KR101992053B1 - SISO-OFDM channel estimation apparatus using deep neural network based on adaptive ensemble supervised learning and method thereof - Google Patents

Abstract

The present invention relates to an SISO-OFDM channel estimation device using a deep neural network based on adaptive ensemble supervised learning, and a method thereof. According to the present invention, provided is an SISO-OFDM channel estimation method which comprises: a step of, in responding to training symbols modulated and sent from a transmission end, separately inputting the training symbols respectively received through a plurality of path channels to a plurality of independent deep neural networks to make the plurality of deep neural networks respectively learn; a step of calculating weighted values to be respectively applied to subcarrier of predetermined symbols by a path channel based on reception power of each subcarrier of the training symbols received by a plurality of path channels; a step of, in responding to first symbols modulated and sent from the transmission end, separately inputting the first symbols respectively received through the plurality of path channels to the plurality of pre-learned deep neural networks to acquire an output value, and deducing a detection symbol with respect to the first symbols from the output value; and a step of coupling detection symbols respectively deduced from the plurality of neural networks by subcarrier to acquire a coupling signal, applying a weighted value corresponding to the subcarrier to the coupling signal, conducting ensemble learning, and detecting each modulation symbol by the subcarrier of the first symbols. According to the present invention, a deep neural network is applied to a receiving end to effectively estimate and compensate a channel, and adaptive ensemble supervised learning is used for overcoming an overfitting problem of the deep neural network, thereby increasing reliability of a system.

Description

[0001] The present invention relates to an SISO-OFDM channel estimation apparatus using a deep neural network based on adaptive ensemble learning learning,

The present invention relates to an apparatus and method for estimating an SISO-OFDM channel using a deep neural network based on adaptive ensemble learning learning, and more particularly, to a SISO-OFDM channel estimation apparatus and method using a deep neural network and an ensemble map learning method, OFDM channel estimation apparatus and method thereof.

Performance for reliability in wireless communication systems continues to be required. In addition, multiple input multiple output (MIMO) technology is used in communication standards such as wireless-fidelity (Wi-Fi) and long term evolution (LTE) to improve the transmission rate.

However, when the channel state is not good due to factors such as the distance between the transmitting end and the receiving end, a signal is transmitted / received using a single input single output (SISO) technique. That is, although MIMO is specified to be selectively implemented in the communication standard, SISO is mandatory to be implemented. SISO is mainly used to guarantee reliability, but channel compensation method in SISO is not mainly studied compared to MIMO. Therefore, an improvement method for the SISO channel estimation and compensation method is required.

Machine learning, a recent issue, in particular the Deep Neural Network, is learned from learning data, away from existing methods that are rigorously programmed to perform specific tasks. Deep neural networks are mainly used mainly in areas such as voice and image recognition and natural language processing.

The field of machine learning can be broadly categorized into instructional learning, non-intelligent learning, and reinforcement learning. Deep neural networks consist of an input layer and an output layer, and one or more hidden layers in between. Each neuron in the hidden layer delivers the output determined by the weight, bias, and activation function to the next layer.

Deep neural networks based on map learning assign labeled data to the input layer during training of the network. Then, the cost function is calculated by comparing the result of the output layer calculated according to the initial weight and bias with the label of training data that is known in advance. Here, the cost function is partially differentiated with respect to the weight and bias that affect the cost function, and the weight and bias are adjusted repeatedly in the direction of decreasing the cost function. When the cost function is lower than the predetermined threshold value, do.

However, the disadvantage of deep neural networks is that the network is over-fitting to the data used for learning. Overfitting means that it is not the data used for learning but that it is not suitable for other data. Therefore, there is a need for a solution to overfitting problems in deep neural networks.

The technique to be a background of the present invention is disclosed in Korean Patent Publication No. 2007-0093557 (published on Sep. 19, 2007).

The present invention relates to an adaptive ensemble map learning-based deep neural network-based SISO-OFDM system capable of solving the over-fitting problem of a deep neural network by estimating and compensating channels using a deep neural network and using adaptive ensemble map learning. And an object of the present invention is to provide a channel estimation apparatus and a method thereof.

The present invention relates to a channel estimation method using an SISO-OFDM channel estimation apparatus, comprising the steps of: inputting training symbols received through a plurality of path channels in correspondence with training symbols sent from a transmitter in a modulated manner, Calculating a weight to be applied to each subcarrier of a predetermined symbol on the basis of the received power of each subcarrier of the training symbol received for each of the plurality of path channels for each of the plurality of path channels; And a first symbol received through each of the plurality of path channels separately from the plurality of learned neural networks to obtain an output value corresponding to a first symbol transmitted from the transmitter, Deriving a detection symbol for one symbol; And combining the detection symbols derived from each of the plurality of NNs on a subcarrier-by-subcarrier basis to obtain a combined signal, applying the weight corresponding to the corresponding subcarrier to the combined signal and performing ensemble learning, And detecting the SISO-OFDM channel, respectively.

Wherein the target for each neuron in the output layer in the neural network is a modulation symbol assigned to each subcarrier of the training symbol sent by the transmitter, and wherein learning the neural network comprises: The weight and bias value applied to each of the plurality of hidden layers between the input layer and the output layer can be finally determined by learning the neural network while inputting the corresponding signal to each neuron of the input layer.

)과 상기 각 뉴런에 대한 타겟 값(

) 간의 유클리디안 거리(E_d)인 아래 수학식의 값이 문턱치 이하가 될 때까지 학습시킬 수 있다.In addition, the step of learning the neural network may include calculating a value output from each neuron in the output layer of the neural network

) And a target value for each neuron (

) It can be learned until the value of the equation is less than the threshold value below between the Euclidean distance (E _d).

Here, N represents the number of subcarriers of the training symbol.

Also, the step of calculating the weights by the path channel may include calculating a weight (W ^{i, j} ) applied to the jth subcarrier of the first symbol received through the i-th path channel, using the following equation have.

는 i에 따른 L개의

중 최대값,

,

는 i번째 경로 채널을 통해 수신된 훈련 심볼의 j번째 부반송파의 전력,

는 L개 경로 채널에 대한

를 모두 합산한 값을 나타낸다.Here, i = {1,2, ... , L}, j = {1,2, ... , N}, L is the number of path channels, N is the number of subcarriers,

Lt; RTI ID = 0.0 > L &

The maximum value,

,

Is the power of the jth subcarrier of the training symbol received on the i < th > path channel,

Lt; RTI ID = 0.0 >

Are added together.

In the ensemble learning step, the j th subcarrier signals of the detection symbols derived respectively corresponding to the plurality of channel paths are combined based on a constellation index, and the N matrices T generating a _j, respectively, targeted to the N matrix T _j, respectively, while varying the j ranging from 1 N, the values summed by the matrix T _j and the weights W ^i, each channel path the product between the ^j deriving a j _{x (x} 1≤J ≤N) of variable j to the maximum, respectively, and the step of respectively detecting the symbol constellation corresponding to the C _Jx on the modulation symbols corresponding to a j-th subcarrier in the first symbol . ≪ / RTI >

Further, the step of generating the matrix T _j is, i but on the basis of Fig index C _x state of a j-th sub-carrier signal of the detected symbol derived in response to the second-path channel constituting the i-th row of the matrix T _j, wherein 1) th column among the N columns of the i-th row and assigning 0 to the remaining columns, and the sum of the elements of each row is 1 for the matrix T _j , The sum can have a range between 0 and L.

Further, the J _x and the modulation symbol can be obtained by the following equations.

는 성분별 곱셈(element-by-element multiplication)을 나타내고,

는 상기 j번째 부반송파에 대응하여 검출된 변조 심볼을 나타낸다.here,

Represents an element-by-element multiplication,

Represents a modulation symbol detected corresponding to the j-th subcarrier.

In addition, the transmitting end transmits the training symbols using one antenna, and the receiving end including the channel estimating device can receive the training symbols through the plurality of path channels using one antenna.

The present invention also provides a neural network learning method for individually inputting training symbols received through a plurality of path channels into a plurality of independent neural networks in correspondence with training symbols sent from a transmitting end, A weight calculator for calculating a weight to be applied to each subcarrier of a predetermined symbol on the basis of the received power of each subcarrier of the training symbol received for each of the plurality of path channels for each of the path channels; A first symbol received through each of the plurality of path channels is separately inputted to the learned plurality of neural networks to obtain an output value and a detected symbol for the first symbol is derived from the output value An initial derivation unit, and a detection symbol derived from each of the plurality of depth- An SISO-OFDM channel estimation apparatus comprising an end detector for detecting a modulation symbol for each subcarrier of the first symbol by performing ensemble learning by applying the weight corresponding to the subcarrier to the combined signal, Lt; / RTI >

According to the present invention, it is possible to effectively estimate and compensate a channel by applying a deep neural network to a receiving end of a SISO-OFDM system, and over fitting problems of a deep neural network can be solved by using adaptive ensemble map learning, .

1 is a diagram illustrating a SISO-OFDM communication system model according to an embodiment of the present invention.
2 is a diagram illustrating a configuration of an SISO-OFDM channel estimation apparatus according to an embodiment of the present invention.
3 is a diagram for explaining a channel estimation method using the apparatus of FIG.
4 is a diagram illustrating a structure of a depth-of-field network for an embodiment of the present invention.
5 is a diagram illustrating a constellation diagram for modulation schemes used in a transmitter.
6 is a diagram illustrating a structure of an adaptive ensemble learning model according to an embodiment of the present invention.
7 is a diagram illustrating a multipath channel and a delay profile according to an embodiment of the present invention.
8 is a chart comparing reliability performance of signal detection between a method according to an embodiment of the present invention and an existing method.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention.

1 is a diagram illustrating a SISO-OFDM communication system model according to an embodiment of the present invention. As shown in FIG. 1, the SISO-OFDM (Single Input Single Output) system includes a transmitter and a receiver. The transmitting end transmits a signal using one antenna, and the receiving end receives a signal using one antenna.

Transmitter converts a data bit to be transmitted into a modulation symbol through a modulator, converts it into an inverse fast Fourier transform (IFFT), transforms it into a time axis, and adds a CP (Cyclic Prefix) . The generated OFDM symbol can be defined as a symbol transmitted by the transmitting end.

An OFDM symbol transmitted from a transmitter is received by a receiver through a multi-path channel between the transmitter and the receiver. Noise is added to the signal received at the receiver. The received OFDM symbol with noise added can be expressed as Equation (1).

Here, Y denotes a received OFDM symbol, H denotes a multi-path channel, X denotes a transmission signal, and N denotes an Additive White Gaussian Noise (AWGN). The multipath channel represents multiple paths formed by scattering or reflecting by obstacles between the transmitting end and the receiving end as shown in FIG.

As shown in FIG. 1, a receiver removes a CP from a received OFDM symbol, and then takes an FFT to convert the CP to a frequency axis. After taking the FFT on the received OFDM symbol, the channel is compensated for signal detection.

In the conventional SISO-OFDM system, a channel is compensated using a previously estimated channel for signal detection. The channel compensation process in the conventional SISO-OFDM system can be expressed by Equation (2).

는 FFT를 통해 주파수축으로 변환된 OFDM 심볼을 나타내며, FFT를 통해 주파수축에서 채널과 송신 신호의 곱셈 연산의 형태로 신호가 변환된다.

는 주파수축에서 추정된 채널,

는 추정된 변조 심볼을 나타낸다. In Equation 2,

Represents an OFDM symbol transformed to a frequency axis through an FFT, and a signal is transformed in the form of a multiplication operation of a channel and a transmission signal on a frequency axis through an FFT.

Is a channel estimated on the frequency axis,

Represents an estimated modulation symbol.

를 복조하여 데이터 비트를 검출한다. Finally, at the receiving end,

To detect a data bit.

In the conventional SISO-OFDM system, there is a reliability problem in the channel compensation process due to amplification of the reception noise generated in the channel estimation error and channel compensation process. However, in the embodiment of the present invention, the above problem can be solved by using a Deep Neural Network (DNN) in channel estimation and compensation.

A deep neural network (hereinafter referred to as a deep neural network) used in this embodiment is based on supervised learning. Map learning implies learning that informs the algorithm what the correct result is for the input.

In a communication system, a channel is estimated using a promised symbol previously known to a transmitter and a receiver. Therefore, the training data used in the depth learning neural network based on the map learning also uses the symbol promised by the transmitter and receiver.

As shown in FIG. 1, a receiver receives an OFDM symbol (hereinafter referred to as a training symbol) for training data generated by a transmitter through a multipath channel, that is, a plurality of path channels. As shown in FIG. 1, the training symbols received through the respective multipath channels are subjected to CP elimination and FFT processes, and then input to respective independent deep neural networks to be used to learn each of the deep neural networks.

The embodiment of the present invention additionally applies an ensemble learning technique for combining multiple neural networks to prevent overfitting occurring in the neural network. In order to make ensemble learning adaptive to the channel state, a weight value to be applied to each subcarrier of a predetermined symbol is calculated for each of the path channels based on the received power for each subcarrier of training symbols received for each of the plurality of path channels I have.

Thereafter, a predetermined symbol received through each path channel is input to each learned neural network, and the derived results are combined to obtain a combined signal. Then, the weights are reflected on the corresponding subcarriers of the combined signal, And performs ensemble learning of enemy relative multiple voting method. Thus, the modulation symbols can be individually detected for each subcarrier of the predetermined symbol.

2 is a diagram illustrating a configuration of an SISO-OFDM channel estimation apparatus according to an embodiment of the present invention. 2, the SISO-OFDM channel estimation apparatus 100 according to the embodiment of the present invention includes a neural network learning unit 110, a weight computing unit 120, an initial derivation unit 130, and a final detection unit 140 do.

The channel estimation apparatus 100 is included in the receiver of FIG. 1, and is used for estimating and compensating a channel and for detecting a signal. Of course, the channel estimating apparatus 100 estimates a channel using a CP-canceled signal and an FFT-processed signal. Hereinafter, for convenience of description, CP removal and FFT processing before channel estimation will be omitted.

The neural network learning unit 110 separately inputs the training symbols received through the plurality of path channels into a plurality of independent deep neural networks corresponding to the training symbols transmitted from the transmitting end, thereby learning each of the plurality of depth neural networks.

The weight calculator 120 calculates a weight to be applied to each subcarrier of a predetermined symbol on the basis of the received power of each subcarrier of the training symbol received for each of the plurality of path channels. The computed weights are used in future ensemble learning.

The initial derivation unit 130 separately inputs first symbols received through a plurality of path channels into a plurality of learned neural networks in response to a predetermined first symbol modulated by the transmitter, thereby obtaining an output value , And derives a detection symbol for the first symbol from the output value.

The final detection unit 140 obtains a combined signal by combining detection symbols derived from a plurality of depth-based neural networks on a subcarrier basis, applies a weight corresponding to the corresponding subcarrier to the combined signal and performs ensemble learning, Respectively.

Hereinafter, a channel estimation method will be described in detail with reference to FIG. 3 is a diagram for explaining a channel estimation method using the apparatus of FIG.

First, a receiver receives a training symbol transmitted from a transmitting terminal through a plurality of path channels (S310).

Then, the neural network learning unit 110 of the in-receiver channel estimation apparatus 100 separately inputs the training symbols received through the plurality of path channels into a plurality of independent deep neural networks to learn each of the plurality of depth neural networks (S320) .

The neural network learning unit 110 inputs the training symbols received through the respective path channels into independent individual neural networks to learn each neural network. Therefore, if there are N path channels, we need N depth NNs and we estimate each N channel NNNs using NNNs.

There is one symbol sent from the transmitting end to the receiving end. However, since the receiving end receives N different channel environments in the receiving end, the neural network is configured and trained for each path channel. Thereafter, the N estimation results by the N depth-of-field neural networks are mutually combined and ensemble-learned, thereby enhancing the reliability of signal detection.

4 is a diagram illustrating a structure of a depth-of-field network for an embodiment of the present invention.

A deep neural network can consist of three parts: an input layer, multiple hidden layers, and an output layer. Each layer may contain multiple neurons.

Learning of the deep neural network means that the input values of each neuron in the input layer and the output values of each neuron in the output layer are determined and the weights and the bias between the layers are optimized, To be able to express well.

In the embodiment of the present invention, a symbol received through a channel is allocated to an input layer, a symbol sent from the transmitting terminal is actually modulated and allocated to an output layer, thereby learning a neural network. This is a corresponding process.

Specifically, a signal corresponding to each subcarrier of the training symbol received through the receiver is input to the input layer of the neural network shown in FIG. Here, of course, a signal obtained by removing the CP and taking an FFT is assigned to the input layer.

The input layer consists of N neurons. N denotes the number of subcarriers in the OFDM symbol. That is, each neuron of the input layer is assigned a signal corresponding to each subcarrier of the received training symbol.

Prior to neural network learning, the weights and bias values applied between neurons in each layer are randomly determined. That is, the initial weight values and the bias values are adjusted and controlled to an optimal value through neural network learning.

As the activation function, sigmoid function, step function, ReLU (rectilinear linear unit) function and tanh (hyperbolic tangent) function can be used, but simulation results show that ERReLU (extended randomized leaky ReLU) Since the best performance is shown, in the embodiment of the present invention, the activation function of the depth neural network uses the ERReLU function.

The target for each neuron in the output layer in the neural network corresponds to the modulation symbol assigned to each subcarrier of the training symbol generated by the transmitter. In this case, since the modulation symbol value is in the form of a complex number, a nonlinear function such that the output converges to 0 or 1 according to the input value, such as the sigmoid function and the tanh function, and the output when the input value is negative ReLU functions such as 0 are not suitable. Therefore, ERReLU is used as an activation function for gains that can be achieved by having a specific complex number output and simultaneously using multiple hidden layers.

In this manner, the neural network learning unit 110 determines a target (a modulation symbol allocated to each subcarrier of the promised training symbol) for each neuron in the output layer, and then determines the weighting factor of each subcarrier of the training symbol actually received on the path channel The signal is input to each neuron part of the input layer of the deep neural network and the neural network is repeatedly learned through it to finally determine the weight (W) and the bias value (b) applied to the multiple hidden layers .

In the case of the iterative learning of the neural network, it is performed until the preset number of times or the error between the output value of the neural network and the target value is within the predetermined value.

는 입력 계층에 입력되는 입력 값,

는 출력 계층을 통해 출력되는 출력 값을 나타낸다. 심층 신경망은 1개의 입력 계층, M-2개의 은닉 계층, 1개의 출력 계층을 포함하여, 총 M개의 레이어로 구성된다.The following describes the activation function in detail. First, in the dense neural network of Fig. 4,

Is an input value input to the input layer,

Represents an output value output through the output layer. The depth neural network consists of a total of M layers including one input layer, M-2 hidden layers, and one output layer.

Of the plurality of hidden layers, the first hidden layer is directly connected to the input layer. The neurons of the first hidden layer receive the result values calculated through the activation function of Equation (3).

The input value of the input layer is multiplied by a weight and then the bias value is added. Similar principles apply to the remaining hidden layers

That is, the neurons of the following hidden layers can receive the result calculated by Equation (4).

Finally, neurons in the output layer can receive the result computed by Equation (5).

The ERReLU function of Equations (3) to (5) is expressed by Equation (6).

The target for the neuron values of the output layer computed by Equations (3) to (5) corresponds to the modulation symbol allocated to each subcarrier of the training symbol transmitted by the transmitting end, as described above. The cost function used to train the weights and bias values is calculated using the Euclidean distance between the neuron value of the output layer and the target.

)과 각 뉴런에 대한 타겟 값(

) 간의 차이를 각 뉴런에 대해 합산하여 유클리디안 거리를 구할 수 있다.The values output from each neuron in the output layer of the neural network (

) And the target value for each neuron (

) Can be summed for each neuron to determine the Euclidean distance.

The Euclidean distance E _d can be calculated by the following equation (7).

는 심층 신경망의 출력 계층을 통해 출력되는 출력 값이고,

는 약속된 신호 즉, 이미 알고 있는 타겟 값이다. 수학식 7을 통해 심층 신경망의 출력 값과 기 설정된 타겟 값 간 오차(유클리디안 거리)를 구할 수 있으며 오차가 문턱 값 이하가 될 때까지 학습을 반복할 수 있다.In Equation (7)

Is an output value output through the output layer of the neural network,

Is a promised signal, i.e., a target value that is already known. The error (Euclidean distance) between the output value of the in-depth neural network and the preset target value can be obtained through Equation (7), and the learning can be repeated until the error becomes less than the threshold value.

The embodiment of the present invention optimizes the weight and bias by using the calculation result of Equation (7) as a cost function and applying a gradient descent algorithm which is a cost minimization algorithm. Optimization of weights and biases is performed repetitively, and learning stops when the cost function is below a predetermined threshold value or reaches the maximum number of iterations.

을 사용하고, 가중치와 바이어스의 조정에 대한 최대 반복 횟수는 5,000번으로 결정하여 사용했다. 또한 학습률(learning rate)은 0.01을 사용했다.In an embodiment of the present invention, the threshold for the cost function is

, And the maximum number of iterations for adjustment of the weight and the bias was determined to be 5,000. The learning rate was 0.01.

Using the above-described method, learning (optimization) of each of the deep-network neural networks is completed for each of the plurality of path channels.

After the optimization of the neural network is completed, the subcarrier signals for the new symbols received corresponding to the predetermined symbols transmitted by the transmitting terminal are input to the respective neurons of the input layer of the neural network, Can be confirmed.

The values output through each neuron in the output layer may be detected as data bits through a demodulation process corresponding to the modulation technique. At this time, it can be adjusted and detected at a close value on the constellation. For example, assuming a BPSK modulation scheme, if the value output from the first neuron of the output layer is [0.2, 1.1], the signal can be detected as a signal of the closest [0, 1] on the constellation .

5 is a diagram illustrating a constellation diagram for modulation schemes used in a transmitter.

BPSK (biphase shift keying) is the modulation technique used in the transmitter. Quadrature phase shift keying (QPSK), quadrature amplitude modulation (16QAM), and 64QAM.

The upper part of FIG. 5 shows the normalization factor used for each modulation scheme (BPSK, 16-QAM, QPSK, 64-QAM) , QPSK, and 64-QAM, respectively.

Higher-order modulation increases the amount of data transmitted by one symbol and increases the data rate. However, since the distance between each constellation point is small, it is vulnerable to noise. Since the transmission power of the transmitter is the same, the constellation for each modulation scheme of FIG. 5 is normalized according to Cm shown in FIG.

Learning of the neural network is performed on a packet basis, and one OFDM symbol (training symbol) for training data required for learning is used per packet. Although each OFDM symbol in the packet passes through the same channel, noise is randomly generated and added to the received signal.

Therefore, even if the cost function is calculated to be less than or equal to the threshold value and the training of the deep neural network is completed, the signal assigned to the neuron of the input layer of the neural network by the reception noise can be varied and the output layer, which is calculated according to the optimized weight and bias, The neuron value of the target may be significantly different from the target.

To overcome this overfitting problem, an embodiment of the present invention further includes an ensemble learning method for calculating a weight for received power from a training symbol received through a multipath, and using the weight at the time of detecting a signal for a predetermined symbol To be applied.

6 is a diagram showing a structure of an ensemble learning model according to an embodiment of the present invention.

Here, it is assumed that the transmitter uses a 64QAM modulation scheme, the OFDM symbol is composed of N subcarriers, and there are four multipath channels between the transmitter and receiver.

These 6 is four independent depth neural network model, corresponding to the four-path channel (1 ^st network, 2 ^nd network, 3 ^rd network, 4 ^th network) and, ensemble weight ^{(Ensemble weight; W 1, j} , W 2 ^{, j} , W ^{3, j} , W ^{3, j} ). Here, j represents a subcarrier index. The four neural networks mentioned in FIG. 6 show learning and optimized state through step S320.

The signal detection process using the ensemble learning model of FIG. 6 corresponds to the steps S340 to S350 of FIG. 3. It is assumed that an unknown received symbol (first symbol) to be detected by the actual receiving terminal, not the training symbol, . Of course, in the case of the weight (W ^{i, j} ) used in the ensemble learning in FIG. 6, the value obtained using the training symbol before step S340, which will be described later in detail in step S330.

In Figure 6 C _x is a constellation of the 64QAM scheme (see (d) in Fig. 5) represents the x-th of the constellation points within the 64 constellation points. Of course, C _x means a value adjusted to the closest constellation point from the neuron value of the output layer to the learned neural network. From the first line to the last line on the constellation diagram, it can be assigned as C ₁ ~ C ₆₄ . For example, C ₁ = 000100 and C ₆₄ = 100000.

Fig each individual depth neural network 6 is for each sub-carrier signal-to-noise ratio: It can be seen that to produce the same _x C both in accordance with (signal to noise ratio SNR), generate a different C _x.

For example, in the case of the first subcarrier (J = 1) in the symbol in FIG. 6, as a result of passing through each path channel, the first, second and fourth deep- ^st, through a 2 ^nd, 4 ^th network), the neural network is a third depth (3 ^rd network that was detected by ₁₁ C, corresponding to a third path, channel) was detected through a C _13. Of course, in this case, the signal of the first symbol time within the first sub-carrier transmitter is actually sent, the probability C ₁₁ days to even higher.

If the SNR is large, the probability of detecting the correct signal increases because the input signal for detection is high in similarity to the data fitted to the learning model of the neural network. Conversely, if the SNR is low, the degree of similarity is low , An over-fitting problem may occur and the probability of detecting the signal may be lowered. Therefore, the embodiment of the present invention calculates the weights from the training symbols received through the multipath, and utilizes the ensemble learning for the relative majority polling method when detecting a signal for a predetermined symbol.

That is, in the embodiment of the present invention, the weight calculator 120 calculates a weight to be applied to each subcarrier of a predetermined symbol on the basis of the received power of each subcarrier of the training symbol received for each of the plurality of path channels, per path channel (S330 ).

Depending on the multipath channel between the transmitting end and the receiving end, the received signal has different received power and delay time for each path channel. In general, the higher the delay, the lower the received power, which may vary depending on the fading.

FIG. 7 is a diagram illustrating a multipath channel and a delay profile according to an embodiment of the present invention. Referring to FIG. FIG. 7 shows a case where there are four path channels between the transmitting end T and the receiving end R. FIG.

7A illustrates that four different path channels are generated by reflection or scattering of signals due to various obstacles existing between the transmitting end T and the receiving end R. FIG. Represents the power magnitude and reception time of the received signal for each path channel.

In the case of path 1, it is seen that the signal goes straight without reflection, and the time delay is the shortest and the reception power is the greatest. In the case of the remaining paths 2, 3, and 4, it can be seen that the signal is reflected at least once and transmitted, and shows more time delay and lower received power.

In order to increase the reliability of the system, it is necessary to give a higher weight to the channel path with high power of the received signal. However, the information of the channel path with high power may be slightly different for every N subcarriers in the OFDM symbol. In the present embodiment, weights can be calculated for each j-th subcarrier by channel path.

That is, the weight calculator 120 calculates a weight to be applied to each subcarrier of a predetermined symbol (hereinafter referred to as a first symbol) for each path channel based on the received power of each subcarrier of the training symbols received for each of the plurality of path channels.

To this end, the weight calculator 120 calculates a weight (W ^{i, j} ) to be applied to the jth subcarrier of the first symbol received through the i-th path channel, using Equation (8) below.

를

로 나누어 정규화된 값으로, 0보다 크고 1 이하의 값을 가진다.Here, i = {1,2, ... , L}, j = {1,2, ... , N}, L is the number of path channels, and N is the number of subcarriers. W ^{i, j} is

To

And is a value that is greater than 0 and less than or equal to 1.

는 i에 따른 L개의

중 최대값을 나타내는 것으로 수학식 9와 같이 정의된다.

Lt; RTI ID = 0.0 > L &

Which is defined as Equation (9). &Quot; (9) "

는 W^i,j 계산을 위해 임시적으로(Temporarily) 저장되는 값으로 아래첨자 T를 부여한 것이며, 이는 수학식 10과 같이 정의될 수 있다.here,

Is a value temporarily stored for the calculation of W ^{i, j,} and is given by the subscript T, which can be defined as shown in Equation (10).

는 i번째 경로 채널을 통해 수신된 훈련 심볼의 j번째 부반송파의 전력,

는 L개 경로 채널에 대한

를 모두 합산한 값이며 수학식 11과 같이 나타낼 수 있다.In Equation (9)

Is the power of the jth subcarrier of the training symbol received on the i < th > path channel,

Lt; RTI ID = 0.0 >

And can be expressed by Equation (11).

In this manner, the weights to be used for the ensemble learning of the j-th subcarrier can be calculated for each channel.

5, a weight applied to a second subcarrier (j = 2) for each i-th path channel is defined as ^{Wi, 2} . Of course i = {1,2, ... , L}. In the following, it is assumed that L = 4.

Among them, a method of obtaining W ^1,2 is exemplified. ^W1,2 represents a weight to be applied to a second subcarrier (j = 2) for a predetermined symbol (first symbol) received through a first path channel (i = 1).

를 구한다. 그리고, 수학식 10과 같이, 각각의 전력 값을

로 나누어 정규화 함으로써, W_T ^1,2, W_T ^2,2, W_T ^3,2, W_T ^4,2 값을 각각 구한다.First, the power values of the training symbols received through the four channel paths are P (Y ^1,2 ), P (Y ^2,2 ), P (Y ^3,2 ), P (Y ^4,2 ) By summing them through Equation (11)

. Then, as shown in Equation 10,

, W _T ^1,2 , W _T ^2,2 , W _T ^3,2 and W _T ^4,2 , respectively.

는 W_T ^3,2가 된다.Then, using the equation (9), the maximum value among the four W _T values is selected. If the maximum value is W _T ³ , 2, then according to equation (9)

^Becomes W _T ^3,2 .

Then, ^W1,2 is obtained through the equation (8). That is, W _T ^1,2 is normalized by dividing it by W _T ^3,2 . Accordingly, W ^1,2 has a value between 0 and 1. Of course, if the maximum value of the four W _T values was W _T ^1,2, then W _T ^1,2 = 1 by Equation (9), and the rest W ^2,2 , W ^3,2 , W ^4,2 1 < / RTI >

After calculating the weights, it is possible to perform reliable signal detection through ensemble learning on the actually received signals at the receiving end.

To this end, the initial derivation unit 130 separately inputs the first symbols received through the plurality of path channels to a plurality of learned neural networks, corresponding to the first symbols modulated by the transmitter, Then, a detection symbol for the first symbol is derived from the output value (S340).

The _Cx values of FIG. 6 represent the detected symbols for each subcarrier for the first symbol derived using four independent deep neural networks, respectively. By the way, it is possible both for the sub-carrier j from the results in Figure 6 confirm that derived the same C _x, or derived some other C _x.

For example, since the 4-th subcarrier derived C ₁₃ all about all four of the paths for channel (i = 4), the final signal to be detected will be the course C _13. However, 2 to ₅₄ C for 1 and 3 times the channel of the four-path channel for sub-carrier (i = 2), it can be seen that have been derived for C ₅₂ times the remaining 2,4-channel, in which case the second It is difficult to know exactly whether the subcarrier signal is C ₅₂ or C ₅₄ .

In this embodiment, reliable signal detection can be performed through ensemble learning using a weight added to each j-th subcarrier for each channel. The weight is generated in step S330.

Accordingly, after step S340, the final detector 140 combines the detected symbols derived from the plurality of neural networks for each subcarrier to obtain a combined signal, and then applies a weight corresponding to the corresponding subcarrier to the combined signal to perform ensemble learning. , The modulation symbol is detected accurately for each subcarrier of the first symbol (S350).

First, the final detection unit 140 combines the jth subcarrier signals of the detected symbols derived respectively corresponding to the plurality of channel paths based on the constellation index, and generates N matrices _Tj, which are combined signals for each subcarrier, do.

Here, j = {1,2, ... , N}, the matrix T _j is generated in correspondence to each subcarrier, as shown in FIG. 6, in this embodiment. This matrix T _j represents a matrix for a relative majority vote corresponding to a modulation scheme for each subcarrier for ensemble learning.

For generating the matrix T _j , the final detector 140 constructs the i-th row of the matrix T _j based on the constellation index C _x of the j-th subcarrier signal of the detected symbol derived corresponding to the i-th path channel, the matrix is generated in such a manner that 1 is assigned to the xth column out of N columns in the ith row and 0 is assigned to the remaining columns.

At this time, for the matrix T _j , the sum of the elements of each row is 1, and the sum of the elements of each row has a range between 0 and L. [

For example, T ₂ obtained by combining detection symbols C ₅₄ , C ₅₂ , C ₅₄ , and C _{52 of the} second subcarrier (j = 2) in FIG. 6 is expressed by Equation (12).

T ₂ has 64 columns corresponding to N = 64, which is the number of subcarriers, and has four rows corresponding to L = 4, which is the number of path channels. In addition, each row is constructed corresponding to each path channel, and only the 54th column in the first row has a value of 1, and the remaining columns have a value of 0 corresponding to C ₅₄ for the first path channel. _Similarly , corresponding to C ₅₂ for the second path channel, the second row has a value of 1 in the 52nd column and a value of 0 in the remainder.

As a result, the sum of the elements of each row becomes 1 unconditionally. In the example of Equation 12, the sum of the 52nd column and the sum of the 54th column are 2, and the sum of the elements of the remaining columns is 0 in the sum of the elements of each column. If all of the detection symbols of the second subcarrier (j = 2) were C ₅₄ , the sum of the elements in the 54th column would be 4 and the sum of the combustion of the remaining columns would be zero.

Thus, in the after generating a T _j, targeting the N matrix T _j, respectively, using equation 13 below, while varying the j ranging from 1 N, the matrix T _j and the weights W ^i, the product between the ^j It derives a j _{x (x} 1≤J ≤N) of variable j to the value of the summation for each channel path to the maximum, respectively.

는 성분별 곱셈(element-by-element multiplication)을 나타낸다. 성분별 곱셈이란 행렬과 일반 값을 대응요소 간 곱셈하는 연산에 해당한다.Here, argmax _j (·) is an operation for deriving a variable j that maximizes the operation value in the parentheses,

Represents an element-by-element multiplication. The multiplication by component corresponds to an operation of multiplying a matrix and a general value by the corresponding elements.

In this case, the final detector 140 detects the symbol corresponding to C _Jx on the constellation diagram as a modulation symbol corresponding to the jth subcarrier in the first symbol, using Equation (14) below.

는 상기 j번째 부반송파에 대응하여 검출된 변조 심볼을 나타낸다.here,

Represents a modulation symbol detected corresponding to the j-th subcarrier.

=C₅₄가 된다. 즉, 제1 심볼 내 3번째 부반송파에 대응하는 변조 심볼

은 성상도 상의 C_Jx에 해당하는 심볼인 C₅₄로 검출한다. C₅₄는 도 5의 성상도에서 [101011] 값에 해당한다.If j, which maximizes the value in the parentheses in equation (13), is 54, then J _x = 54. In this case, according to Equation (14)

= C ₅₄ . That is, a modulation symbol corresponding to the third subcarrier in the first symbol

Is detected as C _{54 which} is a symbol corresponding to C _Jx on the constellation. C ₅₄ corresponds to the value [101011] in the constellation diagram of FIG.

내지

을 검출한다. As described above, the calculation according to Equation (11) is performed separately for N subcarriers constituting an OFDM symbol, and finally, as shown in FIG. 6, N modulation symbols

To

.

The embodiment of the present invention as described above precisely adjusts and controls the weight and the bias using the neural network, thereby reducing the channel estimation error. Further, in order to prevent reliability degradation at the time of channel compensation, the received signal is input to the learned neural network to prevent noise from being amplified and improve reliability.

Also, in order to solve the problem of over-fitting of the depth-of-field neural network to the training data, a plurality of mutually independent deep-layer neural networks are individually learned using each training signal received through the multi-path between the transmitting end and the receiving end, And determines the weight to be applied to each subcarrier of the predetermined signal based on the received power for each subcarrier with respect to the received training signal.

Then, each signal received through the multiple paths is individually applied to a plurality of learned neural networks to acquire respective outputs, and then combine and apply the ensemble learning adaptively by adding related weights, thereby solving the overfitting problem do.

8 is a chart comparing reliability performance of signal detection between a method according to an embodiment of the present invention and an existing method.

 In FIG. 8, the downward signal-to-noise ratio (SNR) is high and the noise is low. The numerical values listed in each case in FIG. 8 indicate a SER (symbol error rate) according to the SNR and an error rate.

 For each of the BPSK, QPSK, 16QAM and 64QAM schemes, the conventional method (Conv) for detection of the signal (Conv), the case of using a single depth neural network (DNN), and the ensemble learning for multiple depth neural networks (DNN + Ensemble learing) using Equation (8), and the SER result for adaptive ensemble learning (DNN + Adaptive Ensemble Learning) using the weight according to Equation (8).

Comparing the results of the conventional channel estimation and compensation method (Conv) and the single neural network (DNN) results, the detection reliability performance of the signal by the neural network at the low SNR, Method and has improved reliability performance over certain SNRs. This is because, at a low SNR, the signal used for learning of the neural network greatly fluctuates due to a large noise relative to the power of the signal, and learning is not performed properly. Also, Therefore, the network model whose learning has been completed is not suitable for the signal to be detected.

According to the results shown in FIG. 8, the signals of 2 dB, 6 dB, 18 dB, and 26 dB for the BPSK, QPSK, 16QAM, and 64QAM modulation techniques are compared with the existing channel estimation and compensation method (Conv) And has improved reliability performance over the large noise ratio. As the signal - to - noise ratio increases over the signal - to - noise ratio for learning the depth - of - field neural networks according to each modulation scheme, the reliability performance improves sharply compared with the conventional method.

12 dB, 20 dB, and 28 dB for each of the BPSK, QPSK, 16QAM, and 64QAM techniques compared to the existing channel estimation compensation method in the case of additional ensemble learning (DNN + ensemble learning) It can be confirmed that the reliability is improved at the noise ratio or more. The reason why the reliability performance is improved at the higher signal-to-noise ratio when the ensemble learning is applied is that the reliability for each model used for the ensemble learning should be more than a certain level. However, it can be seen that as the signal-to-noise ratio increases over the specific signal-to-noise ratio according to each modulation scheme, the reliability performance improves more sharply than the case of using a single deep neural network.

In addition, adaptive ensemble learning (DNN + adaptive ensemble learning) using adaptive ensemble learning that uses the power of the received signal in multipath is applied to each of BPSK, QPSK, 16QAM, and 64QAM techniques And improved reliability over signal-to-noise ratios of 4dB, 10dB, 16dB, and 24dB. This result is almost similar to the case of using a single depth neural network. However, when compared with the above three cases, it can be seen that the reliability performance is improved most dramatically according to the signal-to-noise ratio. Also, since the signal-to-noise ratio is 20 dB or more in a general communication system environment, it can be confirmed that the channel estimation and compensation method provided by the present invention has a great effect on improving the reliability according to the communication system environment.

According to the present invention, it is possible to effectively estimate and compensate a channel by applying a deep neural network to the receiver of the SISO-OFDM system, and to solve the over-fitting problem of the deep neural network using adaptive ensemble map learning, It is possible to increase the reliability.

While the present invention has been described with reference to exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims.

100: channel estimation apparatus 110: neural network learning unit
120: Weight calculation unit 130: Initial derivation unit
140:

Claims (16)

A channel estimation method using a SISO-OFDM channel estimation apparatus,
A step of separately inputting the training symbols received through a plurality of path channels to a plurality of independent depth NNs corresponding to training symbols sent from the transmitting end, thereby learning each of the plurality of depth NNs;
Calculating, for each of the plurality of path channels, a weight to be applied to each subcarrier of a predetermined symbol on the basis of received power of each subcarrier of the training symbol received for each of the plurality of path channels;
A first symbol received through each of the plurality of path channels is input to each of the plurality of learned neural networks in response to a first symbol modulated by a transmitting end to obtain an output value, Deriving a detection symbol for the received signal; And
And combining the detected symbols derived from the plurality of neural networks with each other on a subcarrier basis to obtain a combined signal and applying the weight corresponding to the corresponding subcarrier to the combined signal to perform ensemble learning to obtain a modulation symbol for each subcarrier of the first symbol, And detecting each SISO-OFDM channel. The method according to claim 1,
Wherein the target for each neuron in the output layer in the neural network is a modulation symbol assigned to each subcarrier of the training symbol sent from the transmitter,
The step of learning the deep-
Wherein the training neural network learns a signal corresponding to each subcarrier of the received training symbol in each neuron of the input layer and calculates a weight and a bias applied to each of a plurality of hidden layers between the input layer and the output layer, SISO-OFDM channel estimation method. The method of claim 2,
The step of learning the deep-
A value output from each neuron in the output layer of the neural network

) And a target value for each neuron (

(E _d ) until the value of the following equation becomes equal to or less than the threshold value: SISO-OFDM channel estimation method:

Here, N represents the number of subcarriers of the training symbol. The method according to claim 1,
The step of calculating the weight value for each of the path channels includes:
(W ^{i, j} ) applied to a jth subcarrier of a first symbol received through an i-th path channel, using the following equation: SISO-OFDM channel estimation method:

Here, i = {1,2, ... , L}, j = {1,2, ... , N}, L is the number of path channels, N is the number of subcarriers,

Lt; RTI ID = 0.0 > L &

The maximum value,

,

Is the power of the jth subcarrier of the training symbol received on the i < th > path channel,

Lt; RTI ID = 0.0 >

Are added together. The method of claim 4,
Wherein the ensemble learning step comprises:
Combining the jth subcarrier signals of the detection symbols derived respectively corresponding to the plurality of path channels based on a constellation index to generate N matrices _Tj as the combining signals for the subcarriers;
While variable targeting the N matrix T _j, respectively, the j ranging from 1 N, the matrix T _j and the weights W ^i, the J _x variable j to a value summed for each path channel multiplication between ^j up to (1 &_lt; / = J _x &_lt; N), respectively; And
And detecting a symbol corresponding to C _Jx on the constellation as a modulation symbol corresponding to a j-th subcarrier in the first symbol, respectively. The method of claim 5,
The step of generating the matrix T _j comprises:
th row of the matrix T _j based on the constellation diagram index C _x of the jth subcarrier signal of the detected symbol derived corresponding to the i-th path channel, and only the xth column out of the N columns of the i-th row And assigning 0 to the remaining columns,
Wherein the sum of elements for each row is 1 and the sum of the elements for each column is between 0 and L for the matrix T _j . The method of claim 5,
The J _x and the modulation symbol estimates SISO-OFDM channel is obtained through the following equation by:

here,

Represents an element-by-element multiplication,

Represents a modulation symbol detected corresponding to the j-th subcarrier. The method according to claim 1,
The transmitting terminal transmits the training symbols using one antenna,
Wherein the receiver including the channel estimator receives the training symbols over the plurality of path channels using one antenna. A neural network learning unit for separately inputting the training symbols received through a plurality of path channels to a plurality of independent deep neural networks corresponding to training symbols sent from the transmitting end, thereby learning each of the plurality of depth neural networks;
A weight calculator for calculating a weight to be applied to each subcarrier of a predetermined symbol on the basis of the received power of each subcarrier of the training symbol received for each of the plurality of path channels;
A first symbol received through each of the plurality of path channels is input to each of the plurality of learned neural networks in response to a first symbol modulated by a transmitting end to obtain an output value, An initial derivation unit for deriving a detection symbol for the received signal; And
And combining the detected symbols derived from the plurality of neural networks with each other on a subcarrier basis to obtain a combined signal and applying the weight corresponding to the corresponding subcarrier to the combined signal to perform ensemble learning to obtain a modulation symbol for each subcarrier of the first symbol, And a final detector for detecting the SISO-OFDM channel. The method of claim 9,
Wherein the target for each neuron in the output layer in the neural network is a modulation symbol assigned to each subcarrier of the training symbol sent from the transmitter,
Wherein the neural network learning unit comprises:
Wherein the training neural network learns a signal corresponding to each subcarrier of the received training symbol in each neuron of the input layer and calculates a weight and a bias applied to each of a plurality of hidden layers between the input layer and the output layer, SISO-OFDM channel estimator. The method of claim 10,
Wherein the neural network learning unit comprises:
A value output from each neuron in the output layer of the neural network

) And a target value for each neuron (

(E _d ) until the value of the following equation becomes equal to or less than a threshold value:

Here, N represents the number of subcarriers of the training symbol. The method of claim 9,
The weight computing unit may include:
(W ^{i, j} ) applied to a jth subcarrier of a first symbol received through an i-th path channel, using the following equation: SISO-

Here, i = {1,2, ... , L}, j = {1,2, ... , N}, L is the number of path channels, N is the number of subcarriers,

Lt; RTI ID = 0.0 > L &

The maximum value,

,

Is the power of the jth subcarrier of the training symbol received on the i < th > path channel,

Lt; RTI ID = 0.0 >

Are added together. The method of claim 12,
The final detection unit includes:
Combining the j th subcarrier signals of the detection symbols derived respectively corresponding to the plurality of path channels based on constellation indexes to generate N matrices T _j as the combining signals for the subcarriers,
While variable targeting the N matrix T _j, respectively, the j ranging from 1 N, the matrix T _j and the weights W ^i, the J _x variable j to a value summed for each path channel multiplication between ^j up to (1? J _x? N), respectively,
And detects a symbol corresponding to C _Jx on the constellation as a modulation symbol corresponding to a j-th subcarrier in the first symbol, respectively. 14. The method of claim 13,
It shall be composed of the i-th row of the matrix T _j on the basis of Fig index C _x state of a j-th sub-carrier signal of the detected symbol derived in response to the generation of the matrix T _j, i-th path channels, the i-th row The matrix is generated in such a manner that only the xth column among the N columns is assigned 1 and the remaining columns are assigned 0,
Wherein the sum of the elements of each row is 1 and the sum of the elements of each column is in a range between 0 and L with respect to the matrix T _j . 14. The method of claim 13,
The J _x and the modulation symbol is a SISO-OFDM channel estimation device that is obtained through the following equation:

here,

Represents an element-by-element multiplication,

Represents a modulation symbol detected corresponding to the j-th subcarrier. The method of claim 9,
The transmitting terminal transmits the training symbols using one antenna,
Wherein the receiver including the channel estimator receives the training symbols over the plurality of path channels using one antenna.

Images