JP2526400B2 - Adaptive equalizer using neural network and Viterbi decoder - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive equalization technique required for transmitting digital data through a transmission line having an unspecified amplitude / delay distortion, and a neural network.
It relates to an adaptive equalizer using a network and a Viterbi decoder .

[0002]

2. Description of the Related Art A telephone line is composed of many subscriber lines, exchanges, and relay systems, and it is difficult to make the amplitude and delay characteristics of each call uniform for all calls.
In addition, wireless relay, broadcasting, mobile
In mobile or wireless data terminal communication, it is unavoidable that the transmission path includes temporal fluctuation factors such as fluctuations in the characteristics of the propagation medium and movements of objects.

Therefore, when digital data or an image signal is transmitted through these transmission lines, intersymbol interference, multipath distortion, etc. occur, which causes deterioration of transmission quality such as generation of error bits and ghosts or deterioration of throughput. Therefore, we will automatically compensate for these distortions .
In order, the adaptive equalizer that has been conventionally used.

Conventionally, adaptive equalizers have been used for automatic compensation of transmission line characteristics in the fields of digital data transmission through telephone lines, digital radio relay, television broadcasting and the like. Adaptive equalizers are roughly divided into linear transversal equalizers and decision feedback equalizers. In the linear transversal equalizer, the known training signal is transmitted prior to the actual transmission of the message data or periodically thereafter, and the difference between the known training signal and the received signal sequence, that is, the error signal is generated. Then, the tap gain (weight) of the transversal filter is controlled so as to gradually reduce the squared error. This weight adjustment method is called the least mean square algorithm and maximizes the signal-to-distortion ratio of the output. In some cases, an algorithm (Zero-Forcing Algorithm) that minimizes peak distortion is used. The characteristics are not changed during the message data transmission period and the weights are not adjusted. In the decision feedback equalizer, the decision value of the data is regarded as correct even during the message data transmission period, and the weight is controlled using the difference from the decision value as an error signal. Therefore, the decision feedback equalizer can track the gradual fluctuation of the line characteristics, but on the other hand, if an error occurs in the decision value, error propagation may occur.

An adaptive equalizer using a neural network with a multilayer perceptron configuration has also been proposed. In the neural network, the sum of products element of the transversal filter is made multi-element, and the non-linear element is added to the output of each element,
It can be said that they are made up of multiple layers. However, while the inputs of the transversal filter are limited to time series data, neural networks generally have a large number of inputs, which may be independent of each other.

It is expected that mobile / portable communication in an urban area or indoor wireless data terminal communication will be used under more severe propagation conditions, that is, in an environment where the received signal is subjected to Rayleigh fading and further intersymbol interference is received. In a Rayleigh fading environment, a large number of interfering waves with the same amplitude as the direct wave and with a uniformly distributed phase are superimposed on the direct wave. In a frequency-selective fading environment, intersymbol interference such as adjacent and next-adjacent symbols is added at random phases. Under such severe propagation conditions, the output of the decision feedback equalizer, that is, the decision value is not always correct. The adaptive equalizer technology that overcomes the Rayleigh fading environment or the frequency selective fading environment and enables high-efficiency and high-quality data transmission has not yet been established.

[0007]

SUMMARY OF THE INVENTION The present invention has been made in view of the above,
Its purpose is to provide an adaptive equalizer which allows a high efficiency and high quality digital data transmission even under poor propagation environment.

[0008]

A convolutional coding Viterbi decoding method is a typical technique for realizing high-quality digital data transmission on a transmission line with poor line quality. The message data is convolutionally encoded and then transmitted, and the receiving side performs decoding based on the Viterbi algorithm. If a convolutional code with coding rate R = 1/2 and constraint length K = 5 is used, it is 4.2 dB (bit error rate;
BER = 1 × 10 ^-4 ), 4.6 dB (BER = 1 × 10 ^-5)
It is known that there is a coding gain of

A neutral network according to the present invention
As shown in FIG. 1, the adaptive equalizer using the Viterbi decoder is a tapped delay line or shift register (1) that stores the demodulated output value of the received signal while sequentially shifting it at each symbol interval, a neural network. (2), Viterbi decoder (3), and reconvolutional encoder (4). The received demodulated signal is input to the Viterbi decoder (3) after passing through the forward direction calculation unit (2-1) of the neural network. The output of (3) becomes the decoded output, while the output of (4) becomes the input and is re-encoded. The output of (4) serves as a reference signal for weight adjustment. The demodulation output value is given a time delay corresponding to the time delay of the Viterbi decoder (3) by one side (1) and becomes the input of the forward direction calculation unit (2-2) of the neural network. An error signal is obtained by taking the difference between the output of (2-2) and the output of (4). Based on this error signal, the weight is adjusted in the backward calculation unit (2-3) of the neural network. Using this weight, the signal to the Viterbi decoder is calculated in (2-1). This cycle operation is repeated each time each symbol is received. A backpropagation method (Backpropagation) generally used in neural networks is used as a weight adjustment method performed in the backward calculation unit (2-3).
n Algorithm) is available.

[0010]

[Embodiment] A neutral network according to the present invention
The performance of the adaptive equalizer using the Viterbi compounder was evaluated by computer simulation.

Rayleigh fading and frequency selective fading are applied to the transmission path. If the angle between the moving direction of the moving body and the traveling direction of the radio wave is θ _i , the received wave undergoes a Doppler shift of f _d = (V / λ) cos θ _i . Where V is the moving speed of the moving body and λ is the wavelength of the radio wave. V / λ is called the maximum Doppler frequency. Frequency is the symbol rate of the transmission line R
_When normalized with _s , the normalized maximum Doppler frequency is expressed as F _d = (V / λ) / R _s . Rayleigh fading has θ _i from 0 to 2
Synthesis of N waves uniformly distributed in π range It is represented by. Here, T is the time normalized by the symbol interval (1 / R _s ) and is assumed to take an integer value. φ (t)
Is the phase function corresponding to the modulated signal, α _i is the initial phase, F ₀
Is the normalized carrier frequency, j is the imaginary unit, and the coefficient 1 / N ^1/2 is for normalizing the signal power to 1. i = 1
Is regarded as a direct wave and θ ₁ = 0. Equivalent low range display when F ₀ → 0 Is obtained. When N → ∞, the real and imaginary parts of z (t) can be regarded as the sum of a large number of independent random variables with a mean value of 0. Therefore, each is a Gaussian random variable with a mean value of 0. Therefore, the amplitude (absolute value) of z (t) forms a Rayleigh distribution. 5-1 and 5-2 in FIG. 2 are examples of amplitude (absolute value) and phase time waveforms of z (T), and 6 in FIG. 2 is an example of cumulative distribution of amplitude (absolute value). ． The time waveform of the phase (5-2) is multiplied to display the modulation component φ (T). The cumulative distribution (6) almost agrees with the theoretical value of the Rayleigh distribution. In a frequency-selective fading environment, further delayed waves z _d (T) = z (T-1), z _2d (T) = z (T
-2), ..., etc. are added to z (T). Where z _d (T)
Limited to only. The received signal is expressed as u (T) = [1 / (1 + γ)] z (t) + [γ / (1 + γ)] z _d (T) γ [dB] = -20 log ₁₀ γ ． However, all θ _i of the delayed wave are uniformly random.

FIG. 3 is a block diagram showing the details of the computer simulation carried out. In the embodiment, the coding rate R =
1/2, constraint length K = 5, generator polynomial g ₁ = (10111), g ₂ =
The convolutional code of (11001) is used. Quadrature phase modulation (QPSK) of carrier wave by two outputs of encoder at transmission side
And send. U (T) undergoes fading distortion on the transmission line
Is received. u (T) is orthogonally demodulated on the receiving side, and u
The real and imaginary parts of (t) are in-phase components (I-channel;
I-ch) signal and quadrature component (Q-channel; Q-ch) signal. In addition, pseudo Gaussian random numbers are added as additive noise to each channel independently. Normalized maximum Doppler frequency F _d is 0.000
We set 01 and set N = 1 wave or N = 5 wave. The neural network has a two-layer structure, and the numbers of elements in the first and second layers are 4 and 2, respectively. The number of input terminals is Ic
The number of each of h and Q-ch is 5 and the total is 10. Learning coefficient
μ was 0.2 and the inertia coefficient η was 0.8.

[0013]

The results of the computer simulation carried out under the above conditions are shown in Table 1. For comparison, a. Normal unequalized uncoded BPSK, b. QPSK using only the ordinary Viterbi decoding method, c. The results of BPSK using only the neural network are also shown. d. The columns indicate the neutral network and the Viterbi compound according to the present invention.
This is the result of QPSK with an adaptive equalizer using a combiner. Each method is compared under exactly the same propagation conditions.
One run consists of 49500 symbol (bit) message data following a 500 symbol training sequence. Each run was executed 100 times, and 100 runs were classified according to the number of error bits contained in 49500 bits of each run. For example, N =
1 wave, E _b / N. When (signal energy to noise power density ratio per bit) = 8 dB and γ = ∞ dB, a. In unequalized uncoded BPSK, it is shown that 90 out of 100 runs contained 6 to 50 error bits. The actual number of error bits at this time is 9.6 on average each time.
They are in good agreement with the theoretical value. <E _b > is the average signal energy per bit.

From the computer simulation results shown in Table 1, the neutral network and the Viterbi according to the present invention are shown.
The QPSK system with an adaptive equalizer using a decoder is
When only = 1 wave and γ = ∞ dB, in other words, when only the direct wave is received and there is no interference, it is inferior to the normal Viterbi decoding method, but there is N = 5 wave interference and further γ = 20 dB adjacent symbol interference. 94 times (<Eb / NO> = 3) out of 100 runs even in a poor fading environment
It was shown that error-free transmission is possible from 4 dB) to 90 times (24 dB).

[Table 1]

[Brief description of drawings]

FIG. 1 is a neutral network and a network according to the present invention.
FIG. 7 is a functional block diagram showing a schematic configuration of an adaptive equalizer using a tabby compound device .

FIG. 2 is a diagram showing an example of transmission path characteristics in a computer simulation carried out for performance evaluation.

FIG. 3 is a block diagram showing details of a computer simulation.

[Explanation of symbols]

1 Delay line or shift register with tap 2 Neural network 2-1 Forward calculation unit of neural network 2-2 Reverse calculation unit of neural network 3 Viterbi decoder 4 Reconvolutional encoder 5-1 N wave synthesis Waveform amplitude (absolute value) time waveform example 5-2 Phase time waveform (multiplied to remove modulation component) 6N N-wave composite wave amplitude (absolute value) accumulation Distribution example 7 Transmission line 8 Transmission side convolutional encoder

Continuation of the front page (51) Int.Cl. ⁶ Identification number Office reference number FI technical display location H04L 27/01 H04L 27/00 K

Claims (1)

(57) [Claims] 1. A tapped delay line or to store while sequentially shifting the demodulated output value for each symbol interval of the received signal and the shift register, the taps with a delay line or the Shift register
A first neural network forward calculation unit for inputting the demodulated output value from data, and Viterbi decoder which receives the output of the first neural network forward calculation unit, identical to the transmission-side convolutional encoder With the functional characteristics of
Both inputs a re-convolutional encoder that receives the output of the Viterbi decoder and a demodulated output value that is delayed by a tapped delay line or shift register corresponding to the time delay of the Viterbi decoder. Second neural network forward calculation unit and the second neutral net
Output of the work forward calculation unit and output of the reconvolutional encoder
And a neural network reverse calculation unit which receives the Sadea error signal between force, the reverse calculation unit, and an output of the second neural network forward calculation unit output and re convolutional encoder Is the difference
Error error backpropagation based on the error signal
An adaptive equalizer characterized by adjusting and controlling the weight of a neural network.