CN114598575B - Deep learning channel estimation method based on self-attention mechanism - Google Patents
Deep learning channel estimation method based on self-attention mechanism Download PDFInfo
- Publication number
- CN114598575B CN114598575B CN202210239196.4A CN202210239196A CN114598575B CN 114598575 B CN114598575 B CN 114598575B CN 202210239196 A CN202210239196 A CN 202210239196A CN 114598575 B CN114598575 B CN 114598575B
- Authority
- CN
- China
- Prior art keywords
- matrix
- channel
- self
- attention
- feature map
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 74
- 230000007246 mechanism Effects 0.000 title claims abstract description 43
- 238000013135 deep learning Methods 0.000 title claims abstract description 31
- 230000004044 response Effects 0.000 claims abstract description 45
- 230000001419 dependent effect Effects 0.000 claims abstract description 5
- 238000013507 mapping Methods 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 81
- 230000006870 function Effects 0.000 claims description 11
- 238000005070 sampling Methods 0.000 claims description 9
- 238000013527 convolutional neural network Methods 0.000 claims description 8
- 238000013528 artificial neural network Methods 0.000 claims description 6
- 239000000654 additive Substances 0.000 claims description 4
- 230000000996 additive effect Effects 0.000 claims description 4
- 238000005457 optimization Methods 0.000 claims description 2
- 230000002194 synthesizing effect Effects 0.000 claims description 2
- 230000006872 improvement Effects 0.000 abstract description 2
- 238000004891 communication Methods 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 4
- 238000005562 fading Methods 0.000 description 3
- 230000001965 increasing effect Effects 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 125000004122 cyclic group Chemical group 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000003595 spectral effect Effects 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 1
- 239000000969 carrier Substances 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000010363 phase shift Effects 0.000 description 1
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L25/00—Baseband systems
- H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
- H04L25/0202—Channel estimation
- H04L25/024—Channel estimation channel estimation algorithms
- H04L25/0242—Channel estimation channel estimation algorithms using matrix methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L25/00—Baseband systems
- H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
- H04L25/0202—Channel estimation
- H04L25/024—Channel estimation channel estimation algorithms
- H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Networks & Wireless Communication (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Signal Processing (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Power Engineering (AREA)
- Radio Transmission System (AREA)
Abstract
The invention discloses a deep learning channel estimation method based on a self-attention mechanism under a multi-input multi-output orthogonal frequency division multiplexing system, which is characterized in that the correlation between time-frequency domain channel impulse responses is utilized, channel characteristic information can be effectively extracted by adopting a self-attention mechanism module, so that a global dependent characteristic mapping is constructed, and the correlation between the time-frequency domain channel impulse responses can be deeply learned by utilizing a deep learning method. Compared with the channel estimation result of the existing MIMO OFDM system, the channel estimation result has ideal improvement in accuracy.
Description
Technical Field
The invention belongs to the technical field of wireless communication, and particularly relates to a deep learning channel estimation method based on a self-attention mechanism, which is applicable to a multi-input multi-output orthogonal frequency division multiplexing system.
Background
Emerging telecommunications technology books ("Orthogonal frequency division multiplexing with subcarrier power modulation for doubling the spectral efficiency of, G and beyond networks." in Transactions on Emerging Telecommunications Technologies,2020,31 (4): e 3921.) state that orthogonal frequency division multiplexing technology is widely used for 5G wireless communications due to its robustness to inter-carrier and inter-symbol interference, and will continue to be used in 6G communications as a key technology. Multiple-input multiple-output systems are capable of enhancing channel capacity without increasing bandwidth, thereby accommodating a greater number of users within the available bandwidth, and are therefore often used in conjunction with orthogonal frequency division multiplexing techniques to achieve better data transmission rates, higher spectral utilization, and better resistance to multipath fading. The performance of the mimo ofdm system in the wireless fading environment depends greatly on the accurate channel impulse response, and it is very important to obtain the accurate channel impulse response through channel estimation. The international society of electronic and electrical engineers communication express ("Deep learning-based channel estimation." in IEEE Communications Letters,2019,23 (4): 652-655.) proposes to treat the time-frequency response of a fast fading communication channel as a two-dimensional image, and to estimate the channel impulse response using a channel estimation method based on a Deep learning algorithm. But the accuracy of the channel impulse response obtained by the channel estimation method is not high because the correlation between the time-frequency domain channel impulse responses is not effectively utilized.
Disclosure of Invention
The invention provides a channel estimation method based on self-attention mechanism for deep learning in a multi-input multi-output orthogonal frequency division multiplexing system, so as to effectively utilize the correlation between time-frequency domain channel impulse responses and further improve the accuracy of estimation results.
The invention relates to a channel estimation method based on self-attention mechanism for deep learning in a multi-input multi-output orthogonal frequency division multiplexing system, which is characterized in that:
n for one transmitting antenna number t The number of the receiving antennas is N r Multiple input multiple output orthogonal frequency division multiplexing system of (2), wherein, at the transmitting end, the n-th t A certain orthogonal frequency division multiplexing symbol on the root transmit antenna is denoted as:wherein L is the number of subcarriers, [] T Represents the transpose, for the nth r The symbols on the kth subcarrier above the root receive antenna are denoted as:
wherein,represents the nth t Root transmit antenna and nth r Channel impulse response between root receive antennas corresponding to kth subcarrier +.>Representing additive white gaussian noise on the receiving antenna; therein the pair ofOne pilot symbol, all N r The symbols on the kth subcarrier received by the root receive antenna are represented as:
Y k ＝H k X k +Z k ,(k＝1,...,L)
wherein the method comprises the steps of Respectively representing a receiving matrix of a receiving end, a transmitting matrix of a transmitting end, a noise matrix,representing the frequency domain channel impulse response matrix on the kth subcarrier, H k Expressed as:
for H k Channel estimation is carried out to obtain the channel impulse response estimation result as followsUsing a channel estimation method based on a least squares criterion, a cost function shown by the following formula is minimized:
letting the cost function of the above relation toThe partial derivative of (2) is 0, i.e
Obtaining channel impulse response estimation results under a channel estimation method based on a least square criterion,
wherein ( -1 An inverse matrix representing the matrix;
the specific process steps of the channel estimation method based on the deep learning of the self-attention mechanism are as follows:
step S1: channel state information at pilot symbols and estimated channel matrix by least squares channel estimation algorithm
S2, step: for the channel matrixPerforming up-sampling linear interpolation to obtain frequency domain channel response at the data symbol, combining with frequency domain channel response at pilot frequency to obtain complete frequency domain channel response matrix, and remolding the complete frequency domain channel response matrix into L×N sym ×(N t ×N r X 2), wherein L represents the number of subcarriers, N sym Indicating the number of sub-carriers, N t Indicating the number of transmitting antennas, N r Indicating the number of receiving antennas;
step S3: extracting features of the low-resolution matrix through two convolutional neural networks to obtain a feature map about the low-resolution matrix; obtaining a weighted self-attention mechanism feature map through a self-attention mechanism module according to the feature map, multiplying the self-attention mechanism feature map by a learning coefficient, and adding the learning coefficient to the original feature map to obtain a relationship feature of mutual dependence of any two positions in the feature map; the relation features are respectively passed through a convolution network and a self-attention mechanism module twice, so that a relation dependence feature with more comprehensive learning is obtained;
s4, step: the relation dependence characteristic is subjected to a convolution network to output a matrix consistent with the dimension of a channel matrix which is finally required to be estimated
The invention relates to a deep learning channel estimation method based on a self-attention mechanism in a multi-input multi-output orthogonal frequency division multiplexing system; the channel estimation method effectively utilizes the correlation between the time-frequency domain channel impulse responses, which is not considered in the prior channel estimation work; the invention provides a deep learning channel estimation method based on a self-attention mechanism, which can effectively construct the interdependence relation of any two positions by outputting channel characteristics to a super-resolution neural network of the self-attention mechanism; because of a certain correlation among elements of the channel matrix generated by the adopted channel model, the self-attention mechanism module is adopted to more effectively utilize the correlation among time-frequency domain channel impulse responses, and compared with the channel estimation result of the existing multi-input multi-output orthogonal frequency division multiplexing system, the channel estimation result has ideal improvement in accuracy.
Description of the drawings:
FIG. 1 is a flowchart of a neural network implementation of a deep learning channel estimation method based on a self-attention mechanism according to the present invention;
FIG. 2 is a schematic diagram of a self-attention mechanism module according to the present invention.
FIG. 3 is a graph comparing the Mean Square Error (MSE) performance of the method of the present invention with the channel estimation of an existing MIMO OFDM system at different signal-to-noise ratio (SNR) settings;
fig. 4 is a graph comparing the Bit Error Rate (BER) performance of the method of the present invention with the channel estimation of the existing mimo-ofdm system at different signal-to-noise ratio (SNR) settings.
Detailed Description
The deep learning channel estimation method based on the self-attention mechanism in the mimo-ofdm system of the present invention is described in further detail and specifically illustrated by the following embodiments with reference to the accompanying drawings.
Embodiment case 1:
to facilitate an understanding of the implementation of the method, a detailed description will be given below of how the present invention makes use of deep learning based on self-attention mechanismsThe method performs channel estimation. N for one transmitting antenna number t The number of the receiving antennas is N r Multiple input multiple output orthogonal frequency division multiplexing system of (2), wherein, at the transmitting end, the n-th t A certain orthogonal frequency division multiplexing symbol on the root transmit antenna is denoted as:wherein L is the number of subcarriers, [] T Represents the transpose, for the nth r The symbols on the kth subcarrier above the root receive antenna are denoted as:
wherein,represents the nth t Root transmit antenna and nth r Channel impulse response between root receive antennas corresponding to kth subcarrier +.>Representing additive white gaussian noise on the receiving antenna; for one of the pilot symbols, all N r The symbols on the kth subcarrier received by the root receive antenna are represented as:
Y k ＝H k X k +Z k ,(k＝1,...,L)
wherein the method comprises the steps of Respectively representing a receiving matrix of a receiving end, a transmitting matrix of a transmitting end, a noise matrix,representing the frequency domain channel impulse response matrix on the kth subcarrier, H k Expressed as:
for H k Channel estimation is carried out to obtain the channel impulse response estimation result as followsUsing a channel estimation method based on a least squares criterion, a cost function shown by the following formula is minimized:
letting the cost function of the above relation toThe partial derivative of (2) is 0, i.e
Obtaining channel impulse response estimation results under a channel estimation method based on a least square criterion,
wherein ( -1 An inverse matrix representing the matrix;
the specific process steps of the channel estimation method based on the deep learning of the self-attention mechanism are as follows:
step S1: channel state information at pilot symbols and estimated channel matrix by least squares channel estimation algorithm
S2, step: for the channel matrixPerforming up-sampling linear interpolation to obtain frequency domain channel response at the data symbol, and combining with frequency domain channel response at pilot frequency to obtain complete signalA frequency domain channel response matrix, and remodelling the complete frequency domain channel response matrix into L multiplied by N sym ×(N t ×N r X 2), where L is the number of subcarriers, N sym N is the number of subcarriers t For transmitting the antenna number N r The number of the receiving antennas;
step S3: extracting features of the low-resolution matrix through two convolutional neural networks to obtain a feature map about the low-resolution matrix; obtaining a weighted self-attention mechanism feature map through a self-attention mechanism module according to the feature map, multiplying the self-attention mechanism feature map by a learning coefficient, and adding the learning coefficient to the original feature map to obtain a relationship feature of mutual dependence of any two positions in the feature map; the relation features are respectively passed through a convolution network and a self-attention mechanism module twice, so that a relation dependence feature with more comprehensive learning is obtained;
s4, step: the relation dependence characteristic is subjected to a convolution network to output a matrix consistent with the dimension of a channel matrix which is finally required to be estimated
Fig. 1 shows a flowchart of a neural network implementation of the deep learning channel estimation method based on the self-attention mechanism.
In this embodiment, step S1: channel state information at pilot symbols and estimated channel matrix by least squares channel estimation algorithmThe method specifically comprises the following steps:
by transmitting pilot symbols X with known pilot positions k Using the received signal Y k Using least square methodEstimating channel state information of the symbol at all pilot frequencies to obtain an estimated channel matrixI.e. the channel matrix A0 in fig. 1.
In one embodiment, step S2: for the channel matrixPerforming up-sampling linear interpolation to obtain frequency domain channel response at the data symbol, combining with frequency domain channel response at the pilot frequency to obtain a complete frequency domain channel response matrix, and remolding the complete channel frequency response matrix into L×N sym ×(N t ×N r X 2), where L is the number of subcarriers, N sym N is the number of subcarriers t For transmitting the antenna number N r The method specifically comprises the following steps of:
for a pair ofThe up-sampling linear interpolation, namely the up-sampling operation A1 in the first step of fig. 1, is performed, the method replaces up-sampling by a linear interpolation method, and assuming that the interval between two pilots is L, the frequency domain channel response estimated value at the middle position of the mth pilot and the (m+1) th pilot can be expressed as:
after the frequency domain channel response at the data symbol is obtained, in order to facilitate the convolutional network operation, the real part and the imaginary part of the matrix are respectively used as the channel dimensions of the channel matrix, so that the matrix is remodeled into L multiplied by N sym ×(N t ×N r X 2) is provided.
In one embodiment, step S3: extracting features of the low-resolution matrix through two convolutional neural networks to obtain a feature map about the low-resolution matrix; obtaining a weighted self-attention mechanism feature map through a self-attention mechanism module according to the feature map, multiplying the self-attention mechanism feature map by a learning coefficient, and adding the learning coefficient to the original feature map to obtain a relationship feature of mutual dependence of any two positions in the feature map; the relation features are respectively passed through a convolution network and a self-attention mechanism module twice to obtain a relation dependency feature with more comprehensive learning, which comprises the following steps:
the low-resolution matrix is first passed through two convolutional neural networks A2 and A3 as shown in fig. 1, to obtain a feature map x:
wherein Inter (·) represents the interpolation function, i.e. the upsampling module in the network, W 1 ，W 2 The weights of the convolutional neural network are represented, respectively, and the resulting feature map is then input into a self-attention feature module (i.e., A4 shown in fig. 1). Fig. 2 is a schematic diagram of a self-attention mechanism module structure used in the present invention, wherein a feature map is B0 shown in fig. 2, and then three feature matrices query, key and value are obtained by respectively obtaining 3 feature mapping functions f, g and h through 3 convolution networks with convolution kernel sizes of 1×1:
query(x)＝f(x)＝W f x+b f ，
key(x)＝g(x)＝W g x+b g ，
value(x)＝h(x)＝W h x+b h ，
where W, b denote the weight and bias of the convolutional network, respectively. The three feature matrices query, key, and value described above correspond to B1, B2, and B3, respectively, shown in fig. 2. Then, the query matrix, the key matrix and the value matrix are respectively recombined into the size ofMatrix Query, key, value, B4, B5, B6 shown in FIG. 2, wherein +.>N=H×W, then performing matrix multiplication operation on the Query and Key, and obtaining a block with the size of +.>Inter-dependent Attention mapping matrix attention_map, i.e. B7 shown in fig. 2:
Attention_map j,i representing the degree of consideration dependence on the ith region in synthesizing the jth region, i.e. implementing the interdependence at any two positions in its features,
because the obtained attention_map is a weight matrix with a sum of 1 according to the row, the attention_map is transposed, then multiplied by the Value matrix to obtain weighted sum at each position, finally a convolution network with a convolution kernel size of 1×1 is used to obtain a self-Attention feature map o (i.e. B8 shown in fig. 2), then the self-Attention feature map is multiplied by a coefficient to be added with the original feature map, and the final output is:
y＝γO+x,
where y is B9 shown in fig. 2, γ is a learnable scalar initialized to 0, and since it can be learned more easily in locally adjacent network features, the introduced γ learns from local and then gradually learns to give higher weight to non-local features, which is also consistent with the intuition of the person: the simple task is learned firstly, then the complexity of the task is gradually increased, and the weighted self-attention feature map is continuously overlapped on the original feature map along with the continuous deep learning, so that a globally dependent feature map is finally obtained.
In one embodiment, step S4 above: the relation dependence characteristic is subjected to a convolution network to output a matrix consistent with the dimension of a channel matrix which is finally required to be estimatedThe method specifically comprises the following steps:
by setting the number of convolution kernels, the channel dimension of the final output is converted to be consistent with the estimated channel matrix dimension through convolution (namely A5 shown in fig. 1), and finally, the feedback of the neural network is performed by solving the following optimization problem, so that a better estimation result is obtained:
where Θ represents the training parameters of the whole network.
After the above steps, the recovered estimated channel (i.e., A6 shown in fig. 1) is finally outputted.
The invention relates to a deep learning channel estimation method based on a self-attention mechanism in a multi-input multi-output orthogonal frequency division multiplexing system; the channel estimation method effectively utilizes the correlation between the time-frequency responses of the channels, which is not considered in the prior channel estimation work; the invention provides a deep learning channel estimation method based on a self-attention mechanism, which can effectively construct the interdependence relation of any two positions by outputting channel characteristics to a super-resolution neural network of the self-attention mechanism; because of a certain correlation among elements of the channel matrix generated by the adopted channel model, the self-attention mechanism module is adopted to more effectively utilize the correlation among time-frequency responses of the channels, and compared with the channel estimation result of the existing multi-input multi-output orthogonal frequency division multiplexing system, the accuracy of the channel estimation result is improved.
The simulation is used for comparing the deep learning channel estimation method based on the self-attention mechanism in the MIMO OFDM system of the invention with the existing channel estimation method in the system. The indicators against which the accuracy of the channel estimation is measured by the present invention are Mean Square Error (MSE) and Bit Error Rate (BER).
The simulation of the deep learning channel estimation method based on the self-attention mechanism under the multi-input multi-output orthogonal frequency division multiplexing system in the embodiment is specifically set as follows:
for simulation of different signal-to-noise ratios, the number of transmitting antennas is 2, the number of receiving antennas is 2, the number of subcarriers is 64, a protection interval is a cyclic prefix, the cyclic prefix is 1/4, the modulation mode is quadrature phase shift keying, noise is additive Gaussian white noise, the transmitting power is normalized, and the signal-to-noise ratio is expressed in the form of a logarithmic function.
Fig. 3 shows the mean square error comparison result of the method of the present invention with the existing estimation method under different signal-to-noise ratios, wherein the solid line of the uppermost C3 mark represents that the existing estimation method is based on the least squares method, the solid line of the C2 mark represents that the existing estimation method is based on the super-resolution deep learning method, and the solid line of the lowermost C1 mark represents the method of the present invention. As can be seen from fig. 3, the mean square error of the mimo ofdm system employing the method of the present invention is smaller than that employing the least square method and the super-resolution deep learning method. At high signal-to-noise ratio, the method has a gain of about 1.1dB relative to the super-resolution deep learning method, and at low signal-to-noise ratio, the method has a gain of about 0.8 dB.
Fig. 4 shows the bit error rate comparison result of the method of the present invention with the existing estimation method under different signal-to-noise ratios, wherein the solid line of the uppermost D3 mark represents that the existing estimation method is based on the least squares method, the solid line of the middle D2 mark represents that the existing estimation method is based on the super-resolution deep learning method, and the solid line of the lowermost D1 mark represents the method of the present invention. As can be seen from fig. 4, the bit error rate of the mimo ofdm system employing the method of the present invention is smaller than that employing the least square method and the super-resolution deep learning method. And with the increase of the signal-to-noise ratio, the performance gap between the performance of the method and the performance of the existing estimation method is gradually increased.
By the above embodiment, it is proved that the deep learning channel estimation based on the self-attention mechanism in the mimo orthogonal frequency division multiplexing system has more accurate channel estimation result and ideal performance in terms of mean square error or bit error rate performance because of effectively utilizing the correlation between channels compared with the existing channel estimation method.
Claims (1)
1. A deep learning channel estimation method based on a self-attention mechanism is characterized in that:
n for one transmitting antenna number t The number of the receiving antennas is N r Multiple input multiple output orthogonal frequency division multiplexing system of (2), wherein, at the transmitting end, the n-th t A certain orthogonal frequency division multiplexing symbol on the root transmit antenna is denoted as:wherein L is the number of subcarriers, [] T Represents the transpose, for the nth r The symbols on the kth subcarrier above the root receive antenna are denoted as:
wherein,represents the nth t Root transmit antenna and nth r Channel impulse response between root receive antennas corresponding to kth subcarrier +.>Representing additive white gaussian noise on the receiving antenna; for one of the pilot symbols, all N r The symbols on the kth subcarrier received by the root receive antenna are represented as:
Y k ＝H k X k +Z k ,(k＝1,...,L)
wherein, respectively representing a receiving matrix of a receiving end, a transmitting matrix of a transmitting end, a noise matrix,representing the frequency domain channel impulse response matrix on the kth subcarrier, H k Expressed as:
for H k Channel estimation is carried out to obtain the channel impulse response estimation result as followsUsing a channel estimation method based on a least squares criterion, a cost function shown by the following formula is minimized:
letting the cost function of the above relation toThe partial derivative of (2) is 0, i.e
Obtaining channel impulse response estimation results under a channel estimation method based on a least square criterion,
wherein ( -1 An inverse matrix representing the matrix;
the specific process steps of the channel estimation method based on the deep learning of the self-attention mechanism are as follows:
step S1: channel state information at pilot symbols and estimated channel matrix by least squares channel estimation algorithm
S2, step: for the channel matrixPerforming up-sampling linear interpolation to obtain frequency domain channel response at the data symbol, combining with frequency domain channel response at the pilot frequency to obtain a complete frequency domain channel response matrix, and remolding the complete frequency domain channel response matrix into L×N sym ×(N t ×N r X 2), where L is the number of subcarriers, N sym N is the number of subcarriers t For transmitting the antenna number N r For the number of receive antennas, including;
for a pair ofUp-sampling linear interpolation is performed, the up-sampling is replaced by a linear interpolation method, and assuming that the interval between two pilots is L, the frequency domain channel response estimated value at the middle position of the mth pilot and the (m+1) th pilot can be expressed as:
after the frequency domain channel response of the data symbol is obtained, the real part and the imaginary part of the matrix are respectively used as the channel dimension of the channel matrix, and the matrix is remodeled into L multiplied by N sym ×(N t ×N r X 2) a low resolution matrix;
step S3: extracting features of the low-resolution matrix through two convolutional neural networks to obtain a feature map about the low-resolution matrix; obtaining a weighted self-attention mechanism feature map through a self-attention mechanism module according to the feature map, multiplying the self-attention mechanism feature map by a learning coefficient, and adding the learning coefficient to the original feature map to obtain a relationship feature of mutual dependence of any two positions in the feature map; the relation features are respectively passed through a convolution network and a self-attention mechanism module twice to obtain a relation dependency feature with more comprehensive learning, comprising:
the low-resolution matrix firstly passes through two convolutional neural networks to obtain a feature map x:
wherein Inter (·) represents the interpolation function, i.e. the upsampling module in the network, W 1 ，W 2 Respectively representing the weights of the convolutional neural network, inputting the obtained feature map into a self-attention feature module, and obtaining three feature matrices query, key and value through 3 convolutional networks with the convolution kernel size of 1×1 and 3 feature mapping functions f, g and h respectively:
query(x)＝f(x)＝W f x+b f ，
key(x)＝g(x)＝W g x+b g ，
value(x)＝h(x)＝W h x+b h ，
wherein W, b represent the weight and bias of the convolutional network respectively; then, the query matrix, the key matrix and the value matrix are respectively recombined into the size ofMatrix Query, key, value of (2), whereinN=h×w, then performing matrix multiplication on Query and Key, and obtaining a value of +_n via a softmax module>Is defined by the inter-dependent Attention mapping matrix attention_map:
Attention_map j,i representing the degree of consideration dependence on the ith region in synthesizing the jth region, i.e. implementing its featuresInter-dependency at any two positions,
the attention_map is a weight matrix with the sum of 1 according to the row addition, the attention_map is transposed and multiplied by the Value matrix to obtain the weighted sum of the attention_map at each position, and finally a convolution network with the convolution kernel size of 1×1 is used to obtain a self-Attention feature map, then the self-Attention feature map is multiplied by a coefficient to be added with the original feature map, and the final output y is:
y＝γO+x,
y is B9, gamma is a learnable scalar initialized to 0, the gamma is learned from a part, then the non-local feature is given higher weight by gradual learning, and the weighted self-attention feature map is continuously overlapped on the original feature map along with the continuous deep learning, so that a globally dependent feature map is finally obtained;
s4, step: the relation dependence characteristic is subjected to a convolution network to output a matrix consistent with the dimension of a channel matrix which is finally required to be estimatedComprising the following steps:
the number of convolution kernels is set, the channel dimension finally output is converted to be consistent with the estimated channel matrix dimension through convolution, and finally feedback of the neural network is carried out through solving the following optimization problem, so that a better estimation result is obtained:
where Θ represents the training parameters of the whole network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210239196.4A CN114598575B (en) | 2022-03-11 | 2022-03-11 | Deep learning channel estimation method based on self-attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210239196.4A CN114598575B (en) | 2022-03-11 | 2022-03-11 | Deep learning channel estimation method based on self-attention mechanism |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114598575A CN114598575A (en) | 2022-06-07 |
| CN114598575B true CN114598575B (en) | 2024-01-09 |
Family
ID=81817558
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210239196.4A Active CN114598575B (en) | 2022-03-11 | 2022-03-11 | Deep learning channel estimation method based on self-attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114598575B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115664898B (en) * | 2022-10-24 | 2023-09-08 | 四川农业大学 | OFDM system channel estimation method and system based on complex convolution neural network |
| CN118282811A (en) * | 2024-06-04 | 2024-07-02 | 中科南京移动通信与计算创新研究院 | MIMO channel estimation method based on AI |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113160375A (en) * | 2021-05-26 | 2021-07-23 | 郑健青 | Three-dimensional reconstruction and camera pose estimation method based on multi-task learning algorithm |
| US11115095B1 (en) * | 2020-03-05 | 2021-09-07 | Cornell University | Finite-alphabet beamforming for multi-antenna wideband systems |
| CN114066052A (en) * | 2021-11-15 | 2022-02-18 | 国网江苏省电力有限公司营销服务中心 | Adjustable load power multi-step prediction method based on improved TCN correction accumulated error |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8588321B2 (en) * | 2007-12-25 | 2013-11-19 | Panasonic Corporation | Wireless communication apparatus, wireless communication system and wireless communication method |
| US11405240B2 (en) * | 2020-08-11 | 2022-08-02 | University Of Manitoba | Method for massive unsourced random access |
-
2022
- 2022-03-11 CN CN202210239196.4A patent/CN114598575B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11115095B1 (en) * | 2020-03-05 | 2021-09-07 | Cornell University | Finite-alphabet beamforming for multi-antenna wideband systems |
| CN113160375A (en) * | 2021-05-26 | 2021-07-23 | 郑健青 | Three-dimensional reconstruction and camera pose estimation method based on multi-task learning algorithm |
| CN114066052A (en) * | 2021-11-15 | 2022-02-18 | 国网江苏省电力有限公司营销服务中心 | Adjustable load power multi-step prediction method based on improved TCN correction accumulated error |
Non-Patent Citations (1)
| Title |
|---|
| Deep Learning-Based channel estimation with SRGAN in OFDM Systems;ZhaoSiqiang等;《2021IEEE Wireless Communication and Networking Conference(WCNC)》;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114598575A (en) | 2022-06-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wei et al. | Transmitter and receiver window designs for orthogonal time-frequency space modulation | |
| Savaux et al. | LMMSE channel estimation in OFDM context: a review | |
| CN114598575B (en) | Deep learning channel estimation method based on self-attention mechanism | |
| Liu et al. | Transmit antennae space-time block coding for generalized OFDM in the presence of unknown multipath | |
| TWI463823B (en) | Model-based channel estimator for correlated fading channels and channel estimation method thereof | |
| CN1878159B (en) | A method for transmitting symbols through at least a channel in a telecommunication system | |
| EP1757000B1 (en) | Frequency domain equalization of frequency-selective mimo channels | |
| RU2481712C2 (en) | Mobile communication system, receiving device and signal transmission method | |
| Al-Naffouri et al. | A model reduction approach for OFDM channel estimation under high mobility conditions | |
| CN108881076B (en) | MIMO-FBMC/OQAM system channel estimation method based on compressed sensing | |
| CN106506415A (en) | A kind of method of multiuser MIMO ofdm system channel estimation | |
| WO2008032849A1 (en) | Wireless communication apparatus | |
| Ru et al. | Model-driven channel estimation for OFDM systems based on image super-resolution network | |
| Ponnaluru et al. | RETRACTED ARTICLE: Deep learning for estimating the channel in orthogonal frequency division multiplexing systems | |
| Hussein et al. | Least Square Estimation‐Based Different Fast Fading Channel Models in MIMO‐OFDM Systems | |
| Ramadan et al. | Non-linear equalisation and CFO compensation for MIMO-OFDM communication systems based on DWT | |
| Bhoyar et al. | Leaky least mean square (LLMS) algorithm for channel estimation in BPSK-QPSK-PSK MIMO-OFDM system | |
| Zamiri-Jafarian et al. | Adaptive channel SVD estimation for MIMO-OFDM systems | |
| CN101719816A (en) | Method for realizing low feedback velocity of self-adaptive MIMO-SCFDE system | |
| CN111030741B (en) | Precoding algorithm for interference suppression of multi-carrier system under fast time-varying scene | |
| Wang et al. | Model-based channel estimation framework for MIMO multicarrier communication systems | |
| CN101969417A (en) | Low-return self-adaptive multimode transmission method of MIMO-SCFDE (Multiple Input Multiple Output Single Carrier Frequency Domain Equilibrium) system | |
| Mei et al. | LMMSE channel estimation for OFDM systems with channel correlation function selection | |
| Gao et al. | An Attention-Aided Deep Neural Network Design for Channel Estimation in Massive MIMO Systems | |
| CN102437977B (en) | The efficient channel estimation method of a kind of low complex degree |
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 |