JPH06112931A - Pre-processing method for digital signal - Google Patents

Abstract

PURPOSE:To easily build up a waveform processing unit utilizing the neural network by providing the pre-processing method making position and time axis matching for teacher data and learning data. CONSTITUTION:The distribution of a code interval of deteriorated learning data is obtained (steps 100-102), and the distribution of the code interval and the basic code interval of teacher data are compared to generate a code string of learning data (steps 103, 104). The teacher data and the learning data are compared and retrieved as code strings and the position is matched based on the coincidence of the code strings (steps 105-107). Furthermore, the code length is revised and final position and time axis are matched based on the result of a retrieval code string of an optimized code length (steps 108-112).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a digital signal pre-processing method used for constructing a waveform processing device using a neural network to remove signal deterioration that occurs during recording / reproduction / transmission. is there.

[0002]

2. Description of the Related Art Conventionally, signal processing in a transmission system (recording / reproducing) which is easily affected by non-linear distortion such as recording / reproducing by a high density optical recording medium / magnetic recording medium or transmission / reception in a communication path of a high information amount In that case, select the most suitable modulation method,
Efforts are being made to suppress signal deterioration by using various waveform equalizers (equalization / filtering), noise cancellers, emphasis / de-emphasis, and the like.

Under such a condition, a neural network having a non-linear conversion system is used to learn the characteristics of the recording / reproducing system using the above-mentioned optical / magnetic recording medium and the communication path system of a high information amount, and the transmission system. It is conceivable that the deteriorated digital signal from the device is waveform-equalized and output in real time, and the present applicant has already applied for a patent (Japanese Patent Application No. 3-319).
No. 52 Invention title "waveform equalization apparatus by neural network and design method thereof").

A waveform equalizer using this neural network N delays an input signal (for example, a signal whose waveform has been deteriorated from a reproducing device) by a delay unit connected in series for a predetermined time and temporally. It is configured to input a plurality of input and output signals (values) before and after to a neural network for waveform equalization. The neural network is, for example, a layered (for example, three layers of an input layer, an intermediate layer, and an output layer) neural network using a back propagation as a learning algorithm, and is a combination of many units connected with variable weights. (Body). The unit is a conversion system with non-linear input / output characteristics, and the total sum from the previous layer obtained by multiplying the output value from the previous layer by an independent weight is input to the unit, and this total sum is nonlinearly transformed. It is output to the rear layer.

By utilizing this neural network, the characteristics of the recording / reproducing system and the communication path system described above are learned in advance, and the deteriorated digital signal from the transmission system is waveform-equalized and output. That is, as shown in FIG.
The reproduced signal from the recording medium Mo reproduced by the reproducing device D is A / D converted and taken into the digital memory DM at a predetermined sampling period. Known digital data used at the time of learning are recorded in advance on the recording medium Mo. The data taken in the digital memory DM is
It is transferred to the workstation WS for learning processing and learning calculation processing is performed.

As shown in FIG. 5B, in the workstation WS, a simulation program of a neural network having a learning algorithm by a known backpropagation is prepared, and a learning calculation process is executed on the neural network N. To be done. That is, the captured data (hereinafter referred to as learning data) reproduced by the reproducing device D, captured in the digital memory DM and transferred to the workstation WS is input to the neural network N, and the input signal is input layer → intermediate layer. The output signal is output from the output layer after transmission and calculation processing from the layer to the output layer. This output signal is based on the weight obtained by the learning so far. The calculated output signal is compared with a teacher signal (teacher data based on known digital data recorded in advance on the recording medium Mo), and transmitted / calculated in the order of output layer → intermediate layer → input layer. It is processed and learned.

[0007] Learning means to reduce the difference between the actual output value (output signal) and the desired output value (teacher signal) by setting the weight (coefficient) which is the strength of the coupling between the units of the neural network N. It is variable and converges. The characteristics of the transmission / reproduction system to be learned (learned) include the characteristics of the data itself, the characteristics of the recording device, the characteristics of the reproducing device, the characteristics of the transmission line, etc., and depending on the size of the neural network,
Select appropriately. Based on various learning data, the above-mentioned learning is repeated until the desired correct answer rate (performance) is obtained, and the weight representing the strength of the coupling between the units of the neural network N is changed.

[0008]

In constructing such a waveform equalizer using a neural network, it is necessary to match the timing of the learning data (reproduction data) from the reproducing device with the teacher data. If there is a timing difference between the two, correct learning will not be performed. For this reason, conventionally, a special signal (positioning signal) for adjusting the timing of the teacher data is added, and this timing adjustment (positioning) is performed manually by visual inspection. For example, in FIG.
As shown in, the start end s0 'of the learning data that will correspond to the start end t0 of the teacher data is searched for. However, such a timing adjustment method has problems that the accuracy is low and it takes time.

Further, the learning data (reproduction data) contains a fluctuation of the time axis, that is, a so-called jitter component due to uneven rotation of the device, intersymbol interference and the like. For this reason, there is a time lag between the teacher data and the learning data, which also becomes an obstacle to correct learning. In particular, this jitter component is to be removed by the waveform equalizer to be constructed, and there is no means other than simple linear correction of the time axis between the two as pre-learning processing. .
For example, as shown in FIG.
The time interval of the learning data is corrected (proportionally compressed or expanded) from s0 to se to s0 'to se' so as to correspond to te. However, such time-axis correction is not sufficient, and there is a problem that learning efficiency is extremely poor.

Therefore, according to the present invention, pre-processing of a digital signal for adjusting the timing of learning data (reproduction data) and teacher data and correcting the time axis, which is used in constructing a waveform processing device using a neural network. A method is provided so that learning of a neural network can be performed easily and efficiently.

[0011]

In order to solve the above problems, the present invention provides a first digital signal (for example, teacher data) and a first digital signal which is deteriorated as shown in FIG. A method of preprocessing a digital signal for aligning a position and a time axis with two digital signals (for example, learning data), and obtaining a distribution of code intervals of the second digital signal (steps 100, 101, 102), The distribution of this code interval is compared with the basic code interval of the first digital signal (step 103), and the code string of this digital signal is created from the code interval of the second digital signal (step 104). A search code string having an arbitrary code length is extracted from the created code string of the second digital signal (step 10).
5) The code string of the first digital signal that matches this search code string is searched (step 106), and the first and second digital signals are aligned based on the code string that matched the search (step 106). 107) providing a preprocessing method for the digital signal in which the time axes of the first and second digital signals are adjusted based on the result of the alignment (step 108), and further, the code length of the search code string is set. The position and time axis alignment is repeated by increasing or decreasing (steps 109, 110 and 111), and the result at the code length with the smallest error between the position and time axis aligned first and second digital signals is finally obtained. The present invention provides a preprocessing method of a digital signal which is a result of various position and time axis alignment (step 112).

[0012]

According to the above-described digital signal preprocessing method, the first digital signal and the deteriorated second digital signal are compared and searched as a code string, and alignment is performed based on the matching of the code strings. Is done. Furthermore, position and time alignment is performed according to the result of the search code string having the optimized code length by changing the code length.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a digital signal preprocessing method according to the present invention will be described below with reference to the drawings. FIG. 1 is a flow chart showing the procedure of a digital signal preprocessing method. Figure 2
FIG. 6 is a diagram for explaining coding of learning data. FIG. 3 is a diagram for explaining the alignment of the position and time axis of a digital signal.

This preprocessing method of a digital signal is used when constructing a waveform processing device using a neural network to remove signal deterioration, and learning data SD (second digital signal which is a deteriorated signal from a reproducing device) is used. The position and the time axis of the signal) and the teacher data TD (first digital signal) are matched. The learning data is a deteriorated reproduction signal from the reproduction device D and is sampled by the digital memory DM, as already described with reference to FIG. The learning data and the teacher data are subjected to arithmetic processing (learning processing) on the workstation WS to construct a neural network. The digital signal preprocessing method described in the present embodiment is also one in which arithmetic processing is performed on this workstation WS. The digital signal targeted by the present invention is digital data configured with a code interval of 3T to 11T with respect to the basic data clock T, as shown in FIG. Learning on workstation WS is about 40
It is repeatedly executed for about 0 pieces of digital data (code).

The procedure of the digital signal preprocessing method shown in FIG. 1 will be described in detail below. First, as shown in FIG. 2A, the values of the sampled learning data are integrated and then divided by the sample time to obtain the average value of the learning data (step 100). And
With this average value as the 0 level (= threshold value), the zero cross point (the time when the data crosses the 0 level) is obtained. Since the learning data is sampled discrete values (digital data), it is obtained by linear interpolation from the sampling points before and after. Then, the interval a between the zero cross points of the learning data is obtained (step 101).

Next, as shown in FIG. 2B, a (frequency) histogram of the intervals a of the zero cross points of the obtained learning data is created (step 102). Learning data is from 3T
Since 11T (T is a basic data clock) is distorted, the frequency histogram is an envelope having a maximum value and a minimum value corresponding to the interval of 3T to 11T. Then, the interval of the minimum value an in the frequency histogram is obtained (step 1
03).

Then, the interval of the minimum value an in the frequency histogram is regarded as corresponding to the time width of each data (that is, 3T to 11T), and the learning data is labeled (step 104). For example, a3 ≦ a <a4 is 3T data, a4 ≦ a <a5 is 4T data, ..., a11 ≦ a <a12 is 11T data, and “3T”, “4T” ... “11
A label is attached to the learning data, such as "T".

In this way, by assigning labels to all the learning data, the learning data is labeled with the label "3".
Label group composed of "T", "4T" ... "11T",
It is grasped as a character string group (hereinafter referred to as a learning label group). That is, the learning data sampled as shown in FIG. 3C becomes the labeled learning label groups SL1 to SLe as shown in FIG.

On the other hand, the original data to be the teacher data are digital values and can be regarded as label groups TL1 to TLe (hereinafter referred to as teacher label groups) in which "3T", "4T" ... "11T" are combined. The teacher label group as shown in FIG. 3A corresponds to the teacher data TD that changes with time as shown in FIG. 3B.

Therefore, of about 400 pieces of learning data, about 200 pieces of continuous learning data are searched for in the search label groups SL3 to SL.
As n (label length is 200) (step 105), the teacher label group is compared and searched. That is, from the learning level groups SL1 to SLe shown in FIG. 3D, search label groups SL3 to SLn as shown in FIG. 3E are created, and the same number of teacher label groups as the search label group are created. One is extracted, and the one having the smallest difference (error) is set as the label matching (step 106). FIG. 3 shows a state in which the search label groups SL3 to SLn and the teacher label groups TL2 to TLm are label-matched. The learning data and the teacher data are aligned with each other based on the search label group and the teacher label group that have been label-matched (step 107). That is,
The time t1 of the teacher data corresponding to the teacher label TL2 and the time s2 of the learning data SD corresponding to the search label SL3 are aligned as the same timing.

Further, the time axes of the aligned learning data and teacher data are adjusted. Assuming that the time width of the teacher data is ta, the time width of the learning data is tb, and tb = k · ta, both time axes (entire time length) are linearly corrected by the ratio k and matched. That is, based on the ratio k, the time axis of the learning data (or the teacher data) is corrected (expanded or compressed) to match the time intervals t1 to tm and the time intervals s2 to sn (step 108). Then, the difference b between the temporally corrected learning data and the teacher data is obtained for each digital data (code), and the sum c of the difference b is obtained (step 109). Since the correction of the time axis is only linear correction as a whole, a difference that is an error occurs for each digital data (code). The difference sum c is used as the evaluation value of the label matching of the search label group in the label length. The smaller the total difference c is, the higher the evaluation value is.

Next, in the above label matching, the number of search label groups to be compared is optimized (the label length is optimized). Since the learning data partially contains jitters and errors, the amount of label matching deviation, that is, the sum of differences c changes depending on the number of search label groups.
The larger the number of search label groups, the more accurate the matching may be performed. However, the smaller the number of search label groups, the more accurate the matching may be performed depending on the variance of jitter and error. Therefore, optimization of label length is required.

Therefore, about 200 labels are grouped into 5 search label groups.
The search label group of about 100 to 300 learning data is increased / decreased by about 0%, compared with the teacher label group and searched, and the label matching is performed for each number of label groups to obtain the total difference c (step 110, 11
1). The average of the sum of differences c is 0, and the one with a small variance is set as the optimum number of search label groups, and this is determined as the final correction (step 112).

The learning data is aligned with the position and time axis at the optimum search label length, and finally corrected data is input to the neural network, and the difference between the output value of the neural network and the teacher data is calculated. The weight of the neural network is learned and changed by a known back propagation learning method so that the weight becomes smaller. In this way, the neural network is constructed (see FIG. 5 (B)).

According to this method, since the learning data and the teacher data are time-aligned, extremely efficient learning is performed. In addition, unlike the conventional method, it is not necessary for the neural network builder to adjust the time, and learning can be performed based on many different teacher data (learning data), and a high-performance neural network waveform processing device can be obtained. To be

The digital signal preprocessing method described above may be used not only for the construction of the neural network but also for the design and construction of the adaptive filter.

[0027]

As described in detail above, according to the preprocessing method for a digital signal according to the present invention, a digital signal which is an original signal and a digital signal in which this original signal is deteriorated are compared and searched as a code string. Positioning is performed based on the coincidence of the code strings, and further, the final position and time axis alignment are performed by the search code string having the optimized code length, so the position and time axis alignment of both are easy, And it can be accurate. Further, since it can be processed by software, it can be automatically executed by a computer.

[Brief description of drawings]

FIG. 1 is a diagram showing an embodiment of a digital signal preprocessing method according to the present invention, and is a flow chart showing a procedure of a digital signal preprocessing method.

FIG. 2 is a diagram for explaining coding of learning data.

FIG. 3 is a diagram illustrating alignment of a position of a digital signal and a time axis.

FIG. 4 is a diagram illustrating a configuration of a digital signal.

FIG. 5 is a diagram illustrating the construction of a waveform equalizer using a neural network.

FIG. 6 is a diagram illustrating position and time axis alignment of a digital signal.

[Explanation of symbols]

100 to 102 Steps 103 to 104 for obtaining the distribution of code intervals of deteriorated learning data Steps to create a code string of learning data 105 to 107 Steps to perform alignment based on matching of code strings 108 to 112 Optimized Step of making final position and time axis alignment by the search code string TD First digital signal (teacher data) SD Second digital signal (learning data) a Zero-cross point interval an Minimum value c Sum of differences SLn search Label group (Learning label group) TLm Teacher label group

Claims (3)

[Claims] 1. A preprocessing method for a digital signal, wherein a first digital signal and a second digital signal in which the first digital signal is deteriorated are aligned in terms of position and time, and the second digital signal. The distribution of the code intervals of is calculated, the distribution of the code intervals is compared with the basic code interval of the first digital signal, and the code string of the digital signal is created from the code interval of the second digital signal. A search code string having an arbitrary code length is extracted from the created code string of the second digital signal, and a first code that matches the search code string is extracted.
The digital signal code sequence of the first digital signal is searched, the first and second digital signals are aligned based on the searched code sequence, and the first and second digital signals are aligned based on the alignment result. A method for preprocessing a digital signal, characterized in that the time axes of the signals are matched. 2. A pre-processing method for a digital signal, which aligns the position and time axis of a first digital signal and a second digital signal in which the first digital signal has deteriorated, the second digital signal The distribution of the code intervals of is calculated, the distribution of the code intervals is compared with the basic code interval of the first digital signal, and the code string of the digital signal is created from the code interval of the second digital signal. A search code string having an arbitrary code length is extracted from the created code string of the second digital signal, and a first code that matches the search code string is extracted.
The digital signal code sequence of the first digital signal is searched, the first and second digital signals are aligned based on the searched code sequence, and the first and second digital signals are aligned based on the alignment result. The time axis of the signal is adjusted, the code length of the search code string is increased or decreased, and the position and time axis adjustments are repeated. The error between the first and second digital signals aligned in the position and time axis is minimized. A preprocessing method for a digital signal, characterized in that the result of a small code length is used as a final position and time axis alignment result. 3. The first digital signal is a teacher signal for learning when constructing a neural network, and the second digital signal is a deteriorated learning signal to be subjected to waveform processing by the neural network. A method of preprocessing a digital signal according to claim 1 or 2.