KR100262961B1 - Equalization Apparatus and Method Using Decision Feedback Recursive Neural Network - Google Patents

Abstract

PURPOSE: A decision feedback recurrent neural equalizer and a method thereof are provided to efficiently compensate for an interference between symbols and non-linear distortion, and to reduce calculating amount through simplifying a structure. CONSTITUTION: Delay units delay input signals from a channel for a data frequency unit. Calculating units are connected to the delay lines for calculating the delayed signals from the delay lines and a feedback signal from an output terminal. A decision value detect unit detects digital signals through substituting the output signals from the calculating unit into absolute values. Then, the detect unit supplies the digital signals to the calculating unit after feeding back.

Description

Equalizing Apparatus Using Decision Feedack Recurrent Neural And Method Thereof

The present invention relates to an equalization device, and more particularly, to a decision feedback recursive neural network equalization device and method capable of effectively compensating signal interference in a magnetic recording / reproducing system.

In general, inter-symbol interference due to the characteristics of the channel in digital communication makes it difficult to use the band as well as reliable communication. The equalizer is used to overcome this problem. In particular, in the case of the magnetic recording / reproducing system, the equalization of the complete channel becomes difficult with the general equalizer due to the nonlinearity caused by the physical characteristics of the channel as well as the intersymbol interference.

A typical method used for effective equalization of a channel with nonlinear distortion is an equalization method using neural networks. To date, the neural network used in the equalizer is called a finite impulse response (FIR) in the form of a finite impulse response (FIR), and an infinite impulse response (IIR). Recursive neural network structure is used.

1 is a block diagram showing the configuration of a conventional decision feedback neural equalizer (NDFE).

In the decision feedback neural network equalizer of FIG. 1, a buffer delay line (TDL) structure having a channel output signal sequence as an input forms an input layer 4. The front filter 6 is filled with the channel output signal sequence, and the estimated value of the determined feedback input signal sequence {a _k } enters the input of the rear filter 8. In addition, the output of the front filter 6 { } Weighted through this detector 10 Is weighted using the Back Propagation (BF) algorithm that is detected and fed back to the input layer. By learning, the input signal sequence is estimated.

In this way, the decision feedback neural network equalizer has a good equalization performance because the estimated value of the decision feedback input signal sequence {a _k } enters the input of the rear filter 8 to compensate for the interference by the preceding symbol. However, in the case of having a very deep valley in the channel band or severe nonlinear distortion, satisfactory equalization is difficult to be achieved, and there is a disadvantage in that the structure is complicated and the calculation amount is large.

Referring to FIG. 2, there is shown a structure of a neural recurrent equalizer (NRE) capable of solving the above-described problem of the decision feedback neural network equalizer.

The recursive neural equalizer of FIG. 2 is a structure in which each node is not only connected in a feedforward manner but partially or completely connected to each other, and a hidden layer exists between an external input vector and an external output vector. . In other words, the recursive neural network equalizer calculates the current node output value and the past output values of each node of the hidden layer are fed back.

As a result, the recursive neural equalizer has an IIR structure and therefore has an advantage of having a small amount of computation. However, likewise, a recursive neural equalizer has a problem in that the channel is severely distorted or strong in the nonlinearity, and thus it is not properly trained and very unstable.

It is therefore an object of the present invention to provide an apparatus and method for decision feedback recursive neural network equalization that can effectively compensate for intersymbol interference and nonlinear distortion.

It is another object of the present invention to provide an apparatus and method for decision feedback recursive neural network equalization that can reduce the amount of computation by simplifying the structure using an IIR structure.

It is still another object of the present invention to provide an apparatus and method for decision feedback recursive neural network equalization that can be easily implemented in hardware by simplifying the structure using an IIR structure.

1 is a block diagram showing the configuration of a crystal feedback equalizer using a conventional neural network;

Figure 2 shows the structure of a conventional recursive neural network equalizer.

Figure 3 is a block diagram showing the configuration of the decision feedback recursive neural network equalizer according to the present invention.

<Brief description of symbols for the main parts of the drawings>

2,12 buffer 4: input layer

6: front filter 8: rear filter

14: Hidden Node 16: Crystallizer

18: adder

In order to achieve the above object, the equalization apparatus using the decision feedback recursive neural network according to the present invention comprises at least one delay means for delaying the input signal supplied from the channel by data period, and the delayed signal and output stage supplied from the delay line. A calculation means connected to the delay line by the number of delay lines so as to perform a functional calculation of the feedback signal from the signal; Equipped.

The equalization apparatus using the decision feedback recursive neural network according to the present invention includes at least one delay means for delaying an input signal supplied from a channel in data period units, and a function operation for a delayed signal supplied from a delay line and a feedback signal from an output terminal. A calculation means connected to the delay line by the number of delay lines, a decision value detection means for detecting a digital signal by returning the output signal of the calculation means to an absolute value, and feeding it back to the calculation means; And a weight correction means for correcting the weight in accordance with the output signal of the detection means.

The equalization method using the decision feedback recursive neural network according to the present invention comprises the steps of delaying the input signal supplied from the channel by data period unit, the function of calculating the delayed signal supplied from the delay line and the feedback signal from the output stage, Detecting the digital signal by returning the function-operated signal to an absolute value and feeding it back to the input of the function operation.

The equalization method using the decision feedback recursive neural network according to the present invention comprises the steps of delaying the input signal supplied from the channel by data period unit, the function of calculating the delayed signal supplied from the delay line and the feedback signal from the output stage, Detecting a digital signal by returning the function-operated signal to an absolute value, feeding it back to the input of the function operation, and correcting weights according to the function-operated signal and the digital signal.

Other objects and advantages of the present invention in addition to the above objects will become apparent from the description of the preferred embodiment of the present invention with reference to the accompanying drawings.

Hereinafter, a preferred embodiment of the present invention will be described in detail with reference to FIG.

3 is a block diagram showing the configuration of an equalizer using the decision feedback recursive neural network according to the present invention.

The decision feedback recurrent neural equalizer (DFRNE) of FIG. 3 includes buffers 12 having a TDL structure and delayed input signals to delay the channel input signal X _k in units of a certain period of data. And digitally determined values by substituting the absolute nodes 14, which are functionally operated by the feedback signal fed back from the output terminal, and the output signals Y ₁ , Y ₂ , Y _k of the hidden nodes 14 with absolute values. It is provided with the decision makers 16 for calculating and feeding back this decision value to the hidden nodes 14. Also, the decision feedback recursive neural network equalizer of the present invention includes adders 18 which are commonly connected to the output line of the hidden node 14 and the output line of the determiner 16 to adjust the weight value.

Unlike the conventional recursive neural equalizer (RNE), the decision-feedback recursive neural equalizer of the present invention does not simply add the weight of the past output value to the output value of the current node, but a structure in which the value determined from the past output value is added to the weight. Have Each hidden node 14 is supplied with the output signals of the respective buffers 12 in common, and the output signals of the respective determiners 16 are supplied in common, and these functions are functionally calculated to determine the result of the determination of the determiner 16. Will be supplied as a signal.

In detail, the input signal supplied from the channel in the decision feedback recursive neural network equalizer. X _k As shown in FIG. 3, WB passes through the buffers 12 of the TDL structure and is delayed by data period. w Multiplied by and input to k hidden nodes 14. At the same time, the past output value of each hidden node 14 Y _k Is a value calculated by determining by each determinant 10 as a determined value. Is output. This determined value Is multiplied by the weight to be input to the hidden node 14 again. Accordingly, the output value of each hidden node 14 is as shown in Equation 1 below.

Here, sets I and L represent a set of external inputs and a set of past output values of nodes, respectively. Vector w Is a weight obtained by the real-time recursive learning method proposed by William. Denotes the result of the decision made by the determiner 10 to {± 1}. The desired output value is used in the training mode and the output value of each hidden node 14 in the decision-directed mode. Y ₁ , Y ₂ ,.. , Y _k Use the value determined by. This is expressed as the following equation.

In the decision feedback recursive neural equalizer of the present invention, the learning method differs only in calculating the output value of each node, and the rest of the processes are substantially the same as those of the conventional recursive neural equalizer.

In detail, in step 1, the output values of the current hidden nodes 14 are calculated using Equation 1. The instantaneous error value is then calculated in step 2. In this case, the calculated instantaneous error value is represented as follows.

The sensitivity function is then calculated in step 3. This is expressed as the following equation.

Here, f (·) is an active function of the neural network and modeled as a tanh function. In other words, to be.

Finally, the weight is corrected using the instantaneous error value calculated in step 4. This is expressed as the following equation.

Where α is any positive real number.

Conventional recursive neural equalizers perform learning with only one desired output value, but the proposed decision feedback recursive neural equalizer gives the desired output values to all hidden nodes 14 of the hidden layer, and all of them participate in learning to apply more constraints. In addition to learning, it enables stable learning of the system. Hereinafter, the performance of the decision feedback recursive neural network equalizer according to the present invention will be described in terms of MSE.

First, when the binary signal sequence {l _k } consisting of {+1, -1} is transmitted, the input vector of the equalizer passing through the channel and mixed with white Gaussian noise {n _k } is called {X _k }. Speaking statistically independent of heat, the output values of the conventional recursive neural equalizer and the decision feedback recursive neural network equalizer according to the present invention are as follows.

Each MSE is represented by the following equation.

By using the above equation to obtain each MSE and the difference can be obtained as follows.

The result is squared over all terms, so the value on the right side is always negative. Accordingly, it can be seen that the MSE values of the decision feedback recursive neural network equalizer (DFRNE) and the recursive neural network equalizer (RNE) have the following relationship.

MSE _DFRNE ≤MSE _RNE

As such, it can be seen that the MSE of the decision feedback recursive neural network equalizer according to the present invention is always equal to or smaller than that of the conventional recursive neural network equalizer.

In conclusion, the decision-feedback recursive neural network equalizer according to the present invention can reduce the amount of computation by using the IIR form as in the conventional recursive neural network structure, and make a decision every hour to obtain the effect of extracting past error values in the system. Past error values in the system eliminate the risk of system instability and divergence from occasional bouncing values.

As described above, according to the equalization apparatus and method using the decision feedback recursive neural network according to the present invention, the interference between symbols and nonlinear distortion occurring in the magnetic recording / reproducing system can be effectively removed to restore the distorted signal. Can be. In addition, the equalization device using the decision feedback recursive neural network of the present invention can be easily implemented in real hardware because of its simple structure as a feature of the IIR structure. Further, the present invention can be applied to a magnetic recording medium or all communication channels with particularly high nonlinear distortion, thereby obtaining excellent performance.

On the other hand, it will be appreciated by those skilled in the art that various changes and modifications can be made without departing from the spirit of the present invention. Therefore, the technical scope of the present invention should not be limited to the contents described in the detailed description of the specification but should be defined by the claims.

Claims (7)

At least one delay means for delaying the input signal supplied from the channel in data period units; Arithmetic means connected to the delay line by the number of delay lines so as to function-operate the delayed signal supplied from the delay line and the feedback signal from the output end; And a decision value detecting means for detecting a digital signal by returning the output signals of the calculating means to an absolute value, and feeding it back to the calculating means. The method of claim 1, And said calculating means adds up the past output values with weights. The method of claim 1, The set value of the input signal supplied from the channel is determined as I, the set of signals supplied from the delay line is L, the weight is ω and {± 1}. Speaking of Output of said computing means S _i (t) Is Equalizer using a decision feedback recursive neural network, characterized in that the equation. At least one delay means for delaying the input signal supplied from the channel in data period units; Arithmetic means connected to the delay line by the number of delay lines so as to function-operate the delayed signal supplied from the delay line and the feedback signal from the output end; Determination value detection means for detecting a digital signal by replacing output signals of the calculation means with an absolute value and feeding the feedback signal back to the calculation means; And a weight correction means for correcting a weight in accordance with an output signal of said calculating means and an output signal of said detecting means. The method of claim 4, wherein If the instantaneous error value is e _k (t), the sensitivity function is p ^k (t), and any positive real number is α, the weight Δω (t) corrected by the weight correction means is Equalizer using a decision feedback recursive neural network, characterized in that the equation. Delaying the input signal supplied from the channel in data period units; Functionally calculating the delayed signal supplied from the delay line and the feedback signal from the output stage; And replacing the function-operated signal with an absolute value, detecting a digital signal, and feeding it back to the input of the function operation. Delaying the input signal supplied from the channel in data period units; Functionally calculating the delayed signal supplied from the delay line and the feedback signal from the output stage; Substituting the function-operated signal with an absolute value to detect a digital signal and feeding it back to the input of the function operation; And correcting weights according to the function-operated signal and the digital form signal.