressed as:
Figure BDA0003486510050000142
in step 530, feature extraction is performed on the first historical time slot channel based on the self-attention mechanism processing layer and the first vector matrix, and first historical channel features related to the first historical time slot channel are obtained.
In one embodiment, on the basis of the first vector matrix, feature extraction may be performed on the first historical time slot channel by using a self-attention mechanism based on a self-attention mechanism processing layer, so as to obtain a first historical channel feature about the first historical time slot channel.
To further describe the channel prediction method provided by the present invention, the following will describe a process of performing feature extraction on the first historical timeslot channel based on the self-attention mechanism processing layer and the first vector matrix in conjunction with the following embodiments.
Fig. 6 is a schematic flowchart of feature extraction performed on a first historical timeslot channel based on a self-attention mechanism processing layer and a first vector matrix according to the present invention.
In an exemplary embodiment of the present invention, the encoder may further include a normalization processing layer. Feature extraction for the first historical timeslot channel based on the self-attention mechanism processing layer and the first vector matrix may include steps 610 through 630, each of which is described separately below.
In step 610, a key matrix, a query matrix, and a value matrix for the first historical timeslot channel are determined based on the first vector matrix, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix, and the linear transformation matrix corresponding to the value matrix.
In one embodiment, a key matrix K for a first historical time-slot channel may be extracted (e) Query matrix Q (e) And a value matrix V (e) 。
Wherein the key matrix K (e) Query matrix Q (e) And a value matrix V (e) Can be determined by the following formulas, respectively:
Figure BDA0003486510050000143
Figure BDA0003486510050000151
Figure BDA0003486510050000152
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003486510050000153
respectively representing the linear transformation matrix corresponding to the key matrix, the query matrix and the value matrix.
In step 620, an attention matrix for the first historical time-slot channel is obtained through the key matrix and the query matrix of the first historical time-slot channel based on the self-attention mechanism processing layer.
In step 630, feature extraction is performed on the first historical time-slot channel based on the normalization processing layer, the attention matrix of the first historical time-slot channel, and the value matrix.
In one embodiment, as can be seen in conjunction with FIG. 7, the key matrix K of the first historical time slot channel obtained may be based on the attention-free mechanism processing layer (e) And query matrix Q (e) Obtaining an attention matrix E for the first historical time slot channel (e) . Where the input X in fig. 7 is the first vector matrix and the output O in fig. 7 is the output for the self-attention mechanism. Wherein the attention matrix E (e) The following formula may be used to determine:
Figure BDA0003486510050000154
wherein d represents a key matrix K (e) Of (c) is calculated.
Further, the attention matrix E of the first historical time slot channel can be based on the normalization processing layer (e) And a value matrix V (e) And extracting the characteristics of the first historical time slot channel.
Attention matrix E (e) Is a P × P square matrix with the i-th row and j-th column of elements E (e) [i,j]Indicating the attention weight of the jth input to the ith input. To obtain the internal features of the input, a value-pair matrix V is determined according to the attention weights in the attention matrix (e) Summing, the output from the attention mechanism can be obtained.
However, it is difficult to train the transform prediction model by simple operations in the self-attention mechanism because the instability of data distribution leads to the gradient disappearance problem. In one embodiment, residual concatenation and layer normalization methods may be employed in the encoder to solve this problem.
The present invention will be described with reference to the following embodiments for performing a feature extraction process on a first historical time slot channel based on a normalization processing layer and an attention matrix of the first historical time slot channel.
Fig. 8 is a schematic flowchart of feature extraction performed on the first historical time slot channel based on the attention matrix of the normalization processing layer and the first historical time slot channel.
In an exemplary embodiment of the present invention, the encoder may further include a full connection layer. As shown in fig. 8, the feature extraction for the first historical timeslot channel based on the normalization processing layer and the attention matrix of the first historical timeslot channel may include steps 810 and 820, which will be described separately below.
In step 810, based on the normalization processing layer, an implicit variable related to the first historical time slot channel is obtained through the attention matrix and the value matrix of the first historical time slot channel.
In step 820, feature extraction is performed on the first historical timeslot channel based on the fully connected layer and the hidden variables of the first historical timeslot channel.
In one embodiment, considering residual concatenation and layer normalization, the input and normalization can be added to obtain a hidden variable representation of the input (corresponding to the first historical timeslot channel):
Figure BDA0003486510050000161
wherein the layers are normalized
Figure BDA0003486510050000162
μ j And
Figure BDA0003486510050000163
are each E (e) [:,j]The expectation and variance of (c).
Figure BDA0003486510050000164
Representing a first vector matrix, V, for a first historical time-slotted channel (e) A matrix of values representing a channel with respect to a first historical time slot.
In yet another embodiment, to avoid the denominator being zero, a small amount e is added to the denominator. In addition, considering a two-layer fully connected network (FC), further extracting and synthesizing the features of each historical input can be expressed as:
Figure BDA0003486510050000165
in this embodiment, residual join and layer normalization are considered again.
In the application, after the encoder finishes extracting the features of the first historical time slot channel, the utilization of the features extracted by the encoder by the decoder is considered. First, a sequence is input at the encoder
Figure BDA0003486510050000171
Middle sample with length G (G)<P) (corresponding to the second historical timeslot channel), which is used as input to the decoder together with a zero padding of length L, can be expressed as follows:
Figure BDA0003486510050000172
the decoder aims at the characteristics of the historical channel and the encoder
Figure BDA0003486510050000173
Generating a predicted channel sequence of length L at zero-padding positions
Figure BDA0003486510050000174
Similar to the encoder, in order to obtain the decoder input
Figure BDA0003486510050000175
The auto-supervision mechanism is used by the decoder to derive a second historical channel characteristic for a second historical time-slotted channel in the following manner.
In an exemplary embodiment of the invention, the decoder may include a masked self-attention mechanism processing layer. Feeding the second historical time-slot channel to the decoder, deriving a second historical channel characteristic for the second historical time-slot channel may comprise the following steps.
Determining a sequence of channel state information for a second historical time-slot channel;
vectorizing the channel state information sequence of the second historical time slot channel to obtain a second vector matrix related to the second historical time slot channel;
and performing feature extraction on the second historical time slot channel based on the mask self-attention mechanism processing layer and the second vector matrix to obtain second historical channel features related to the second historical time slot channel.
In an exemplary embodiment of the present invention, the decoder may further include a normalization processing layer. Based on the mask self-attention mechanism processing layer and the second vector matrix, the feature extraction of the second historical time slot channel can be realized by the following modes:
determining a key matrix, a query matrix, and a value matrix for a second historical timeslot channel based on the second vector matrix, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix, and the linear transformation matrix corresponding to the value matrix;
based on the mask self-attention mechanism processing layer, obtaining a post-mask attention matrix related to the second historical time slot channel through the key matrix and the query matrix of the second historical time slot channel;
and extracting the characteristics of the second historical time slot channel based on the normalization processing layer, the masked attention moment array of the second historical time slot channel and the value matrix.
In one embodiment, a channel state information sequence for the second historical timeslot channel may be determined and a vectoring operation may be performed on the channel state information sequence for the second historical timeslot channel to obtain a second vector matrix for the second historical timeslot channel. Further, key matrices K for the second historical time-slot channels may be separately extracted based on the second vector matrices (d) Query matrix Q (d) Sum matrix V (d) And is determined by the following formula:
Figure BDA0003486510050000181
Figure BDA0003486510050000182
Figure BDA0003486510050000183
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003486510050000184
the matrix is composed of a second vector matrix of a second historical time slot channel and filled zeros. For convenience of explanation, the following may be mentioned
Figure BDA0003486510050000185
A second vector matrix, equivalent to a second historical time-slot channel, is illustrated.
Figure BDA0003486510050000186
Representing a linear transformation matrix corresponding to the key matrix,
Figure BDA0003486510050000187
A linear transformation matrix corresponding to the query matrix,
Figure BDA0003486510050000188
A linear transformation matrix corresponding to the matrix of values is represented.
Further, a mask-based self-attention mechanism processing layer and a key matrix K for a second historical time-slot channel (d) Query matrix Q (d) An attention matrix for the second historical timeslot channel is derived and determined by the following equation:
Figure BDA0003486510050000189
and an attention matrix E for the first historical timeslot channel (e) In contrast, a Mask (Mask) operation is added to avoid the attention of the historical channel to the future channel, obey the causal rule and avoid the disorder of the training process. During the application process, the attention matrix E about the second historical time slot channel can be used (d) Further, the hidden variable input by the decoder is obtained and determined by the following formula:
Figure BDA0003486510050000191
further, a second historical channel characteristic for the second historical timeslot channel may be derived based on a hidden variable. Due to hidden variable Z (d) And is positively correlated with the second historical channel characteristic, and for convenience of explanation, the second historical channel characteristic may be expressed as Z (d) 。
Based on encoder derived information in order to exploit long-term historical channel information at the decoder
Figure BDA0003486510050000192
And decoder derived Z (d) The full attention mechanism between the encoder and the decoder is realized, and the long-term historical channel information is fully utilized by paying attention to the characteristics provided by the encoder to obtain accurate prediction of a future channel.
The present invention will be described with reference to the following embodiments for a process of predicting a future time slot channel in real time by a decoder.
Fig. 9 is a schematic flow chart of predicting a future time slot channel in real time by a decoder according to the present invention.
In an exemplary embodiment of the present invention, the decoder may further include a full attention mechanism processing layer. As shown in fig. 9, predicting the future time slot channel in real time by the decoder may include steps 910 to 930, each of which will be described separately below.
In step 910, a key matrix, a query matrix, and a value matrix for the future time slot channel are obtained based on the first historical channel characteristics, the second historical channel characteristics, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix, and the linear transformation matrix corresponding to the value matrix.
In one embodiment, obtaining the key matrix, the query matrix, and the value matrix for the future time slot channel based on the first historical channel characteristic, the second historical channel characteristic, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix, and the linear transformation matrix corresponding to the value matrix is implemented by the following formulas:
Figure BDA0003486510050000193
Q＝W q Z (d) (20)
Figure BDA0003486510050000201
wherein K represents a key matrix for a future time slot channel; w is a group of k A linear transformation matrix corresponding to the key matrix;
Figure BDA0003486510050000202
representing the first historical channel characteristic; q represents a look-up matrix for a future time-slot channel; w is a group of q Representing a linear transformation matrix corresponding to the query matrix; z is a linear or branched member (d) Representing a second historical channel characteristic; v represents a value matrix of a future time slot channel; w v And the value matrix corresponds to a linear transformation matrix.
In step 920, based on the full attention mechanism processing layer and the normalization processing layer, the hidden variables about the future time slot channel are obtained through the key matrix, the query matrix and the value matrix of the future time slot channel.
In step 930, a future time-slot channel is predicted based on the hidden variables of the future time-slot channel.
In one embodiment, it can be obtained by an encoder respectively
Figure BDA0003486510050000203
And decoder derived Z (d) A key matrix K, a query matrix Q and a value matrix V are obtained. For the specific calculation formula, please refer to formula (19) to formula (21).
Further, by establishing and encoding the output
Figure BDA0003486510050000204
Parallel attention mechanisms, each future channel will naturally have a long-term correlation with the historical channel. The predicted channel sequence (corresponding to the future slot channel) can be obtained as follows:
Figure BDA0003486510050000205
Figure BDA0003486510050000206
due to the fact that
Figure BDA0003486510050000207
The length of the sequence is G + L, and L vectors after the sequence are the predicted channel
Figure BDA0003486510050000208
The matrix-form channel can be obtained by real number-to-complex number and de-directional quantization. With this embodiment, predicted channels (corresponding to future slot channels) for L slots in the future can be obtained.
To further describe the channel prediction method provided by the present invention, the following embodiments are described below.
Fig. 10 is a schematic structural diagram of a channel prediction model of a transformer provided in the present invention.
In an exemplary embodiment of the present invention, the channel prediction method applied in the present invention may be applied to a channel prediction architecture, where the channel prediction architecture may include a transformer encoder and a transformer decoder.
In the application process, the input of the transformer encoder is the channel state information sequence H1 (comprising H (T-P + 1), H (T-G + 1) \ 8230; \8230; H (T)) of the first historical timeslot channel. The channel state information sequence H1 of the first historical time slot channel is sequentially processed by the self-attention mechanism to obtain an attention matrix E about the first historical time slot channel (e) . However, it is difficult to train the transformer prediction model due to the instability of the data distribution resulting in the gradient disappearance problem. The method of residual concatenation and layer normalization can be adopted in the encoder to solve the problem, and obtain an implicit variable related to the input H1, and obtain a channel characteristic H1 (e) related to the first historical timeslot channel through the implicit variable related to the input H1, where the channel characteristic H1 (e) corresponds to the above-mentioned method
Figure BDA0003486510050000211
The transform decoder input H2 (comprising H (T-G + 1), H (T) \8230; 0) comprises the channel state information sequence of the second historical timeslot channel. Here, a length G (G < P) history channel sequence (corresponding to the second history slot channel) may be sampled in the encoder input sequence H1, and the sequence and the length L zero padding are used together as the input H2 of the transform decoder. The goal of the decoder is to generate a predicted channel sequence H2 (d) of length L at the zero-padded position based on the historical channel and the encoder's characteristics H1 (e), where H2 (d) includes H2 (T + 1) \ 8230; \ 8230; H2 (T + L).
Similar to the encoder, to obtain the characteristics of the decoder input H2, an auto-supervision mechanism is used by the decoder, extracting the key matrix K on the decoder input H2 (d) Query matrix Q (d) Sum matrix V (d) . And based on a key matrix K (d) Query matrix Q (d) Sum matrix V (d) Determining an attention matrix E (d) . Further, based on the attention matrix E (d) An implicit variable Z (d) of the decoder input H2 is determined.
In order to utilize long-term historical channel information at the decoder, a full attention mechanism between the encoder and the decoder is realized based on the encoder-derived H1 (e) and the decoder-derived Z (d), and accurate prediction of a future channel is obtained by fully utilizing the long-term historical channel information by paying attention to the characteristics provided by the encoder. Specifically, the key matrix K, the query matrix Q, and the value matrix V are obtained by H1 (e) obtained by the encoder and Z (d) obtained by the decoder, respectively. For the specific calculation formula, please refer to formula (19) to formula (21).
By establishing an attention mechanism in parallel with the encoder output H1 (e), each future channel will naturally have a long-term correlation with the historical channel. The predicted channel sequence can be obtained as follows
Figure BDA0003486510050000221
H2(d)＝LN(Z+FC(Z)) (25)
Since the length of the H2 (d) sequence is G + L, L vectors after the sequence are the predicted channel H2 (T + 1) \ 8230; \ 8230; H2 (T + L). The matrix-form channel can be obtained by real number-to-complex number conversion and de-directional quantization.
As shown in fig. 11, the sum-rate performance of the channel prediction scheme (corresponding to the temporal Neural ODE in the diagram) for large-scale MIMO proposed by the present invention is improved by 50% compared with the conventional auto-regression or deep learning method, and the throughput performance under the channel change scene is comprehensively improved.
According to the above description, the channel prediction method provided by the invention can realize accurate channel prediction by training the channel prediction model offline and predicting the future time slot channel on the basis of the encoder and the decoder in the trained channel prediction model online in real time, and effectively solve the problems of rate performance loss caused by channel obsolescence.
Based on the same conception, the invention also provides a channel prediction device.
The following describes the channel prediction apparatus provided by the present invention, and the channel prediction apparatus described below and the channel prediction method described above may be referred to correspondingly.
Fig. 12 is a schematic structural diagram of a channel prediction apparatus provided in the present invention.
In an exemplary embodiment of the present invention, the channel prediction apparatus may be applied to a MIMO base station, where the MIMO base station may be a distributed MIMO base station or a centralized MIMO base station. As shown in fig. 12, the channel prediction apparatus may include a creation module 1210, a processing module 1220, and a prediction module 1230, each of which will be described below.
The creation module 1210 may be configured for creating a channel prediction model.
The processing module 1220 may be configured to train a channel prediction model based on a channel training data set, resulting in a trained channel prediction model, where the trained channel prediction model may include an encoder and a decoder.
The prediction module 1230 may be configured to predict future time slot channels in real time based on the encoder, decoder, and historical time slot channels.
In an exemplary embodiment of the present invention, the historical time-slot channels may include a first historical time-slot channel and a second historical time-slot channel, wherein the first historical time-slot channel includes the second historical time-slot channel, and the prediction module 1230 may predict the future time-slot channel in real time based on the encoder, the decoder and the historical time-slot channels in the following manner:
generating beams on a plurality of strongest channel angle lattice points based on energy with respect to a channel angle domain of the MIMO base station; collecting a first number of first historical time slot channels based on the beam, and feeding the first historical time slot channels to the encoder in time slot order to obtain first historical channel characteristics about the first historical time slot channels; feeding the second historical time slot channel to a decoder, obtaining second historical channel characteristics about the second historical time slot channel; a first number of first historical channel characteristics is input to a decoder, and a future time slot channel is predicted by the decoder in real time in combination with a second number of second historical channel characteristics, wherein the first number is greater than the second number.
In an exemplary embodiment of the invention, the encoder may include a self-attention mechanism processing layer, and the prediction module 1230 may feed the first historical timeslot channel to the encoder in the following manner, resulting in a first historical channel characteristic for the first historical timeslot channel:
determining a sequence of channel state information for a first historical time-slot channel; vectorizing a channel state information sequence of a first historical time slot channel to obtain a first vector matrix about the first historical time slot channel; and performing feature extraction on the first historical time slot channel based on the self-attention mechanism processing layer and the first vector matrix to obtain first historical channel features related to the first historical time slot channel.
In an exemplary embodiment of the present invention, the encoder may further include a normalization processing layer, and the prediction module 1230 may perform feature extraction on the first historical timeslot channel based on the self-attention mechanism processing layer and the first vector matrix in the following manner:
determining a key matrix, a query matrix, and a value matrix for a first historical time slot channel based on the first vector matrix, a linear transformation matrix corresponding to the key matrix, a linear transformation matrix corresponding to the query matrix, and a linear transformation matrix corresponding to the value matrix; obtaining an attention matrix related to the first historical time slot channel through the key matrix and the query matrix of the first historical time slot channel based on the self-attention mechanism processing layer; and extracting the characteristics of the first historical time slot channel based on the normalization processing layer, the attention matrix and the value matrix of the first historical time slot channel.
In an exemplary embodiment of the present invention, the encoder may further include a full connection layer, and the prediction module 1230 may perform feature extraction on the first historical timeslot channel based on the normalization processing layer, the attention matrix of the first historical timeslot channel, and the value matrix in the following manner:
on the basis of the normalization processing layer, obtaining an implicit variable related to the first historical time slot channel through the attention matrix and the value matrix of the first historical time slot channel; and extracting the characteristics of the first historical time slot channel based on the fully connected layer and the hidden variables of the first historical time slot channel.
In an exemplary embodiment of the invention, the decoder may include a masking self-attention mechanism processing layer, and the prediction module 1230 may feed a second number of second historical time-slot channels to the decoder, resulting in a second historical channel characteristic for the second historical time-slot channel, in the following manner:
determining a sequence of channel state information for a second historical timeslot channel; vectorizing the channel state information sequence of the second historical time slot channel to obtain a second vector matrix related to the second historical time slot channel; and performing feature extraction on the second historical time slot channel based on the mask self-attention mechanism processing layer and the second vector matrix to obtain second historical channel features related to the second historical time slot channel.
In an exemplary embodiment of the present invention, the decoder may further include a normalization processing layer, and the prediction module 1230 may perform feature extraction on the second historical timeslot channel based on the masking autofocusing mechanism processing layer and the second vector matrix in the following manner:
determining a key matrix, a query matrix and a value matrix for the second historical timeslot channel based on the second vector matrix, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix and the linear transformation matrix corresponding to the value matrix; obtaining a post-mask attention matrix related to the second historical time slot channel through the key matrix and the query matrix of the second historical time slot channel based on the mask self-attention mechanism processing layer; and extracting the characteristics of the second historical time slot channel based on the normalization processing layer, the masked attention moment array of the second historical time slot channel and the value matrix.
In an exemplary embodiment of the present invention, the decoder may further include a full attention mechanism processing layer, and the prediction module 1230 may predict the future time slot channel in real time through the decoder in the following manner:
obtaining a key matrix, a query matrix and a value matrix about a future time slot channel based on the first historical channel characteristic, the second historical channel characteristic, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix and the linear transformation matrix corresponding to the value matrix; based on the full attention machine mechanism processing layer and the normalization processing layer, obtaining hidden variables related to the future time slot channel through a key matrix, a query matrix and a value matrix of the future time slot channel; and predicting the future time slot channel based on the hidden variable of the future time slot channel.
In an exemplary embodiment of the present invention, the predicting module 1230 may obtain the key matrix, the query matrix and the value matrix for the future time slot channel based on the first historical channel characteristic, the second historical channel characteristic, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix and the linear transformation matrix corresponding to the value matrix by using the following formulas:
Figure BDA0003486510050000251
Q＝W q Z (d) (27)
Figure BDA0003486510050000252
wherein K represents a key matrix of a future time slot channel; w k A linear transformation matrix corresponding to the key matrix;
Figure BDA0003486510050000253
representing a first historical channel characteristic; q represents a look-up matrix for future time slot channels; w q Representing a linear transformation matrix corresponding to the query matrix; z (d) Representing a second historical channel characteristic; v represents a value matrix of a future time slot channel; w is a group of v And the value matrix corresponds to a linear transformation matrix.
In an exemplary embodiment of the present invention, the processing module 1220 may train the channel prediction model based on the channel training data set in the following manner to obtain a trained channel prediction model:
dividing a channel training data set into a channel training sample and a channel identification verification sample; calculating a loss function of the normalized mean square error based on the prediction result of the channel training sample and the channel identification verification sample; adjusting parameters of the channel prediction model based on a gradient postback algorithm and a loss function to obtain a trained channel prediction model, wherein the loss function is determined by the following formula:
Figure BDA0003486510050000261
wherein L (Θ) represents a loss function,
Figure BDA0003486510050000262
representing a sample of the channel identification verification,
Figure BDA0003486510050000263
representing the prediction results for the channel training samples.
Fig. 13 illustrates a physical structure diagram of an electronic device, and as shown in fig. 13, the electronic device may include: a processor (processor) 1310, a communication Interface (Communications Interface) 1320, a memory (memory) 1330, and a communication bus 1340, wherein the processor 1310, the communication Interface 1320, and the memory 1330 communicate with each other via the communication bus 1340. Processor 1310 may invoke logic instructions in memory 1330 to perform a channel prediction method, wherein the channel prediction method is applied to a MIMO base station, the method may comprise: creating a channel prediction model; training a channel prediction model based on a channel training data set to obtain a trained channel prediction model, wherein the trained channel prediction model comprises an encoder and a decoder; the future time slot channel is predicted in real time based on the encoder, decoder and historical time slot channel.
In addition, the logic instructions in the memory 1330 may be implemented in software functional units and stored in a computer readable storage medium when sold or used as a stand-alone product. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.
In another aspect, the present invention also provides a computer program product, which includes a computer program, the computer program can be stored on a non-transitory computer readable storage medium, when the computer program is executed by a processor, a computer can execute the channel prediction method provided by the above methods, wherein the channel prediction method is applied to a MIMO base station, and the method can include: creating a channel prediction model; training a channel prediction model based on a channel training data set to obtain a trained channel prediction model, wherein the trained channel prediction model comprises an encoder and a decoder; based on the encoder, the decoder and the historical time slot channel, the future time slot channel is predicted in real time.
In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, the computer program being implemented by a processor to perform the channel prediction method provided by the above methods, wherein the channel prediction method is applied to a MIMO base station, and the method may include: creating a channel prediction model; training a channel prediction model based on a channel training data set to obtain a trained channel prediction model, wherein the trained channel prediction model comprises an encoder and a decoder; the future time slot channel is predicted in real time based on the encoder, decoder and historical time slot channel.
The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.
It is further to be understood that while operations are depicted in the drawings in a particular order, this is not to be understood as requiring that such operations be performed in the particular order shown or in serial order, or that all illustrated operations be performed, to achieve desirable results. In certain environments, multitasking and parallel processing may be advantageous.
Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims (12)

1. A channel prediction method is applied to a MIMO base station, and the method comprises the following steps:
creating a channel prediction model;
training the channel prediction model based on a channel training data set to obtain a trained channel prediction model, wherein the trained channel prediction model comprises an encoder and a decoder;
predicting, in real-time, a future time slot channel based on the encoder, the decoder, and a historical time slot channel, wherein,
the historical time slot channels comprise a first historical time slot channel and a second historical time slot channel, wherein the first historical time slot channel comprises the second historical time slot channel, and the predicting future time slot channels in real time based on the encoder, the decoder and the historical time slot channels comprises:
generating beams on a plurality of strongest channel angle lattice points based on energy for a channel angle domain of the MIMO base station;
collecting a first number of the first historical time-slot channels based on the beams and feeding the first historical time-slot channels to the encoder in time-slot order, resulting in first historical channel characteristics for the first historical time-slot channels;
feeding a second number of the second historical time-slot channels to the decoder, resulting in second historical channel characteristics for the second historical time-slot channels;
inputting a first number of the first historical channel characteristics to the decoder, and predicting the future time slot channel in real time by the decoder in combination with a second number of the second historical channel characteristics, wherein the first number is greater than the second number.
2. The channel prediction method of claim 1, wherein the encoder comprises a self-attention mechanism processing layer, and wherein feeding the first historical time-slot channel to the encoder results in a first historical channel characteristic for the first historical time-slot channel, comprising:
determining a sequence of channel state information for the first historical time-slotted channel;
vectoring the channel state information sequence of the first historical time slot channel to obtain a first vector matrix related to the first historical time slot channel;
and performing feature extraction on the first historical time slot channel based on the self-attention mechanism processing layer and the first vector matrix to obtain first historical channel features of the first historical time slot channel.
3. The channel prediction method of claim 2, wherein the encoder further comprises a normalization processing layer, and wherein the feature extraction of the first historical timeslot channel based on the self-attention mechanism processing layer and the first vector matrix comprises:
determining a key matrix, a query matrix and a value matrix for the first historical timeslot channel based on the first vector matrix, a linear transformation matrix corresponding to the key matrix, a linear transformation matrix corresponding to the query matrix and a linear transformation matrix corresponding to the value matrix;
obtaining an attention matrix about the first historical time slot channel through the key matrix and the query matrix of the first historical time slot channel based on the self-attention mechanism processing layer;
and performing feature extraction on the first historical time slot channel based on the normalization processing layer, the attention matrix of the first historical time slot channel and the value matrix.
4. The channel prediction method of claim 3, wherein the encoder further comprises a full connection layer, and the feature extraction for the first historical timeslot channel based on the normalization processing layer, the attention matrix of the first historical timeslot channel, and the value matrix comprises:
obtaining an implicit variable related to the first historical time slot channel through the attention matrix and the value matrix of the first historical time slot channel based on the normalization processing layer;
and extracting the characteristics of the first historical time slot channel based on the fully-connected layer and the hidden variables of the first historical time slot channel.
5. The channel prediction method of claim 2, wherein the decoder comprises a masked self-attention mechanism processing layer, and wherein feeding the second number of the second historical timeslot channels to the decoder results in a second historical channel characteristic for the second historical timeslot channel, comprising:
determining a sequence of channel state information for the second historical time-slotted channel;
vectorizing the channel state information sequence of the second historical time slot channel to obtain a second vector matrix related to the second historical time slot channel;
and performing feature extraction on the second historical time slot channel based on the mask self-attention mechanism processing layer and the second vector matrix to obtain second historical channel features related to the second historical time slot channel.
6. The channel prediction method of claim 5, wherein the decoder further comprises a normalization processing layer, and the feature extraction of the second historical timeslot channel based on the mask self-attention mechanism processing layer and the second vector matrix comprises:
determining a key matrix, a query matrix and a value matrix for the second historical timeslot channel based on the second vector matrix, a linear transformation matrix corresponding to the key matrix, a linear transformation matrix corresponding to the query matrix and a linear transformation matrix corresponding to the value matrix;
obtaining a masked attention matrix related to the second historical time slot channel through the key matrix and the query matrix of the second historical time slot channel based on the masked self-attention mechanism processing layer;
and extracting the characteristics of the second historical time slot channel based on the normalization processing layer, the masked attention matrix of the second historical time slot channel and the value matrix.
7. The channel prediction method of claim 6, wherein the decoder further comprises a full attention mechanism processing layer, and wherein the predicting the future time slot channel in real time by the decoder comprises:
obtaining a key matrix, a query matrix and a value matrix about the future time slot channel based on the first historical channel characteristic, the second historical channel characteristic, a linear transformation matrix corresponding to the key matrix, a linear transformation matrix corresponding to the query matrix and a linear transformation matrix corresponding to the value matrix;
obtaining hidden variables related to the future time slot channel through a key matrix, a query matrix and a value matrix of the future time slot channel based on the full attention mechanism processing layer and the normalization processing layer;
predicting the future time slot channel based on the implicit variable of the future time slot channel.
8. The channel prediction method of claim 7, wherein the deriving the key matrix, the query matrix, and the value matrix for the future time-slot channel based on the first historical channel characteristics, the second historical channel characteristics, the linear transformation matrix corresponding to the key matrix, the linear transformation matrix corresponding to the query matrix, and the linear transformation matrix corresponding to the value matrix is implemented by the following formulas:
Figure FDA0003930410360000041
Q＝W q Z (d)
Figure FDA0003930410360000042
wherein K represents a key matrix of the future time slot channel; w is a group of k A linear transformation matrix corresponding to the key matrix is represented;
Figure FDA0003930410360000043
representing the first historical channel characteristic; q represents a look-up matrix for the future time-slot channel; w q Representing a linear transformation matrix corresponding to the query matrix; z is a linear or branched member (d) Representing the second historical channel characteristics; v represents a matrix of values for the future time slot channel; w is a group of v And the value matrix corresponds to a linear transformation matrix.
9. The channel prediction method of claim 1, wherein the training the channel prediction model based on the channel training data set to obtain a trained channel prediction model comprises:
dividing the channel training data set into channel training samples and channel identification verification samples;
calculating a loss function of a normalized mean square error based on the prediction results for the channel training samples and the channel identification verification samples;
adjusting parameters of the channel prediction model based on a gradient postback algorithm and the loss function to obtain the trained channel prediction model, wherein the loss function is determined by the following formula:
Figure FDA0003930410360000051
wherein L (Θ) represents the loss function,
Figure FDA0003930410360000052
representing the channel identification verification sample,
Figure FDA0003930410360000053
representing a prediction result with respect to the channel training samples.
10. A channel prediction apparatus, wherein the channel prediction apparatus is applied to a MIMO base station, and the apparatus comprises:
a creation module for creating a channel prediction model;
the processing module is used for training the channel prediction model based on a channel training data set to obtain a trained channel prediction model, wherein the trained channel prediction model comprises an encoder and a decoder;
a prediction module, configured to predict a future time slot channel in real time based on the encoder, the decoder, and a historical time slot channel, where the historical time slot channel includes a first historical time slot channel and a second historical time slot channel, where the first historical time slot channel includes the second historical time slot channel, and the prediction module predicts the future time slot channel in real time based on the encoder, the decoder, and the historical time slot channel by:
generating beams on a plurality of strongest channel angle lattice points based on energy for a channel angle domain of the MIMO base station;
collecting a first number of the first historical time-slot channels based on the beams and feeding the first historical time-slot channels to the encoder in time-slot order, resulting in first historical channel characteristics for the first historical time-slot channels;
feeding a second number of the second historical time-slot channels to the decoder, resulting in second historical channel characteristics for the second historical time-slot channels;
inputting a first number of the first historical channel characteristics to the decoder, and predicting, by the decoder, the future time slot channel in real-time in combination with a second number of the second historical channel characteristics, wherein the first number is greater than the second number.
11. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program performs the steps of the channel prediction method according to any of claims 1 to 9.
12. A non-transitory computer readable storage medium, having stored thereon a computer program, which, when being executed by a processor, carries out the steps of the channel prediction method according to any one of claims 1 to 9.
CN202210082374.7A 2022-01-24 2022-01-24 Channel prediction method, device, electronic equipment and storage medium Active CN114422059B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210082374.7A CN114422059B (en) 2022-01-24 2022-01-24 Channel prediction method, device, electronic equipment and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210082374.7A CN114422059B (en) 2022-01-24 2022-01-24 Channel prediction method, device, electronic equipment and storage medium

Publications (2)

Publication Number Publication Date
CN114422059A CN114422059A (en) 2022-04-29
CN114422059B true CN114422059B (en) 2023-01-24

Family

ID=81277817

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210082374.7A Active CN114422059B (en) 2022-01-24 2022-01-24 Channel prediction method, device, electronic equipment and storage medium

Country Status (1)

Country Link
CN (1) CN114422059B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117479316A (en) * 2022-07-18 2024-01-30 中兴通讯股份有限公司 Channel state information determining method, electronic equipment and storage medium
CN117639862A (en) * 2022-08-10 2024-03-01 维沃移动通信有限公司 CSI prediction processing method, device, communication equipment and readable storage medium
CN117879737A (en) * 2022-09-30 2024-04-12 维沃移动通信有限公司 Channel prediction method and device and communication equipment
CN115802401B (en) * 2022-10-19 2024-02-09 苏州大学 Wireless network channel state prediction method, device, equipment and storage medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110912598A (en) * 2019-11-22 2020-03-24 中原工学院 Large-scale MIMO system CSI feedback method based on long-time attention mechanism
CN111080032A (en) * 2019-12-30 2020-04-28 成都数之联科技有限公司 Load prediction method based on Transformer structure
WO2020220278A1 (en) * 2019-04-30 2020-11-05 华为技术有限公司 Channel estimation model training method and device
CN112152948A (en) * 2019-06-28 2020-12-29 华为技术有限公司 Wireless communication processing method and device

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110621026B (en) * 2019-02-18 2023-09-05 北京航空航天大学 Multi-moment prediction method for base station flow
US11799570B2 (en) * 2020-06-30 2023-10-24 Qualcomm Incorporated Techniques for cross-band channel prediction and reporting

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020220278A1 (en) * 2019-04-30 2020-11-05 华为技术有限公司 Channel estimation model training method and device
CN112152948A (en) * 2019-06-28 2020-12-29 华为技术有限公司 Wireless communication processing method and device
CN110912598A (en) * 2019-11-22 2020-03-24 中原工学院 Large-scale MIMO system CSI feedback method based on long-time attention mechanism
CN111080032A (en) * 2019-12-30 2020-04-28 成都数之联科技有限公司 Load prediction method based on Transformer structure

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Realistic Channel Models Pre-training;Yourui Huangfu等;《2019 IEEE Globecom Workshops (GC Wkshps)》;20200305;第1-6页 *
注意力机制CNN的毫米波大规模MIMO系统信道估计算法;刘紫燕等;《系统工程与电子技术》;20211011;第307-312页 *

Also Published As

Publication number Publication date
CN114422059A (en) 2022-04-29

Similar Documents

Publication Publication Date Title
CN114422059B (en) Channel prediction method, device, electronic equipment and storage medium
CN111901862B (en) User clustering and power distribution method, device and medium based on deep Q network
CN113411110B (en) Millimeter wave communication beam training method based on deep reinforcement learning
Yang et al. Machine learning enabling analog beam selection for concurrent transmissions in millimeter-wave V2V communications
CN107171708B (en) Channel tracking and hybrid precoding method for large-scale MIMO system
CN112260737B (en) Multi-beam satellite communication robust precoding method with total energy efficiency and minimum energy efficiency balanced
CN113965233B (en) Multi-user broadband millimeter wave communication resource allocation method based on deep learning
Elbir et al. Federated learning for physical layer design
KR102510513B1 (en) Deep learning based beamforming method and apparatus for the same
CN110611526A (en) Millimeter wave hybrid analog/digital beam forming method based on improved Riemann manifold optimization
CN113381790A (en) AI-based environment knowledge assisted wireless channel feedback method
Xia et al. Deep learning based beamforming neural networks in downlink MISO systems
CN113438002A (en) LSTM-based analog beam switching method, device, equipment and medium
CN113011448B (en) Multi-user downlink channel estimation method and device based on transfer learning
CN111010222B (en) Deep learning-based large-scale MIMO downlink user scheduling method
He et al. Joint user scheduling and beamforming design for multiuser MISO downlink systems
Li et al. A GAN-LSTM based AI framework for 6G wireless channel prediction
Yang et al. Transmit antenna selection for millimeter-wave MIMO system based on GBDT
CN114567358A (en) Large-scale MIMO robust WMMSE precoder and deep learning design method thereof
Jiang et al. Multitask deep learning-based multiuser hybrid beamforming for mm-wave orthogonal frequency division multiple access systems
CN117240331A (en) No-cellular network downlink precoding design method based on graph neural network
Wang et al. Joint beamforming and power allocation using deep learning for D2D communication in heterogeneous networks
CN114826833A (en) Communication optimization method and terminal for CF-mMIMO in IRS-assisted MEC
Schynol et al. Deep unfolding in multicell mu-mimo
Wang et al. New Environment Adaptation with Few Shots for OFDM Receiver and mmWave Beamforming

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