JPH07283741A - Viterbi decoding method and viterbi decoder - Google Patents

Abstract

PURPOSE:To perform an effective decoding operation even to a Markov information source by deciding the transmission signal with addition of an offset value correspond ing to the state transition probability of the signal series set based on the Markovian nature of the information source. CONSTITUTION:In Viterbic decoding mode, the offset value corresponding to the state transition probability of the signal series set based on the Markovian nature of a signal source that is obtained by an offset value generating means is added to the inter-code distance calculated by an inter-code distance calculation means 11 between the transmission and reception signals. Then a transmission signal is decided by a deciding means 12. That is, the state of the register constructing a coder is defined as the internal state of a Markov information source and the transition of this state is observed. Then the transition probability is applied as the internal state when the maximum likelihood coding is carried out. In regard of a voice signal, no redundancy is recognized at the output side of the coder and the approximately same occurrence probability is secured between 0 and 1. However the redundancy remains in the time axis direction owing to the Markov information source. Thus the noise influences are corrected by the redundancy of the information source so that the correct Viterbi decoding can be effectively carried out.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a Viterbi decoding method and a Viterbi decoder used in a communication system.

For example, in a communication system such as a satellite communication system or a mobile communication system, it is necessary to perform error correction in order to improve the reliability of data to be transmitted. As the most effective error correction method, the maximum likelihood decoding method represented by the Viterbi decoding method is known, but this assumes that the information source is random and However, it cannot be said that efficient decoding is performed because the redundancy is not used.

Therefore, it is necessary to realize more efficient decoding for such an information source, for example, an information source having a correlation with a signal sequence (Markov information source).

[0004]

2. Description of the Related Art In the title of the invention, "Maximum-Likelihood Decoding Method and Maximum-Likelihood Decoder for Communication Systems", filed by the applicant on November 15, 1991, if the information source has redundancy, It was shown that the use of redundancy improves the decoding characteristics of a normal Viterbi decoder that does not consider redundancy.

In this method, the redundancy caused by the different appearance probabilities of binary digital signals of 0 and 1 is determined by the branch metric (that is, the likelihood function) of the Viterbi decoding method.
The maximum posterior probability decoding method is realized by using it when calculating

That is, the code system generated from the information source (for example,
The transmission signal is determined by adding an offset value corresponding to the occurrence probability of 0, 1).

Therefore, the more different the occurrence probabilities are, the more efficient decoding (correct decoding even if there is much noise) corresponding to the redundancy of the information source can be performed.

[0008]

As described above, the decoding method described in "Maximum-Likelihood Decoding Method and Maximum-Likelihood Decoder of Communication System" is large if the difference between the appearance probability of 0 and the appearance probability of 1 is large. Since the redundancy remains, the decoding can be performed more effectively than the normal Viterbi decoder and the performance is improved.

[0009] For example, in the case of a voice signal, a Gaussian distribution is obtained when the appearance probabilities for various amplitude levels of the voice signal are obtained, but 0 (or 1) is assigned to many amplitude levels with a high appearance probability, and a low amplitude level is assigned. By assigning a small number, the appearance probabilities of 0 and 1 are different, and a large effect can be obtained in the Viterbi decoder by using the redundancy of the audio signal.

However, in order to reduce the transmission bandwidth, a voice encoder is provided, and for example, the bit rate is changed from 64 Kb / s to 8 Kb / s.
When the compression is performed, since the redundancy is used for this compression, the appearance probabilities of 0 and 1 become almost equal on the output side of the speech coder.

In such a state, the transmission signal must be determined only by the inter-code distance between the transmission signal and the reception signal, which is shown in the above "maximum likelihood decoding method and maximum likelihood decoder of communication system". Even if such a decoder is used, the same decoding characteristic as that of a normal Viterbi decoder that does not consider redundancy is obtained.

That is, the above decoder does not perform efficient decoding on an information source (Markov information source) having the same appearance probability of the signal itself but having a correlation with the signal sequence like a voice signal. There was a problem called.

[0013]

FIG. 1 is a block diagram showing the principle of the first and second aspects of the present invention. In the figure, 2 is an offset value generating means for transmitting an offset value corresponding to the state transition probability of a signal sequence based on the Markov property of an information source, and 11 is an intersymbol distance for obtaining an intersymbol distance between a received signal and a transmitted signal. The calculating means 12 is a judging means for judging the transmission signal by adding an offset value to the inter-code distance.

According to the first aspect of the present invention, at the time of Viterbi decoding, an offset value corresponding to a state transition probability of a signal sequence based on the Markov property of an information source is added to the inter-code distance between the transmission signal and the reception signal for transmission. The signal is judged.

According to a second aspect of the present invention, the Viterbi decoder comprises an inter-code distance calculating means, an offset value generating means and a judging means. According to a third aspect of the present invention, the offset value generating means generates a value corresponding to −2σ ² log _e p (s | S _a ) as an offset value.

[0016]

First, the nth-order Markov information source is such that the appearance probability of the current signal (0, 1) depends on at most the past n signals, and does not depend on older signals. Is. Note that past signal sequences s ₀ , s ₁ , ..., s
_The collection of _n-1 is regarded as inside information of the information source. In other words, the fact that a signal sequence of s ₀ , s ₁ , ..., S _n-1 has been generated in the past means that the current state S _a of the information source is S _a = s ₀ , s ₁ ,. It can be regarded as s _n-1 .

In such a current state, the conditional probability qs _a (s _n ) that the signal s _n will appear next is qs _a (s _n ) = p (s _n │ s ₀ , s ₁ , ... , s _n-1 ) (1) can be written.

Further, after the appearance of the signal s _n , a new s _n is added to the signal sequence, the oldest signal s ₀ is deleted, and the current state S _b of the information source is S _b = s ₁ , s _2,..., a transition to s _n.

Here, the state S _a transits to the state S _{b because the} signal s _n appears, and the state S _a changes to the state S _a.
The probability of transition to S _b , that is, the transition probability p _ab is equal to the probability of occurrence of the signal s _n qs _a (s _n ).

Further, in the case of a voice signal, as described above, the redundancy disappears on the output side of the encoder and the appearance probabilities of 0 and 1 are almost equal, but in the time axis direction due to the Markov information source. Redundancy remains.

On the other hand, the Viterbi decoding method is an algorithm for performing maximum likelihood decoding of a convolutional code (coding rate = R, constraint length = K). In the first and second aspects of the present invention, the state transition of the shift register constituting the encoder is used as the internal state of the Markov information source, and its state transition is observed, and maximum likelihood decoding is performed. Apply.

Accordingly, when the constraint length K> the Markov information source order n, the following equation is used as the branch metric function b (s | S _in ). b (s | S _in ) ＝ d ² (s, r) −2σ ² log _e p (s | S _a ) (2) where d ² (s, r): the received signal r and the transmitted signal s in Euclidean distance between section affected by noise S _in: the internal state of the shift register sigma ^2: noise variance p (s | S _a): the probability current state is s appears when S _a Note, constraint length When using all K, S _in = S _a , but
S _a is included _in S _in when partially used.

That is, by using the equation (2), the first term on the right side affected by noise is corrected by the second term on the right side using the redundancy of the information source. It becomes possible to perform correct decoding and efficient decoding is performed.

[0024]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 2 is an example of a digital communication system using a convolutional code, FIG. 3 is an example of a convolutional encoder in FIG. 2, and FIG.
(b) is an operation explanatory view of (a). Further, FIG. 4 is an example of an explanatory diagram of the Viterbi decoding method, FIG. 5 is a configuration diagram of main parts of the Viterbi decoder in FIG. 2, and FIG. 6 is a configuration diagram of main parts of the first, second and third embodiments of the present invention. Is.

Here, the difference units 111a and 111b, the squarers 112a and 112b, the adder 113, and the convolutional encoder 114 in FIG. The multiplier 23 and the multiplier 24 are components of the offset value generating means 2.
The adders 121 and 122 and the old path metric value 123 are the determination means 12
Is a constituent part of.

2 to 6 will be described below. In FIG.
QPSK is explained as 4-phase PSK. First, the digital communication system shown in FIG. 2 uses a convolutional code with a coding rate R = 1/2 and a constraint length K = 3 as an error correction code in the four-phase PSK method.
The information source 30 is a second-order Markov process, and the internal state of the shift register forming the convolutional encoder 31 is S _xy (x, y is 0, 1
, And the state transition probabilities of this internal state are p (0 | S ₀₀ ) = 3/4, p (1 | S ₀₀ ) = 1/4 p (0 | S ₀₁ ) = 1/3 , p (1 ｜ S ₀₁ ) ＝ 2/3 p (0 ｜ S ₁₀ ) ＝ 2/3, p (1 ｜ S ₁₀ ) ＝ 1/3 p (0 ｜ S ₁₁ ) = 1/4, p (1 | S ₁₁ ) = 3/4.

That is, the internal state of the shift register is S _00.
, The probability that the input data is 0 is 3/4, the probability that 1 is 1/4
Therefore, the former probability is three times larger than the latter probability.

Now, the binary data sent from the information source 30 is converted by the convolutional encoder 31 of R = 1/2 and K = 3 into I,
Converted to 2-channel Q data and 4-phase PSK modulator 32
Added to. The 4-phase PSK modulator 32 generates a 4-phase PSK wave by using the data of 2 channels and sends it to the transmission path 14, so that 2-bit data is transmitted by 1 symbol.

On the other hand, since noise of the transmission line is added to the four-phase PSK wave while propagating through the transmission line and is input to the four-phase PSK demodulator 35 on the receiving side, the demodulator 35 receives the input four-phase PSK wave. Is demodulated and two channels of demodulated outputs (I ', Q'in the figure) are extracted and sent to the Viterbi decoder 1.

Therefore, the Viterbi decoder uses the transition probability p (s | S
_xy ) table 21 and the output of the line quality detection portion 22 to perform decoding in consideration of the Markov property of the information source and take out the decoded data (details will be described later).

In FIG. 3, coding rate R = 1/2, constraint length K =
As shown in (a), the convolutional encoder of 3 has a two-stage shift register (x and y portions in the figure) 123 and an exclusive OR of two inputs 121 and an exclusive OR of three inputs. The shift register is considered as a state, and the state transition diagram is written as shown in (b).

Incidentally, the state surrounded by the circle in (b) 00 → state
The symbol "1/11" near 01 is 1, 2 as I, Q output if data 1 is input when the state of shift register x, y is 00.
It indicates that the state of the shift registers x and y transits to 01 by sending out 1. Hereinafter, the same applies to other states.

Further, FIG. 4 shows Viterbi decoding (coding rate R = 1 /
2, constraint length K = 3) This is a part of the trellis diagram for explaining the state transition of Fig. 3 (b) with the horizontal axis as the time axis. The solid line branches depend on the input data 0. A transition is shown, and a dotted line branch shows a transition by the input data 1. Also, the numbers "00", "01", "10", "11" written along the branches
Is the output of the encoder when the transition of that branch occurs.

When data 1 is input to the encoder at time 0, the initial state S ₀₀ changes to S ₀₁ at time 1 and the output (code sequence in the figure) of the encoder becomes "11". Since the received sequence at that time is "11", the number of errors is 0, so the branch metric is 0. However, when data 0 is input, the code sequence is the same as S ₀₀ at “00”, the number of errors with the reception sequence of “11” is 2, and the branch metric is 2. It should be noted that these processes are performed in the branch metric calculation part 11 of FIG.

Further, when the branch metric at the time point 2 is obtained in the same manner as above, it is 2 when S ₀₀ , 0 when S ₀₁ , and S ₁₀
Is 1 when, and 1 when S ₁₁ . Therefore, the sum of branch metrics from time 0 to time 2 (called path metric) is 4 for S ₀₀ → S ₀₀ → S ₀₀ , 2 for S ₀₀ → S ₀₀ → S ₀₁ , and S ₀₀ → S ₀₁ → S. It becomes 1 at ₁₁ .

[0036] In addition, the S ₀₀ of point 3, and S ₀₀ of point 2
The branch from S ₁₀ enters, but the branch metric of this is 1 for S _{00 and} 1 for S ₁₀ , so
The path metric of S ₀₀ → S ₀₀ → S ₀₀ → S ₀₀ is 5, S ₀₀ → S
The path metric of ₀₁ → S ₁₀ → S ₀₀ is 2, the latter is smaller than the former, and the latter is the surviving path.

Similarly, when the survivor paths entering each state are obtained for each time point, four survivor paths remain at the final point (when the encoder output is 2).
The path with the smallest path metric (maximum likelihood path) is selected from these, and the information sequence corresponding to the selected path is output as the result of maximum likelihood decoding of the received sequence.

Note that these operations are performed by the ACS (add-comp
are-select) part. Further, the path memory 14 stores the surviving path, and the path metric memory 13 stores the path metric of the surviving path in each state.

Here, the above description of Viterbi decoding using FIG. 4 is a case where the Markov property of the information source is not taken into consideration, but a decoding method considering the Markov property will be described below. For example, as described above, the state S ₀₀ at the time point 3 in FIG. 4 is more likely to be the case where the data 0 is input to the state S ₀₀ at the time point 2 and the case where the data 0 is input to the state S ₁₀ . Select. I, Q of the encoder at this time
In the output, the former is (0,0) = s ₁ and the latter is (1,1) = s ₂ .

Therefore, as shown in equation (2), the sum of the Euclidean distance between the actual received signal r and the transmitted signals s ₁ and s ₂ and the offset value according to the state transition probability is used as the branch metric value. If the state transition probability set in step 2 is used, b (0 | S ₀₀ ) ＝ d ² (s ₁ , r) -2σ ² log _e p (0 ｜ S ₀₀ ) ＝ d ² (s ₁ , r) -2σ ² log _e 3/4 (3) b (0 | S ₁₀ ) ＝ d ² (s ₂ , r) -2σ ² log _e p (0 ｜ S ₁₀ ) ＝ d ² (s ₂ , r) -2σ ² log _e It becomes 2/3 (4).

Similar calculations are performed for the other paths to determine the maximum likelihood path. Next, FIG. 6 shows the path metric calculator in the state S ₀₀ at the time point 3, and the operation thereof will be described below.

The difference units 111a and 111b shown in FIG.
The demodulation outputs I ', Q'from the phase PSK demodulator 35 and the outputs I', Q'of the convolutional encoder 114 when the states S ₀₀ and 0 at time 2 are added
When Q and are input, the differences (I-I ') and (Q-Q') are calculated and added to the corresponding squarers 23 and 25. Therefore,
These squarers calculate the difference between the I and Q components by 2
Then, the adder 113 adds these to obtain the square of the distance between the demodulator output and the encoder output. As a result, the Euclidean distance of one item on the right side of equation (3) is obtained.

Here, the line quality detector 22 determines the S / N of the signal.
The noise variance σ ² is output using the (E _b / N ₀ ) ratio.
In the case of SK wave, σ ² = P ² / (2 × 10 ^A ) as is well known. However, A = [(E _b / N ₀ ) −3] / 10, P: the amplitude of the input signal.

Note that the noise variance σ is derived from the above equation using the lower limit of operation (E _b / N ₀ ) at the time of system design when a line quality detector is provided or when no means for detection is available. You can use one.

Further, it is assumed that the above state transition probability is stored in a memory (not shown). On the other hand, the line quality detector 22
Sends the noise variance σ ² corresponding to the S / N (E _b / N ₀ ) ratio (dB) of the applied signal to the multiplier 24 via the doubler 23. P (0 ｜ S ₀₀ ) read from
Since 3/4 of the state transition probability of is also applied, the two items on the right side of the above equation (2) are obtained from the multiplier 24.

Since the multiplier 24 sends the multiplication result to the adder 121, the adder 121 adds the offset value to the Euclidean distance between the demodulator output and the encoder output, and the branch between time points 2 and 3 is performed. The metric value is obtained. This value is obtained by correcting one item on the right side of the equation (2) affected by noise and correcting the effect by two items on the right side.

Further, the adder 122 adds the old path metric value (S ₀₀ ) up to time point 2 (this value is also corrected for the influence of noise) and the above branch metric value to add the path metric up to time point 3. The value is obtained.

The lower half of the chain line in FIG. 6 is the time point 2 when the data 0 is input to the state S ₁₀ at the time point 2.
This is a path metric calculation unit that obtains a branch metric value between time point 3 and time point 3 and a path metric up to time point 3. Since the operation of this part is the same as when data 0 is input to the state S ₀₀ , it is omitted.

That is, for the Euclidean distance between the transmission signal and the reception signal affected by noise, the effect of noise is corrected by using the offset value according to the state transition probability, so that the Markov information source can be more accurately corrected. Decoding is possible, and efficient decoding can be performed.

[0050]

As described in detail above, according to the present invention, there is an effect that more efficient decoding can be realized for a Markov information source.

[Brief description of drawings]

FIG. 1 is a principle configuration diagram of first and second aspects of the present invention.

FIG. 2 is an example of a digital communication system explanatory diagram using a convolutional code.

FIG. 3 is an example of an explanatory diagram of a convolutional encoder in FIG.
Is a block diagram of an encoder main part, and (b) is an operation explanatory view of (a).

FIG. 4 is an example of a Viterbi decoding method explanatory diagram.

5 is a configuration diagram of the main part of the Viterbi decoder in FIG.

FIG. 6 is a main part configuration diagram of an embodiment of the first, second and third present inventions.

[Explanation of symbols]

 2 offset generation means 11 intersymbol distance calculation means 12 determination means

Claims (3)

[Claims] 1. A Viterbi decoding method for Viterbi decoding a received signal, which corresponds to an inter-code distance between a transmitted signal and a received signal at the time of Viterbi decoding and a state transition probability of a signal sequence based on the Marco property of an information source. A Viterbi decoding method characterized in that a transmission signal is determined by adding an offset value. 2. An intersymbol distance calculating means (11) for obtaining an intersymbol distance between a received signal and a transmitted signal, and an offset value corresponding to a state transition probability of a signal sequence based on the Markov property of an information source. A Viterbi decoder comprising an offset value generating means (2) and a judging means (12) for judging a transmission signal by adding an offset value to an inter-code distance. 3. The offset value generating means is: −2σ ² log _e p (s | S _a ), where p (s | S _a ): probability σ ² that state s appears when the current state is S _a The Viterbi decoder according to claim 2, having a function of generating a value corresponding to the variance of noise as an offset value.