FI101759B - Changed Viterbide coding method - Google Patents

Description

101759

Modified viterbide decoding method - Förändrad viterbidekoderingsmetod

The invention relates to a method for decoding a convolutionally encoded digital bit sequence.

Convolutional coding and its viterbide decoding used as its inverse operation are well-known methods for improving the reliability of digital data transmission. The transmitting device has a convolutional coding block, i.e. a 10 convolutional encoder, which receives a certain bit sequence for processing. In encoding, that block produces a new bit sequence, the shape of which depends on both the original bit sequence and the structure of the encoder, and in which the effect of a particular original bit is distributed over several consecutive bits or bit combinations. The viterbide decoding performed in the receiver belongs to the so-called maximum likelihood methods and in which the receiver decoder tends to convert the convolutionally encoded bit sequence back to the original sequence. Transmission errors may have occurred between the transmitting and receiving devices, so the viterbi decoder selects the bit or bit sequence that most likely corresponds to the portion of the encoded sequence it receives as decoded information. In order to find out the background of the invention, a simple example case is described below, in which the so-called convolutional coding. the constraint length is three.

Figure 1 shows a convolutional encoder which receives one bit at a time as input data and forms a two-bit output data combination from it and the contents of the registers it contains. Registers 12 and 14 each represent one-bit delay combiner ports 16 and 18 of the exclusive OR type (XOR). Initially, the contents of both registers 12 and 14 are zero. For example, encoding the bit sequence 011011 with the encoder of Figure 1 results in 30 0 bit pairs (00), (11), (10), (11), (01) and (10), which can be interpreted as an output sequence. ) or vector 001110110110. In many applications, the output data sequence is punctured, i.e. the bits according to the known pattern, i.e. the puncturing vector, are removed from it. If the output data sequence of the example is punctured with the vector 111111101010, the final output data sequence is a bit string 35 001110101. The puncturing vector is used by removing from the original (unpunctured) bit sequence the bits for which the puncturing vector has zero.

101759 2

A receiver that receives and indicates a signal transmitted over a transmission path may apply either absolute or sliding detection. Hard decision detection means that the detector part of the receiver interprets each received bit as either one or zero. In soft decision de-5 tection, the receiver detector portion produces, for each received bit, a sample having a plurality of bits corresponding to the probability that the received bit is zero or one. For example, if a sample has four bits, it can express 16 different probability values, which are usually described as a certain integer range of the decimal system, which can be, for example, [-8, 7]. For example, the two-complement method can be used to describe the values of a four-bit sample 10 for said range, with the sample value "0000" corresponding to the decimal system number 0, the sample value "0111" corresponding to the decimal system number 7 and the sample value "1000" corresponding to the decimal system number -8. For the sake of brevity, the following decimal numerical values are used instead of the four-bit values of the samples.

15

According to a general embodiment, the value of the sample 7 corresponds to a situation in which the received bit is, in the opinion of the detector, with a very high probability one. Correspondingly, the sample value of -8 corresponds to a situation where the received bit is, in the opinion of the detector, very likely to be zero. Values of -1 and 0, which are located in the middle of the 20 sample intervals, mean that the detector cannot interpret the received bit as either better than zero with very high certainty. The task of the Viterbide encoder is to reconstruct the original bit sequence so that it is correct with the highest possible probability. The operation of the decoder differs somewhat depending on whether it receives bits (absolute detection) or 25 samples (sliding detection) from the detector at its input port.

In its operation, the Viterbi decoder utilizes the information that only certain allowed transitions between different states can take place in the convolutional encoder of the transmitter. Figure 2a shows the so-called a trellis map illustrating all possible states of the convolutional encoder 30 of Figure 1. Each column of the trellis map corresponds to one encoder clock cycle: during which the encoder receives one input data bit and results in a corresponding output data combination. The two-bit values in the lines of the trellis map correspond to the contents of the encoder registers, so that the left bit of each two-bit value corresponds to the contents of register 12 and the right bit corresponds to the contents of register 14. At the beginning, the contents of each of the 35 registers are zero. In the next column, the contents of register 14 are zero, but the contents of register 12 may be zero or one, depending on which bit the encoder has received as input. After this, the registers always have four possible states.

101759 3

On a trellis map, a particular bit sequence to be convolutionally encoded corresponds to an unambiguous path that starts from the left column and passes from left to right, going through exactly one state in each column. Figure 2b shows the path produced by the above-mentioned bit sequence 011011. Viterbi decoding is a process called

5 trellis search, where the decoder tries to find the right path and follow it through the entire trellis map. The search is facilitated by the knowledge that there is only a limited number of possible transitions from one particular encoder register state to the next. In the above embodiment, the encoder receives only one bit at a time as input data, and since one bit can have only two possible values, it can cause two possible transitions from each state 10. Figure 3 shows possible transitions between states in the encoder according to Figure 1. For example, if the contents of the registers are "01", where the contents of the register 12 are zero and the contents of the register 14 are 1, the next state can only be "10" or "00". Exactly two arrows representing possible transitions leave each state, whereby, according to symmetry, exactly two arrows arrive at each new state.

Figure 4a further shows the path produced by the same example sequence on the trellis map. The combination of output data produced by the convolutional encoder during each transition is added to the figure. The first bit of the example sequence is zero, whereby the convolution-20 thiocoder transitions from the state "00" (leftmost column) to the state "00" (second leftmost column) and the output data combination it produces is "00". The second bit is one, in which case the convolutional encoder switches from the state "00" (second leftmost column) to the state "10" (third leftmost column) and the output data combination produced by it is "11". Similar markings continue until the end of the path 25. In Figure 4b, alongside the right path, a second, arbitrary path is drawn with dashed arrows, which has nothing to do with the original bit sequence but in which all transitions are allowed transitions according to Figure 3. Corresponding source data combinations are also indicated by arrows describing arbitrary path transitions.

30 ·: Figure 4c corresponds to Figure 4b in that the output data combinations corresponding to the transitions have been replaced by a certain metric, i.e. a measure, which indicates how many correct bits there are in the output data combination compared to the original. On the right path, the metric for each transition is, of course, 2 because all the bits are correct. On an arbitrary 35a path, the measure of a given offset may be 0, 1, or 2, depending on whether the output data combination is completely incorrect, half-correct, or completely correct. Even with an arbitrary path, a given offset can randomly become a metric 2. In decoding, the aggregate metric of a particular path is considered, in which case, in the situation of Figure 4c, the total metric of the correct path is 12 and the total metric of the wrong path is 5. No path can get a larger metric than the real path. . If the structure of the convolutional encoder is chosen appropriately, the metric sum of each wrong path is smaller than that of the correct path. Thus, the search for the correct path is 5 the same thing as the search for the path that gives the largest metric sum.

At the receiver, calculating a metric for an individual transition depends on whether the expression is an absolute or a sliding expression. In absolute detection, the receiver compares the bit pair corresponding to the offset with the received bit pair as described above. It should be noted that in this context "bit pair" is an exemplary concept and consideration can easily be generalized to cases where a convolutional encoder produces single bits, bit triples, quads, etc. In the case of sliding detection, the receiver calculates a metric based on how likely the sample matches the correct bit pair. In the above case, where the samples are described between [-8.7] and 15 where -8 corresponds to a certain zero and 7 to a certain one, the sample pair may be, for example, -4.2 and the bit pair "01" corresponding to the offset. The calculation formula for the metric should be such that it gives the largest metric for the sample pair -8.7 and the smallest for the sample pair 7, -8. One such calculation formula is to add each sample separately to the description value corresponding to the certain correct bit, take an absolute value from the result, and add up the two absolute values produced by the sample pair 20. In this case, the metric of the sample pair -4.2 is | (-8) + (- 4) +1 (7) + (2) I -21. If the second bit of a bit pair is punctured in one of the transitions, the receiver calculates the metric only for the remaining bit, a procedure common to both absolute and sliding detection cases.

25 Because the metric calculation is based on the received and expressed bit pair, some offsets may obtain an incorrect metric due to bit errors. In the case of the example, if the unstructured bit sequence produced by the absolute expression is 111110111110 including three bit errors, the metric of the correct path is as shown in Fig. 4d. The error-free bit sequence was thus 001110110110. It can be seen from Figure 4d that the total of 30 metrics on the right path is 9, so the receiver cannot be sure that • this is the correct path. Another path with a larger metric may branch off from the path shown in the figure for transitions for which the associated metric value is less than 2. In Figure 4e, another possible transition and associated metrics are drawn with dashed arrows at possible branch points. 35 It should be noted that there are always only two possible transitions from one state, so there is only one other option at the junctions in addition to the correct path. As shown in Figure 4e, the first offset from the right path gives the metric 101759 5 as value 2 when the right path gives 0. At the second branch point, both offsets (both right and wrong) give the metric 1.

In Fig. 4f, at the first branch point, the upward branching path is extended according to the allowed transitions so that it joins the right path in the fourth column from the left. The wrong path cannot coincide with the correct path until the bit that caused the wrong offset has passed through the convolutional encoder shift register, i.e., successive registers 12 and 14 in Figure 1. The length of the shift register is the same as the constraint length of the convolutional encoder. The longer the shift-10 tor register in the convolutional encoder, the longer the distance the wrong path must travel separately before it can join the correct path. Even if the metric of the wrong offset at the branch point is larger than the metric of the correct one, the wrong path usually loses more of its metric sum when it travels separately than it wins at the bifurcation point (provided there are not many consecutive bit errors). In the case of Example-15, the total metric of the wrong path is 9, i.e. the same as the correct path. If the shift register is long, the path that gives the largest metric sum is most likely the correct path.

In theory, the receiver can search for the maximum path by examining all possible paths of the trellis map and calculating their metric sums. However, the number of different paths increases rapidly as a function of the number of states and columns in the trellis map, so exploring all possible paths would require an unreasonable computational capacity. Figure 5a schematically shows the partial paths (solid arrows) ending in a status column of a trellis map and the branching possibilities continuing • (dashed arrows). The metric associated with solid lines is not the metric of a single transition but the aggregate metric of the entire partial path from the beginning of the trellis map to the state shown. When searching for the maximum path, the one with the smaller total metric can be pruned out of the two partial paths entering each state. If two equal paths of equal value enter one state, the other path 30 can be pruned off arbitrarily, whereby the probability of a correct guess is •: 50%. Figure 5b shows the same state column, in which only one arrow appears in each state after pruning. Similarly, only one arrow leaves each state, because a similar qualifier has also been made in the next status column, which is not shown in the figure.

35 The prior art viterbide decoding algorithm presented so far can be summarized according to the following four steps: 101759 6 1. For each state of the current column of the trellis map (4), compute the metrics corresponding to the following transitions (two possible transitions from each state).

2. Sum the transition metrics in each state with the sum of the partial path metrics ending in the current column. The result is the metric sums of the partial paths (8 at this stage) ending in the next trellis column.

3. From the bicycles ending in the spaces in the next column, prune those with a less than 10 metrics.

4. Save the metric sum of the remaining paths to the memory along with the information from which state the saved path came from in the previous column. This information is needed to finally trace the maximum path back to the beginning of the 15 trellis map.

However, this algorithm requires a memory space equal to the length of the trellis map multiplied by the number of states in one of its columns. An improvement to the algorithm is known, which will be described next.

20

Figure 6a shows a few possible paths on a particular trellis map. The path that starts at point B and ends at point C is the correct path. Path AF is a parallel path. In addition, there are a few branches of the right path between the three rightmost columns. Because the total metric grows much faster on the right path 25 than on the other paths, the total metric of the right path at point D is likely to be significantly larger than the total metric of the parallel path at point E. The subpaths DF and EF are both off the right path, so their combined metrics are likely of the same order of magnitude. However, if the total metric of the interval BD is sufficiently 30 larger than the total metric of the interval AE, it can be quite safely assumed that when performing the pruning at point F, the path AF is pruned and the path BDF remains. Since the difference in the combined metrics of the paths BD and AE is likely to be greater the more state columns there are in this range, it can be concluded that sooner or later the path that started from A, which has traveled a long distance from the right path, will be eliminated. Figure 6b shows the situation after pruning. Thus, only the amount of memory space required for the current status column and some previous status columns is required. The preceding part of the trellis map has only one path that can be output from the viterbide decoder.

101759 7

The prior art viterbide decoding algorithms described above are relatively inefficient in a situation where the transmitted signal contains multiple consecutive bit errors in a single burst. Viewed on a trellis map, successive bit errors cause several consecutive points where the correct path may branch, leaving the algorithm 5 to handle a large number of parallel path options with aggregate metrics of the same order of magnitude.

It is an object of the present invention to provide a method for decoding a convolutionally encoded digital bit sequence, which effectively eliminates decoding errors due to burst bit error distribution and saves memory and computation capacity. It is a further object of the invention to provide a digital receiver which applies the convolutionally encoded digital bit sequence decoding method according to the invention.

The objects of the invention are achieved by utilizing the repetition of the transmitted information and dividing the viterbide decoding into periods according to the link length, between which the repetition of the transmitted information from which the final decoded bit sequence is formed can be exchanged.

The method according to the invention is characterized in that at least the first and second versions of the data sequencer to be decoded are produced and the versions are decoded substantially simultaneously to form one decoded sequence.

The invention also relates to an apparatus for decoding a convolutionally encoded digital bit sequence. The device according to the invention is characterized in that it comprises a memory for storing the decoded data sequence in at least two different decoded versions and input to a viterbid decoder, and the viterbid decoder comprises means for processing said versions simultaneously to form one decoded sequence.

The prior art viterbi decoding algorithms are characterized in that the decoder receives one bit stream as its input data and generates one bit stream as its output data. The invention differs from such a solution already at a fundamental level, because the decoder according to the invention receives at least two bit streams as input data, even though it provides only one bit stream as its output data. The bit streams as input data are identical in information content but different in error distribution. For example, a situation may be considered in which a transmitter transmits the same data file or other information sequence twice. Upon arrival at the receiver, these information sequences, i.e., repetitions of the same data file, have passed through the same transmission path, but at different times. Therefore, the locations of the bit errors caused by the transmission path do not correlate with each other between repetitions. For the sake of brevity, the following describes input data sequences with different error distributions with different color designations, whereby 5 sequences can be called different color sequences and the whole method as a color viterbid decoding method. Input data sequences that are the same color are identical in both information content and error distribution.

In the prior art description above, it has been found that the effect of the erroneous offset 10 in the trellis map extends a distance equal to the binding length of the convolutional encoder of the transmitter. The method according to the invention utilizes a different error distribution of different colors most preferably by decoding the bit streams of different colors independently of the distance of the link length, after which the metric sum of the different colored paths entering the state of each trellis column is compared. From the paths of different colors entering the state of Sa-15, the one with the largest metric sum is selected and other paths entering the state of the same state are pruned out. When, after pruning, a new independent link length viterb decoding is started for different colors, all colors are used in each mode of the column under consideration. In this case, the color on the basis of which the final output data sequence is formed may change evenly every 20 bond lengths.

The invention will now be described in more detail with reference to exemplary preferred embodiments and the accompanying figures, in which Figure 1 shows a known simple convolutional encoder, Figure 2a shows a known trellis map principle, Figure 2b shows a path in the trellis map of Figure 2a, Figure 3 shows known state transitions in Figure 1 , 30 Figures 4a to 4f show some paths and related features in the lane map of Figure 2a, Figures 5a and 5b show known path pruning in the status column of the lath map, Figures 6a and 6b show another known path pruning in the lane map, Figure 7a shows the flow of the method according to the invention in the trellis map, 35 Fig. 7b shows output bits related to state transitions to facilitate reading of Fig. 7a, Fig. 8a shows an application example of the invention, 101759 9 Fig. 8b shows a comparative example using and Fig. 9 schematically shows a receiver device in which the decoding according to the invention is applied.

5

In the above description of the prior art, reference has been made to Figures 1 to 6b, so in the following description of the invention and its preferred embodiments, reference is made mainly to Figures 7a to 9.

10 Assume that the transmitter transmits an information sequence consisting of 12 consecutive zeros. It is further assumed that the transmitter includes the convolutional encoder of Figure 1 described above in connection with the prior art description, with which it encodes said sequence before transmitting it. The encoder outputs an encoded output data sequence of 12 zero input data sequences, which is also only 15 zeros, i.e. 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00

Assume further that this encoded and transmitted sequence has passed through the transmission path 20 twice and the detector applying the absolute detection of the receiver has produced the following sequences at different iterations: 00 00 10 01 11 01 00 00 10 00 00 00 (green sequences 25 00 00 00 10 00 00 00 10 01 11 00 00 (yellow sequence).

For simplicity, only two sequences of different colors are used here. However, the invention is readily generalizable for handling multiple colors simultaneously. The Hamming distance of the green sequence, i.e. the number of bit-30 drops where it differs from the error-free sequence, is 6 and the yellow sequence is 5. There is an error burst at the beginning of the green sequence which the prior art viterbide decoding method cannot collapse. The rest of the yellow sequence also has a similar error burst.

35 In that regard, the convolutional encoder described above has been chosen as an illustrative example for its simplicity and its structure is deficient in that the long set of ones possibly contained in the original bit sequence produces consecutive zero pairs in the coded sequence in exactly the same way as the 101759 10 long set of zeros. Thus, it would be difficult for any decoder to distinguish between a zero sequence and a first sequence.

Figure 7a shows a trellis map with 13 columns and four status rows. To clarify the description, the columns are numbered 21-33. Viterbid decoding proceeds from column 21 to column 33 according to the color algorithm described above, i.e., the different colored bit sequences are decoded independently of the link length, after which the metric sum of the different colored paths entering each trellis column state is compared. Of the different colored po-10s entering the same state, the one with the largest metric sum is selected and the other paths entering the same state are pruned. When, after pruning, a new independent link length viterb decoding is started for different colors, all colors are used in each mode of the column under consideration. Columns number 21, 24, 27, 30, and 33 are for each link length. The arrow describing each offset is marked with the value of the metric corresponding to that offset -15 va so that Kx (xe [0,1,2]) corresponds to the metric kaa of the yellow bit sequence and Vy (ye [0,1,2]) corresponds to the metric of the green bit sequence y. Certain transitions are common to both colors, with metrics corresponding to both colors marked on the arrow describing the transition.

Figure 7b shows the allowable state offsets of the convolutional encoder and the associated output data bits to facilitate tracking of metrics in the offsets drawn on the trellis map. In the received bit sequence of each color, the first bit pair is "00", so that when moving from column 21 to column 22, the metric of the state transition "00" to "00" is for both colors 2 and the metric for the state transition "00" to "10" is 25 for both colors. In both colors, the second bit pair is "00", whereby from column 22 to column 23 the metric of the state transition "00" - »" 00 "is in both colors 2 and the metric of the state transition" 00 "- *" 10 "is in both colors 0. The state transition" 10 "-» "01" corresponds to the output data bits "01" and the state offset "10" to "11" corresponds to the output data bits "10", with the metric corresponding to these offsets having the color 1 of either law.

Between columns 23 and 24 there is a first qualifier. The third bit pair of the green sequence is "10" and the third bit pair of the yellow sequence is "00", where the metrics of the allowed transitions are according to the following table: 35 101759 11 offset; corresponding output green yellow bit pair metric metric "ir ->" ir; "oo" ι__2

"11" -> "01"; ”11” 0__O

"ΟΓ '->" 10 ";" OI "__ 1__1" 01 "->" 00 ";" 10 "2__1" 10 "->" 11 ";" ΙΟ "2__1" 10 "->" 01 ";" OI "__1__1

"00" -> "10"; "11" O__O

,, 00 "->" 00 ";" 00 "__ 1__2

Each state in column 24 would be preceded by two green and two yellow paths before any qualifying could be performed, which can be described as the Saijo of the states written in succession. For example, state "11" would include both green and yellow paths 00-10-11-11 and 00-00-10-11 before pruning. The metric sum of the green path 00-10-11-11 would be 2 and the metric sum of the green path 00-00-10-11 would be 4, so the internal pruning of the green color leaves only the latter path. The metric sum of the yellow path 00-10-11-11 would be 3 and the metric sum of the yellow path 00-00-10-11 would be 3, so the internal pruning of the yellow color arbitrarily leaves one of these paths 10 left. According to the invention, a pruning is also performed between the colors, leaving only the path with the largest metric sum, which is the green path 00-00-10-11, of the paths entering the state "11". In Figure 7, this is seen in such a way that only one arrow coming from the state "10" in column 23 enters the state "11" in column 24 and represents; green color (V). Similar contests can be made for all the spaces in column 24 15 and as a result only one path per space remains in the paths entering column 24, namely the green paths 00-00-10-11 and 00-00-00-10 and the yellow paths 00-00- 10-01 and 00-00-00-00.

Between columns 24 and 27, the decoding of each color initially proceeds independently, whereby the internal pruning of each color in columns 25 and 26 leaves only the one with the larger metric sum of the paths of the same color coming to a certain state in columns 25 and 26. It should be noted that the metric sum of a path of a certain color entering a given state is divided into two parts. The first part is the metric sum in the state of the previous color bar even column (here column 24) where that path originates, regardless of the color transition through which that state was entered. The second part is only the sum of the metrics related to the color under consideration from the state of said color qualifying column onwards. In column 24, the metric sum of the state "11" is 4 (calculated along the green path 00-00-10-11) according to the above description 101759 12, and a similar view gives the metric sums 3, 5 for the states "01", "10" and "00", respectively. and 6.

Consider the state "00" in column 25 as an example. In terms of yellow, it would have two possible transitions, namely from the states "01" and "00" in column 24. The output bits of the convolutional encoder corresponding to the transitions would be "10" and "00". The corresponding bit pair in the yellow bit sequence is "10", so the metric for the transition from state "01" in column 24 is 2 and the metric for the transition from state "00" is 1. A direct comparison between these would leave the first transition. However, since in column 10 24 the metric sum of the state "01" which does not depend on color is 3 and the corresponding metric sum of the state "00" is 6, the internal pruning of the yellow color compares the metric ww / wmcrar of the transition "01" -> "00" 5, the transition "00" -> "00" to a sum of 7, and selects the latter transition. A similar examination can be made for both colors in all the spaces in columns 25 and 26. In the processing of column 27, 15 internal pruning of colors is also performed first, but in addition, pruning between colors is performed in the same way as in column 24.

The decoding of the green and yellow bit sequences continues as shown above through the entire trellis map from left to right. Figure 7a shows only those transitions that remain after the pruning operations. In the dashed columns spaced by the binding lengths, only one color appears in each space. Other spaces come in two colors. Some transitions do not move forward from some states. This indicates that all the states in the next column to which it would be possible to move from that state will have a larger metric sum from some other state; 25 paths.

Next to the states in column 33 is the sum of the metrics of the path that remains after the last qualifier at the end of that state. When tracks are traced back from these states, it should be noted that the color of the path can only change at color grading-30 for rakes 24, 27, and 30. For example, in column 33, state "11" becomes only the i · green offset. It is known that between columns 30 and 33 this path can * · consist only of green transitions, so the state "11" in column 31 does not present a problem: even if there are two transitions in that state, the backward tracing process knows that the path to be traced comes from the state of column 30 " 11 "because this transition is green-35. The paths ending in the states of column 33 and the associated metric sums are as follows: 101759 13 path ending in state "11": 00-00-00-10-11-11-11-11-11-11-11-11-11, metric sum 22, path ending in state "01": 00-00-00-10-11-11-11-11-11-11-11-11-01, metric sum 20, 5 - path ending in state "10" : 00-00-00-00-00-00-00-00-00-00-10-01-10, metric sum 20, - path ending in "00": 00-00-00-00-00- 00-00-00-00-00-00-00-00, metric sum 22.

10 It will be seen that the metric of the path ending in both state "11" and state "00" is 22. The receiver must decide which path it chooses as the correct path.

The upper path mostly represents consecutive ones and the lower one consecutive zeros. In this case, the above-described shortcomings in the structure of the convolutional encoder are known, so that the receiver can estimate that a sequence of consecutive zeros is more likely, in which case it selects the lower path as the correct path.

This path gives the original information sequence without error. If the structure of the convolutional encoder is better chosen, the occurrence of draw situations between different paths is quite unlikely.

One suggestion for an alternative decoding method instead of the color decoding method according to the invention is to average the corresponding bits in separately received ("different colored") bit sequences describing the same information sequence. The following is a brief example comparing the operation of the mean wedge and the colored method. Assume that a convolutional encoder; · 25 is still the same as above and that the receiver has received the bit sequences 00 01 00 (color A) and 11 10 01 (color B). Figure 8a shows the decoding of bit sequences in the form of a trellis map in the same way as described above with reference to Figure 7a. Column 50 is the initial column with only one state, and column 53 is the column at the link length from which the decoding according to the short example ends. The color decoding method of Kek-30 according to the invention produces four paths, of which the metric sum of the path ending in state "10" of column 53 is 4 and the metric sum of the other paths is 5.

If the bit sequences according to colors A and B are averaged bit by bit, a sequence of number 35 is obtained 0.5 0.5 0.5 0.5 0.0 0.5. The numbers in the sequence can be treated as samples between [0,1], with the metric calculated for each sample being, for example, the formula l- | b-n | the value, where b is the value of n of the "correct" bit used to compute the metric, is the value of that sample. When the decoding of the read sequence by the viterbi algorithm according to the prior art 101759 14 is described in a trellis map, the pattern according to Fig. 8b is obtained. The result is four paths, all of which have the same metric sum (3.5) and can be named so that the path ending in "11" is P1, the path ending in "01" is P2, the path ending in "10" is P3, and the path ending in "00" is "00". "ending path is P4.

5 The bit sequences corresponding to these paths at the output of the convolutional encoder would be as follows:

Pl 11 10 00 P2 00 11 01 10 P3 110101 P4 00 00 00

It has not yet been stated which of the above paths represents the correct, original bit sequence, the reception repetitions of which are represented by colors A and B 15. Assuming that a particular path is correct, and calculating the Hamming distances of the corresponding convolutionally coded sequence and both color sequences, the following table is obtained: correct path A hamming- B hamming- corresponding distance distance sequence _P1_4_J_ P2_2_3_ P3_3_2 When comparing the result based on averaging and ordinary viterbide decoding with the result according to the invention, it is found that the method according to the invention also gives paths P1 and P4, but eliminates paths P2 and P3. In the research that led to the invention, it has been found that the "colored" decoding method according to the invention eliminates paths where the Hamming distances of the corresponding bit sequences relative to the correct sequence in a given bit length of a binding sequence are close to each other (i.e. colors are equally erroneous at the same point) and favored. paths whose Hamming distances of the corresponding bit sequences with respect to the correct sequence in a given bit-length bit period differ significantly from each other (i.e., one color is almost error-free for that bit period compared to other colors). Instead, the procedure 101759 15 based on averaging and ordinary viterbide decoding equally favors all paths that have the same sum of Hamming distances of the corresponding bit sequences. In practice, this means that the method according to the invention works better than the averaging and the ordinary viterbi algorithm if the error bursts contained in the different colors are different in the bit sequence in question. This is usually true when the bit error rate (BER) of the communication link is small enough.

It has been stated above that in the method according to the invention, the trimming of the paths between colors takes place most preferably at intervals of the binding length. This is due to the fact 10 that the effect of the erroneous bit extends in the convolutionally encoded bit sequence over the path length. The method according to the invention can be modified in such a way that the pruning takes place, for example, between two bond lengths or at a different density, but it is most likely to increase the required memory space and computing capacity and to impair the functionality of the method.

15

Next, the application of the method according to the invention in a digital broadcast receiver will be discussed. Standard 20 ETS 300 401 of the European Broadcasting Union (EBU) and the European Telecommunications Standards Institute (ETSI) defines the DAB (Digital Audio Broadcasting) system, which transmits audio data in digital form, but also other information, which can be organized into files. The file is transferred segment by segment. According to the standard, a segment is a single piece of a file of arbitrary size. The transmitting hardware cuts the file into segments, which it numbers in sequence so that each segment is unambiguously identifiable. Each segment is also associated with a 16-bit CRC (Cyclic Redundancy Check) checksum, against which the receiver can determine if there have been any errors in the file transfer. The files are usually sent more than once so that the receiver can try to receive at subsequent repetitions the segments that it has rejected on previous occasions due to errors detected by the CRC.

•

In order to make optimal use of the capacity of the radio channel, the number of repetitions of the file should be kept to a minimum. The following is a method based on the use of the "colored" decoding method according to the invention, in which it is assumed that each transmitted file is played only once. The above discussion is not limited to the DAB system only, but can be applied to all communication systems in which the transmitting device uses convolutional coding and in which the file information is transmitted in segments or equivalent portions. The steps of the method are numbered to illustrate the description.

1. The receiver receives the file segments-5 of the first transmission of the file and decodes them using the prior art viterbide decoding. It tries to compile the submitted file from those segments whose CRC checksum indicates that they do not contain errors. The receiver stores the numbers of the segments found to be invalid, ie their location in the file.

10 2. Simultaneously with step 1, the receiver stores the received segments as sample queues, the samples of which it reads from the received data stream before the viterbide decoding. If the viterbid-decoded segment does not pass the CRC check, the receiver stores the sample string and the viterbid-decoded bit sequence corresponding to the segment in memory. Because a segment contains bit errors, the segment number in it may be incorrect, so the receiver checks the segment number by comparing it to the number of segments received so far (segments are numbered with a sequential sequence number starting from zero). Because convolutional coding and viterbide coding share the effect of the original bits over a range of bits, the receiver stores "extra" samples of at least the trun-20 cation length at the end of the sample sequence so that the entire segment can be decoded with the stored samples. If the segment header or other information received by the receiver does not indicate convolutional coding initialization information, "extra" samples must also be stored at the beginning of the segment to avoid generating errors due to an incorrect initial state at the beginning of the segment to be decoded. In addition to the sample queues 25, the receiver must store information about the location of the segment relative to the logical frame of the DAB system or other structure that controls the use of the puncturing vector in order to know how the depuncture should be performed.

3. The receiver waits for the file to play again.

30 4. During file playback, the receiver receives the segments that were first detected as incorrect. The receiver again stores both the viterbid decoded segment and the sample sequence describing the segment before decoding. If the decoded segment passes the CRC check this time, it can be stored as part of the file and the corresponding sample strings and the bit sequence stored for the first reception can be deleted from the memory.

17 101759 5. If the second time CRC check indicates that there are errors in the segment, the receiver will proceed as follows: 5a. The receiver performs color viterbide decoding according to the invention.

5 The sample strings stored in the first and second reception act as input data sequences of different colors.

5b. The receiver checks whether the CRC checksum included in the bit sequence given by the color decoding indicates errors. If the segment is error-free, the receiver 10 proceeds as in step 4.

5c. If the segment is still incorrect, the receiver compares the two segments produced by normal viterbide decoding and marks the points where the bits of those segments differ from each other. It also means one bit difference on each side of the 15 points. If there are relatively few entries (for example, less than 10; a suitable boundary condition can be found by experimentation), the receiver experiments with possible combinations of all the marked bits and calculates the checksum of the segment obtained in each case. If a modified segment passes the CRC check, it is assumed to be error-free and proceeded as in step 4.

20 5d. If none of the above produces an error-free segment, the receiver waits for the third play of the file. If there are still segments that have been received incorrectly on all three iterations, the receiver can process all three stored segment versions simultaneously as shown in steps 5a to 5c. Before that, however, it can perform the so-called. majority voting, i.e., comparing the stored, viterbid-decoded sequences of my bits with each other and selecting the bit value for each bit position. If the result is an error-free segment, proceed as in step 4. Otherwise, it may cast a majority vote on the sample queues, for example, by selecting the two samples in the same direction (or all samples if they: · are very close to each other) at each sampling position, and averaging them, after which it decodes the generated average sample sequence using standard viterbi decoding. The color decoding method according to the invention functions as a conventional viterbide decoding method if there is only one color.

Although a decoding device has been referred to above as a receiver, it will be apparent to one skilled in the art that the same operation applies to all devices that decode convolutionally encoded digital information.

35 101759 18

Figure 9 shows a simple block diagram of a digital receiver 60 suitable for carrying out the method according to the invention. It has a decoding means 61 equipped to process at least two sample or bit strings to be decoded simultaneously in the manner described above with particular reference to Fig. 7a. The signal to be processed before the decoding element 5 has passed through the prior art receiving parts 62 per se, whereby it has been converted from radio frequency oscillation propagating as electromagnetic radiation into bits (absolute detection) or samples (sliding detection). In order to receive bit or sample strings at different times and input to the decoding member 61 substantially simultaneously, the receiver comprises at least two memory blocks 63 and 64 capable of storing information before the decoding member.

The memory blocks may be physically located on the same memory circuit or there may be several separate memory units for them. After the decoding means, the receiver preferably has an error checking block 65 equipped to examine whether the decoded information is error-free or needs to be further processed to improve error-free, for example by waiting for repetition and applying a color decoding method to the original and repeated sequence. Preferably, the decoded data line is also connected to memory 66 so that the decoded sequences can be stored for later processing. The portion of the receiver after the decoding section is prior art and is designated 67. The control block that controls the decoding operation is shown at 68.

For the first time, the invention provides a method that utilizes two versions of the same convolutionally coded information period containing different error distributions to perform decoding in a manner that eliminates errors. Because the invention utilizes trimming trellis map paths both within and between "colors", it requires only a relatively small amount of memory space and computational capacity compared to a method in which a receiver must store a large number of trellis map paths for comparison. The method of the invention works better than averaging parallel bit sequences, especially if the error bursts contained in the bit sequences are not aligned.

Claims (9)

101759 A method for decoding a convolutionally encoded data sequence, characterized in that at least a first and a second version of said data sequence are produced and the versions are decoded substantially simultaneously to form one decoded sequence 5. A method according to claim 1, characterized in that the decoding operation is divided into periods (21-24; 24-27; 27-30; 30-33), the length of which - measured as the assumed state transitions of the convolutional encoder used in the decoding - is equal to 10 the binding length of the convolutional encoder used to encode the convolutionally encoded data sequence. A method according to claim 2, characterized in that it comprises the steps of 15. decoding both the first and the second version of the data sequence from the beginning (21) to the end (24) of the period independently of the other versions and by trimming among the paths ending in the same state always excluding non-metric-largest paths, - at the end of the period, truncate the found trellis map paths ending in 20 in the same state and formed as a result of decoding different versions, excluding other than the most metric-summed path, at the beginning of the next period at the beginning of both the first and second data sequences decoding the version of all trellis map states, whereby the metric sum of the path starting from a certain trellis map state is added to the metric sum of that state here- ;: 25 degrees, and - at the end of the decoding (33) the correct path is found y is the path of the trellis map with the largest metric sum, and a decoded data sequence is generated by tracing the selected correct path through the trellis map. A method according to claim 1, characterized in that said en-; the first and second versions of said data sequence are produced in a radio receiver by receiving the same information two different times. A method for receiving and decoding a segmented, convolutionally encoded, and repetitive digital file at a radio receiver, characterized in that at least a first and a second version of said segment is received and decoded substantially simultaneously to form a single error-free decoded segment to correct erroneously received segment errors. . A method according to claim 5, characterized in that it comprises the steps of - receiving all the segments of the file on the first play of the file, - storing the received segments as decoded as error-free and the segments received as erroneous without decoding, - receiving at least those segments on the second play. 10 segments were received as erroneous, - the segments received as erroneous on the second repetition are stored decoded, and - the segments received as erroneous on both the first and second repetitions are decoded, using both the first and second repetitions of the segment simultaneously in decoding. The method according to claim 6, characterized in that it further comprises the steps of: generating a third version by receiving said segment on the third repetition of the file, - storing the segments received in error on the third repetition as decoded, and - decoding the segments received as erroneous on the first, second and third repetitions, using the segment versions received on all repetitions simultaneously in decoding. A method according to claim 6, characterized in that it further comprises the steps of: , generating a third version by receiving that segment on the third play of the file, - storing the segments received in error on the third play decoded decoded, and 101759 - undecoded format. 5 An apparatus (60) for decoding a convolutionally encoded data sequence, the apparatus comprising a viterbid decoder (61), characterized in that it comprises a memory (63, 64) for storing said data sequence in at least two different undecoded versions and input to the viterbid decoder (61), said viterbid decoder comprising 10 means to process said versions simultaneously to form a single decoded sequence.