KR20020085302A - Decision feedback recurrent neural network equalizer and learning method for the equalizer - Google Patents

Abstract

PURPOSE: A decision feedback recurrent neural network equalizer and a learning method being used in the same are provided to obtain a stable equalizing performance by performing a neural network learning effectively although an interference and a non-linear distortion are strong. CONSTITUTION: The first buffers(10) include a tap delay line structure for delaying a channel input signal(Xk) by a predetermined period of data. Nodes(20) perform a function calculation by inputting the delayed channel input signal(Xk), decision feedback delay signals, and feedback outputs of each node. Decision devices(30) calculate and output a presumptive value of a digital form by substituting output signals(Yk) of each node(20) with absolute values. The second buffers(40) delay and output the presumptive value being output from the decision devices(30) as a decision feedback delay signal. A switching unit(S) selects one out of the output of the decision devices(30) and a learning signal(dn) and inputs the selected one to the second buffer(40). Accumulators(50) accumulate a difference between each output signal(yn) and the learning signal(dn) of the nodes(20) and calculate an instantaneous error(en).

Description

DECISION FEEDBACK RECURRENT NEURAL NETWORK EQUALIZER AND LEARNING METHOD FOR THE EQUALIZER}

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an equalization device for digital communication, and more particularly, a decision feedback recurrent neural network equalizer (DFRNNE) used to compensate for interference and channel distortion between signals in a digital communication system. It's about learning.

Channel equalization can be considered as a mode classification problem. Since neural networks have excellent mode classification characteristics, various neural network structures are being applied to adaptive channel equalization with the development of neural network technology. Each of the various structures and algorithms of the equalizers has its own advantages and disadvantages. For example, multilayer perceptrons can be applied to any complex nonlinear channel equalization if sufficient degrees of freedom are provided. However, in implementing various structures and algorithms of neural equalizers, there is a contradiction between the complexity of the implementation and the features of the algorithm. The larger the structure, the longer the time required for calculation, the smaller the data transfer rate.

Recurrent Neural Networks are widely used because of their small size and excellent performance, and because of the advantages of alleviating such contradictions in channel equalization. A recursive neural network is similar to an infinite impulse response (IIR) filter, so you get good equalization performance, and even complex nonlinear maps can be completed with just a few nodes.

1 illustrates a typical Recurrent Neural Network Equalizer (RNNE) configuration. As shown in Fig. 1, the output of each node in a general recursive neural network is fed back, which in turn impairs the consistency and stability of the recurrent neural network equalizer. After setting the initial values for learning, the network is trained and then switched to the data transmission mode. In general, RNNE is very sensitive to various initial values set for performing the learning mode. If the gender is strong there is a problem that normal learning is not made. Such instability of learning may cause delays in system operation and may cause communication failure.

Accordingly, an object of the present invention was created to solve the above-described problems of the prior art, and the decision feedback recursive neural network equalizer capable of obtaining stable equalization performance by effectively performing neural network learning even in case of strong inter-symbol interference and nonlinear distortion. And learning methods used therefor.

It is still another object of the present invention to provide a decision feedback recursive neural network equalizer and a learning method used therein, in which an adaptive equalizer can quickly and stably converge a learning algorithm through adaptive adjustment of a learning stage.

1 is a diagram illustrating a general recurrent neural network structure.

2 is a block diagram illustrating a decision feedback recursive neural network configuration having a serial output according to a preferred embodiment of the present invention.

3 is a block diagram illustrating a decision feedback recursive neural network configuration with parallel outputs in accordance with another embodiment of the present invention.

4 is a diagram illustrating a learning curve of a general recursive neural equalizer having different learning stages (α).

5 to 8 are examples of learning curves for explaining performance comparison between a decision regression recursive neural network equalizer (S-DFRNNE) and a general recursive neural network equalizer (RNNE) having a serial output according to an embodiment of the present invention.

Decision return recursive neural network equalizer according to a preferred aspect of the present invention for achieving the above object;

First delay buffers in which a plurality of buffers for delaying a channel input signal form a line;

Arithmetic means for receiving the channel input signals, the decision feedback delay signals, and the feedback output of each node, which are delayed in the delay buffers, and for calculating the respective weights and functions;

Determinators for calculating and outputting an estimated value in digital form by substituting an absolute value of each output signal of the computing means;

Second delay buffers for forming a line with a plurality of buffers for delaying the digital form estimated values outputted from the determiners as the decision feedback delay signal;

Switching means for selecting one of an output of the determiners and a learning signal and inputting it to the second delay buffer;

And accumulating the difference between the output signal and the learning signal of each of the nodes to calculate an instantaneous error for the weight correction.

Also, an equalizer according to an additional aspect of the present invention substitutes an absolute value of the output signal of each of the computing means to calculate an estimated value in digital form, and inputs it to the computing means as a decision feedback delay signal through the second delay buffers. Second determinants for,

Second switching means for selecting one of an output of the second determinants and a learning signal and inputting it to the second delay buffer;

And a second accumulator for accumulating the difference between the output signal and the learning signal of each of the nodes to calculate an instantaneous error for the weight correction.

Hereinafter, the configuration and operation of a decision feedback recursive neural network equalizer according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

2 is a block diagram illustrating a configuration of a decision feedback recursive neural network equalizer (S-DFRNNE) having a serial output according to a preferred embodiment of the present invention. First buffers 10 having a tap delay line structure and a delayed channel input signal Nodes and the output signals of the nodes 20 and the nodes 20, which function Replace with an absolute value to estimate the digital form ( Determinants 30 for calculating and outputting?), Second buffers 40 for delaying and outputting the digital estimated values outputted from the determinators 30 as decision feedback delay signals, and the determiner 30 Output and learning signal Switching means (S) for selecting one of the inputs to the second buffer 40 and an output signal of each of the nodes 20 ) And learning signal ( ) By accumulating the difference It is composed of accumulators (50) yielding.

In FIG. 2, n, m, and l are defined as the number of internal nodes, the number of delay inputs, and the number of decision feedback delay inputs, respectively.

Referring to FIG. 2, first, an input signal supplied from a channel in a data transmission process Is delayed in units of data periods after passing through the first buffers 10 having the tap delay line structure, multiplied by a weight, and input to the n nodes 20. At the same time, the previous output values of the nodes 10 are calculated and output by the determinants 30 as estimated values. This estimated value is passed through the serially connected second buffers 40, delayed by data period, and multiplied by a weight to be input to the n nodes 20 again. Accordingly, each node 20 outputs a value according to Equation 1 below. Equation 1 shows the sum of input weights of the p (p = 1,2, .., n) th node.

In Equation 1 Is the delay input signal, Is the output of the i th node, (d is the channel delay) for the feedback delay signal, Is the weight from the i th node to the p th node, Is the weight from each delayed input signal (j = n + 1, ...., n + m) to the p-th node, Denotes the weight from each decision feedback delay signal (h = n + m + 1, ..., n + m + l) to the p-th node. The decision feedback delay signal may be a learning signal ( In the direct decision mode, that is, in the data transmission mode, the output value of each node 20 is determined.

For reference In this case, f (x) can be expressed as Equation 2 below, and the estimated value is expressed as Equation 3 below.

3 is a block diagram illustrating a configuration of a decision feedback recurrent neural network equalizer with parallel outputs (P-DFRNNE) according to another embodiment of the present invention. Each node 20 is connected to two determinants 30 and 35, and each determinant 30 and 35 is also determined by switching of the switching means S as shown in FIG. 35) return each output to the second buffer 40. The sum of the input weights of the p-th node in the P-DFRNNE is shown in Equation 4 below, and the remaining equations are the same as in Equation DFRNNE having the serial output.

In Equation 4 Denotes the decision feedback input signal of the i-th node with d + 1 time delay.

On the other hand, a learning method in a recursive neural equalizer generally calculates output values of nodes, calculates instantaneous error values, calculates a sensitivity function, and calculates weights using the calculated instantaneous error values and sensitivity functions. It consists of a repetition of the step of correcting.

In the present invention, the weights must be corrected in real time through the above four steps of repetition for each of the two neural network equalizer structures (S-DFRNNE, P-DFRNNE) described above. In this case, the real-time recursive learning (RTRL) algorithm Modifications must be made. In the present invention, the learning algorithm can be converged more quickly and stably due to the adaptive adjustment of the learning stage α, which will be described in more detail below.

First, in the present invention, as step 1, the output values of the nodes 20 are calculated by the above equation (1).

In step 2, the instantaneous error value is calculated. The instantaneous error of the h-th neuron may be defined by Equation 5 below, and the total instantaneous error of the neural network may be obtained by Equation 6 below.

On the other hand, as step 3, a sensitivity function is calculated. In order to calculate this sensitivity function, (Kronecker delta) is defined as Equation 7 below, and the derivative of Equation 2 (Equation 8) should be obtained. The sensitivity function may be expressed as in Equation 9 below.

Then, each instantaneous error calculated as the last step is corrected by the following equation (10) using the sensitivity function.

In Equations 9 and 10 Denotes the learning phase of the adaptive equalizer. And > 0, p = 1,2, ..., n; i = 1,2, ..., n; j = 1,2, ..., n, n + 1, ...., n + m, n + m + 1, ...., n + m + l. Is the internal nodes (j = 1, 2, ..., n) of the decision feedback recursive neural network equalizer having the serial output shown in FIG. + m) or the output of the feedback input signal (j = n + m + 1, ..., n + m + l). Moreover, the moment coefficient u (0 <u <0.001) is derived to update the weights.

In decision-recursive recursive neural network equalizer (P-DFRNNE) with parallel output, 1, 2, ..., n, n + 1, ..., n + m, n + m + l, ..., n + Except for j defined by m + n, the equations described in S-DFRNNE are the same.

4 is a diagram illustrating a learning curve of a general recursive neural equalizer having different learning stages α, and FIGS. 5 to 8 are decision feedback recursive neural network equalization having a serial output according to an embodiment of the present invention. An example of a learning curve for explaining the performance comparison between the S-DFRNNE and the general recursive neural network equalizer (RNNE) is shown.

First, the learning step α shown in Equation 10 generally affects the convergence speed of the algorithm. As shown in Fig. 4, the convergence speed of the algorithm increases within a certain range with the increase of the learning stage α. As α increases, the algorithm converges quickly but is relatively oscillating and unstable. The smaller α is, the more slowly the algorithm converges and falls to a local minimum.

Therefore, the proper selection of the learning phase α is one way to improve the convergence speed of the equalizer. To this end, in the present invention, by continuously adjusting the learning step α after every iteration, the learning algorithm skips the local minimum value to improve the convergence processing speed and solve the algorithm instability.

First of all the total error in the system It is defined as And we use exponential function to adjust α adaptively. These functions are considered En as an independent variable , Set to. The learning algorithm will then adaptively adjust the learning phase α during the iterative process of learning. If the total error is large, α will decrease; if the total error is small, α will increase. Thus, the overall error will be small despite the increase in repetition, and α will gradually maintain a certain level.

Hereinafter, a simulation result of a general recursive neural network equalizer (RNNE) and a decision feedback recursive neural network equalizer (S-DFRNNE) having a serial output according to an embodiment of the present invention will be described.

First weight , Initial sensitivity And the initial output of each node Is These are random numbers with absolute values equal to or less than, and the general RNNE and S-DFRNNE parameters are m = 2, n = 1, α = 0.5, and m = 2, n = 1, l = 3, = 0.5, two groups of learning curves are obtained according to different distortion channels.

In Figure 5 (learning curve), the distortion channel is LCH, SNR = 16 dB. In FIG. 6 (LCH + NLCH learning curve), the distortion channel is LCH cascaded with NLCH and SNR = 18 dB. According to FIG. 5, S-DFRNNE converges at about -30 dB, while RNNE carries greater vibration and converges at about -10 dB. According to FIG. 6, the S-DFRNNE converges at about −30 dB, and the RNNE does not converge. In short, it can be seen that S-DFRNNE has better learning performance than general RNNE.

Comparing the bit error rate BER below, two BER curves as shown in FIGS. 7 and 8 can be obtained for the above-described distortion channel. According to these curves, it can be seen that the performance of the S-DFRNNE is superior to the RNNE performance. During the simulation process, it was found that for two given distortion channels, the BER of RNNE always showed inconsistencies and was accompanied by some vibration. And BER performances decrease with increasing SNR, which is a drawback of RNNE applied to channel equalization, while S-DFRNNE compensates for the drawback and shows consistent BER.

In conclusion, the present invention uses the exponential function to vary the adaptive learning phase α after every iteration, so that neural network learning can be performed stably and quickly even with severe intersymbol interference and nonlinear distortion, resulting in stable equalization performance. You will get it.

As described above, the present invention provides an exponential function of instantaneous error for each iteration for learning the neural network. By varying in proportion to), even if severe intersymbol interference and nonlinear distortion are strong, neural network learning can be performed stably and quickly so that normal communication can be achieved. In addition, the present invention also has an effect of improving the compatibility of the structure of the recursive neural equalizer by providing a decision feedback recursive neural network equalizer having a parallel output.

On the other hand, the present invention has been described with reference to the embodiments shown in the drawings, which are merely exemplary, and those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention should be defined only by the appended claims.

Claims (6)

In the crystal feedback recursive neural network equalizer, First delay buffers in which a plurality of buffers for delaying a channel input signal form a line; Arithmetic means for receiving the channel input signals, the decision feedback delay signals and the feedback output of each node, which are delayed in the delay buffers, and for calculating the respective weights and functions; Determinators for calculating and outputting an estimated value in digital form by substituting an absolute value of the output signal of each computing means; Second delay buffers for forming a line with a plurality of buffers for delaying the digital form estimated values output from the decision units as the decision feedback delay signal; Switching means for selecting one of an output of said determiners and a learning signal and inputting it to said second delay buffer; And an accumulator for accumulating the difference between the output signal and the learning signal of each of the nodes and calculating an instantaneous error for the weight correction. The decision feedback recursive neural network equalizer of claim 1, wherein the output of the computing means has a value calculated according to Equation 11 below. Is a delayed input signal. Is the output of the i node. (d is the channel delay) is the feedback delay signal Is the weight from the i th node to the p th node Is a weight from each delayed input signal (j = n + 1, ...., n + m) to the p th node. Is the weight from each decision feedback delay signal (h = n + m + 1, ..., n + m + l) to the pth node. The apparatus of claim 1, further comprising: second determinators for substituting an absolute value of an output signal of each of the computing means to calculate an estimated value in digital form and inputting the estimated value to the computing means as a decision feedback delay signal through the second delay buffers. and; Second switching means for selecting one of an output of said second determiners and a learning signal and inputting it to said second delay buffer; And a second accumulator for accumulating the difference between the output signal and the learning signal of each of the nodes and calculating an instantaneous error for the weight correction. 4. The decision feedback recursive neural network equalizer of claim 3, wherein the output of the computing means has a value calculated according to Equation 12 below. Is the decision feedback input signal of the i-th node with d + 1 time delay. In the neural network learning method of the decision feedback recursive neural network equalizer, A first step of receiving the channel-delayed input signals, the decision feedback delay signals of the output terminal, and the feedback output of each node, and outputting the result by performing a weight and a function operation; Calculating an instantaneous error of each node necessary for the weight correction; Calculating a sensitivity function based on Equation 13 below; The neural network of the decision feedback recursive neural equalizer, comprising the fourth step of calculating a weight based on Equation 14 using the calculated instantaneous error value and the sensitivity function and correcting the weight based on the weight of the first step. Learning method. , Moment factor u: 0 <u <0.001 Instantaneous error on each node: Learning signal: The method according to claim 5, wherein the learning step α of the fourth step is varied by a value calculated based on Equation 15 at every iteration for neural network learning. , En is the neural network full instantaneous error.