Active user detection and channel estimation method and device, using deep neural network
View Patent ↗Provided is an active user detection and channel estimation method of a base station in a wireless communication system based on grant-free uplink transmission. The method, according to one embodiment, comprises the steps of: receiving superimposed signals ( ) from k active terminals; by using a first artificial neural network and using the received signals ( ) as input, calculating, for all terminals in a cell of a base station, an estimated probability ({circumflex over (Ω)}) of each of the terminals being an active terminal; and estimating channels of the active terminals by using a second artificial neural network and using the received signals ( ) and an active user detection result value as input.
1. An active terminal detection and channel estimation method of a base station in a wireless communication system based on grant-free uplink transmission, the method comprising:
receiving superimposed signals ( ) from k active terminals;
calculating, using a first artificial neural network and using the received signals ( ) as input, an estimated probability ({circumflex over (Ω)}) that each of all terminals in a cell of the base station is an active terminal; and
estimating channels of the active terminals using a second artificial neural network with the received signals ( ) and an active terminal detection result value as input,
wherein the first artificial neural network and the second artificial neural network are each an artificial neural network based on long short-term memory (LSTM) networks, and
wherein the first artificial neural network is established by learning a direct mapping according to Equation 5, and the second artificial neural network is established by learning a direct mapping according to Equation 8:
{circumflex over (Ω)}= g and ( :θ A ) [Equation 5]
wherein θ A is a parameter of the first artificial neural network used for the active user detection,
ĥ {circumflex over (Ω)} =g ce ( ,{circumflex over (Ω)};θ C ) [Equation 8]
wherein θ C is a parameter of the second artificial neural network used for the channel estimations.
2. The active terminal detection and channel estimation method of claim 1 , wherein the received signals ( ) input to the first artificial neural network are changed into a hidden layer representation ( ) by passing through a fully-connected layer, the changed hidden layer representation (yaud) is calculated into a hidden layer representation (o j ) of the active user detection result by sequentially computing a plurality of LSTM cells, and the hidden layer representation (o j ) of the active user detection result is converted into the estimated probability ({circumflex over (Ω)}) by means of a sigmoid function.
3. The active terminal detection and channel estimation method of claim 2 , wherein each of the plurality of LSTM cells performs a computation according to Equation 6:
=σ g ( W f +U f o j-1 +b f )
=σ g ( W i +U i o j-1 +b i )
=σ g ( W o +U o o j-1 +b o )
=σ g ( W c +U c o j-1 +b c )
= º + º
o j = º tanh( ) [Equation 6]
wherein W f , W i , W o , W c , U f , U i , U o , U c , and b f , b i , b o , b c are weights and deviations of the hidden layers, respectively, σg is a sigmoid function, and tanh is a hyperbolic tangent function.
4. The active terminal detection and channel estimation method of claim 1 , wherein the second artificial neural network is trained to activate a channel characteristic of a specific terminal as an input gate and to deactivate channel characteristics of remaining terminals as a forget gate.
5. A base station device for an active user detection and channel estimation in a wireless communication system based on grant-free uplink transmission, the base station device comprising:
a receiver configured to receive superimposed signals ( ) from k active terminals; and
one or more processors configured to control an operation of the receiver,
wherein the one or more processors:
calculates, using a first artificial neural network and using the received signals ( ) as input, an estimated probability ({circumflex over (Ω)}) that each of all terminals in a cell of the base station is an active terminal; and
estimates channels of the active terminals using a second artificial neural network with the received signal ( ) and an active user detection result value as input,
wherein the first artificial neural network and the second artificial neural network are each an artificial neural network based on long short-term memory (LSTM) networks, and
wherein the first artificial neural network is established by learning a direct mapping according to Equation 5, and the second artificial neural network is established by learning a direct mapping according to Equation 8:
{circumflex over (Ω)}= g aud ( :θ A ) [Equation 5]
wherein θ A is a parameter of the first artificial neural network used for the active user detection,
ĥ {circumflex over (Ω)} =g ce ( ,{circumflex over (Ω)};θ C ) [Equation 8]
wherein θ C is a parameter of the second artificial neural network used for the channel estimations.
6. The base station device of claim 5 , wherein the received signals ( ) input to the first artificial neural network are changed into a hidden layer representation ( ) by passing through a fully-connected layer, the changed hidden layer representation ( ) is calculated into a hidden layer representation (o j ) of the active user detection result by sequentially computing a plurality of LSTM cells, and the hidden layer representation (o j ) of the active user detection result is converted into the estimated probability ({circumflex over (Ω)}) by means of a sigmoid function.
7. The base station device of claim 6 , wherein each of the plurality of LSTM cells performs a computation according to Equation 6:
=σ g ( W f +U f o j-1 +b f )
=σ g ( W i +U i o j-1 +b i )
=σ g ( W o +U o o j-1 +b o )
=σ g ( W c +U c o j-1 +b c )
= º + º
o j = º tanh( ) [Equation 6]
wherein W f , W i , W o , W c , U f , U i , U o , U c , and b f , b i , b o , b c are weights and deviations of the hidden layers, respectively, σ g is a sigmoid function, and tanh is a hyperbolic tangent function.
8. The base station device of claim 5 , wherein the second artificial neural network is trained to activate a channel characteristic of a specific terminal as an input gate and to deactivate channel characteristics of remaining terminals as a forget gate.