JPH0562365A - Demodulation method for magnetic recording and reproducing device - Google Patents

Abstract

PURPOSE:To demodulate data with a high reliability by judging the state of a data string consisting of inputted samples and learning corresponding demodulated data and using weight data of an obtained neuron layer. CONSTITUTION:Data before encoding corresponding to a sample value is outputted to a perceptron 20 consisting of N input elements 22, L hidden elements 23, M output elements 24, and two stages of neuron layers 26 connecting them. A serial reproduced signal is inputted to a delay element array whose extent of delay is equal to or shorter than the period of data bits connected in series, and the reproduced signal is converted to N kinds of signal shifted from one another by the extent of delay. In the next sampling stage, a parallel signal is obtained. The state of the data string is judged by this learning, and weight data of neuron layers obtained by the learning routine is used to perform demodulation. Thus, a decoder is used to perform demodulation of high reliability even if the resolution is degraded by high-density recording or the signal is degraded by noise.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a demodulation method of a magnetic recording / reproducing device for sequentially demodulating a reproduction signal from a recording medium, and more particularly to a reproduction signal from a recording medium having a regularity in the reproduction signal, for example, a magnetic field. The present invention relates to a demodulation method for a magnetic recording / reproducing device in which a reproduction signal of a disk device is demodulated by using a neural net.

With the increase in speed of computer systems, there has been a demand for higher speed and larger capacity of magnetic disk devices as external storage devices. Therefore, the frequency of the signal handled by the demodulation circuit of the magnetic disk device becomes high, the recording density (BPI) on the medium increases, and the signal quality deteriorates. Therefore, not only the improvement of the head-medium system,
There is a need for measures for improving the signal quality in the demodulation circuit system of the reproduced signal from the magnetic disk device.

[0003]

2. Description of the Related Art Generally, a magnetic disk device records data sent from a host computer on a magnetic disk. Actually, at the time of recording, the sent data is rearranged so as to be suitable for magnetic recording. Encoding is performed and the encoded data is recorded on the magnetic disk. When demodulating a signal reproduced from a magnetic disk, first, encoded data is extracted from the reproduced signal, decoded, and processed into data that can be sent back to a host computer.

FIG. 5 shows a conventional structure of a demodulation circuit in such a magnetic disk device. The signal reproduced by the magnetic head 70 is amplified by a preamplifier 71 arranged near the magnetic head 70, and then an automatic gain control circuit (AGC) 72, a waveform equalization circuit 73, and a low pass filter (LPF). After passing through 74, the amplitude reaches a point P with a sufficient amplitude for performing the demodulation treatment.

At point P, this signal is divided into two,
One to the differentiation circuit 75, the other to the level slice circuit 77.
Be led to. The signal differentiated by the differentiating circuit 75 is input to the zero-cross detecting circuit 76 via the Q point, and a zero-cross pulse corresponding to the zero-cross point of the differential waveform is generated. Then, the pulse from the zero cross detection circuit 76 is input to the gate circuit 78 via the point R. On the other hand, the output at point P input to the level slice circuit 77 is compared with a predetermined slice level voltage set in advance, and when the signal waveform exceeds the level slice voltage, the level slice gate is opened and the signal is level sliced. The signal is input from the circuit 77 to the gate circuit 78 via the point S. The gate circuit 78 removes a pulse irrelevant to the peak of the reproduction signal by the zero-cross pulse and the gate signal, and generates pulsed data corresponding to the peak point of the analog signal input from the preamplifier 71.

FIGS. 6A to 6E show waveforms at points P to T shown in the conventional demodulation circuit of FIG. Figure 6 (a)
Is a reproduced waveform signal of the head 70 at point P, and FIG.
(b) is a waveform at point Q, which is a differential waveform of the waveform at point P. Therefore, this waveform contains a pulse that is not related to the peak of the reproduced waveform. Further, FIG. 6C shows an output waveform of the zero-cross detection circuit 76 at the R point, which is a signal which becomes a high level "H" when the waveform at the Q point passes the zero point. Further, FIG. 6 (d) shows the output waveform of the level slice circuit 77 at the S point, and the slice level voltages VH and VL (shown in FIG. 6 (a)) set in the level slice circuit 77 are reproduced at the P point. It goes high when the signal crosses. FIG. 6E shows the gate circuit 78 at the point T.
Shows the output waveform from the demodulated data in which a pulse irrelevant to the peak of the reproduced waveform is removed.

As described above, according to the conventional circuit, it is possible to obtain pulsed data corresponding to the peak of the reproduction signal.

[0008]

However, in the conventional demodulation circuit configured as described above, the recording data
The peak corresponding to "1" is detected by the level slice and the zero cross point detection of the differential waveform. Therefore, when observing the waveform like the center peak in the 3-bit pattern, even if it is expected to be a peak from the front-back relationship, if it falls below the slice level, it may not be recognized as a peak and a demodulation error may occur. , Also, recorded data "0"
There is a problem in that even a portion corresponding to (1) is recognized as a peak when the slice level is exceeded.
As described above, in the conventional demodulation method, when the level of the portion corresponding to "0" of the recording data exceeds the peak level corresponding to "1" of the recording data, it becomes impossible to demodulate. Therefore, when performing higher density recording,
The quality of the signal to be handled deteriorates more and more, resulting in a decrease in S / N and a decrease in resolution.

The present invention solves the problem in the method of demodulating a reproduction signal having regularity in the reproduction signal as in the conventional magnetic recording / reproducing apparatus described above, and "1" which is lower than the conventional slice level in the reproduction signal. Exceeding signal or slice level
A "1" signal is a "1" even if it contains a "0" signal,
"0" signal is demodulated as "0", and further, the demodulation data of the magnetic recording / reproducing device with high reliability that can obtain the demodulated data that can be directly returned from the reproduced signal to the host computer is provided by using the neural net. The purpose is to do.

[0010]

FIG. 1 shows a principle configuration of a demodulation method of a magnetic recording / reproducing apparatus of the present invention for achieving the above object. As shown in FIG. 1, the present invention uses a signal reproduced from a magnetic recording medium in a magnetic recording / reproducing apparatus that encodes data from a high-order computer and stores the encoded data in a magnetic recording medium. A method for demodulating data from a high-order computer before conversion, wherein in learning stage 1 of the perceptron, N input elements and L hidden elements, M output elements, and a two-stage neuron layer connecting them From a magnetic recording medium to a perceptron consisting of
Learning is performed so as to output uncoded data corresponding to the sample values. In the data conversion step 2, the serial reproduction signal from the magnetic recording medium is input to N-1 delay element trains having the same or shorter delay amount Td as the data bit period connected in series, and the reproduction signal is reproduced. Are converted into N types of signals that are shifted by the delay amount Td by time intervals. In the sampling stage 3, parallel sampling is performed on the N types of signals for a sampling period Ts that is N times the delay amount Td to obtain a parallel signal. Then, in the data demodulation stage 4, the obtained N sample values are input to the perceptron, the values of the M output elements are compared with a threshold value, and the state of the input data string consisting of N samples is determined. Then, the demodulated data corresponding to this is output.

The above-described perceptron performs demodulation by writing optimum neuron layer weights into the storage means by learning before shipment of the magnetic recording / reproducing apparatus and applying weight data to the neuron layers from this storage means during operation. To do so. It should be noted that the demodulator of the magnetic recording / reproducing apparatus performs learning a plurality of times by a plurality of learning patterns at the time of operation of the apparatus or at regular intervals, and the weight data of the neuron layer obtained by this learning routine. Alternatively, the weight data of the neuron layer stored in the storage unit may be rewritten.

[0012]

According to the demodulator of the magnetic recording / reproducing apparatus of the present invention,
The reproduction signal from the magnetic recording / reproducing apparatus is output as N kinds of signals which are shifted by the delay amount Td by the time interval by N-1 delay elements having the same or shorter delay amount Td as the data bit period. The signals are sampled in parallel at a sampling period Ts which is N times the delay amount Td and input to the neural network. The neural network stores in advance a learning pattern for converting the data sequence of the reproduced signal into data from the host computer. In the neural network, the input data sequence consisting of N input samples is stored. The state of is determined and the demodulated data corresponding to this is output. At this time, learning is performed a plurality of times by a plurality of learning patterns, and demodulation is performed using the weight data of the neuron layer obtained by this learning routine.

[0013]

Embodiments of the present invention will be described in detail below with reference to the accompanying drawings. FIG. 2 shows the configuration of an embodiment of the demodulator of the magnetic recording / reproducing apparatus of the present invention. The present invention is a demodulator of a magnetic recording / reproducing apparatus that demodulates recorded data from a reproduced signal from a magnetic recording medium. In the following, a reproduced signal of a magnetic disk device in which positive and negative peaks always appear alternately in the reproduced signal will be described. An example will be described later. In the magnetic disk device, as described above, when recording the data sent from the host computer on the magnetic disk, the code that rearranges the data sent from the host computer to be suitable for magnetic recording. The encoded data is recorded on the magnetic disk. The reproduction signal reproduced from the magnetic disk directly corresponds to this encoded data.

FIG. 2 shows a case where demodulation is performed based on N consecutive sample values, and N consecutive sample values obtained from the characteristics of the magnetic recording and the conditions such as the code used. Shows the case where the total number of possible states is M. In the figure, reference numeral 21 denotes a delay element sequence, which is composed of N-1 delay elements having a delay time τ = Td (Td is a data period). The reproduction signal from the magnetic disk device (not shown) input to the delay element array 21 is divided into N-1 signals that are delayed by the delay time Td and output from the delay element array 21. ..

A total of N-1 signals of the reproduced signal and the N-1 signals which are shifted in time by Td from each other are taken as N samples, and N input elements 22 and L hidden elements 2 are provided.
3 and M output elements 24, and two-stage neuron layers 26 and 27 connecting these output elements 24 are input to the perceptron 20. The number M of output elements 24 of the perceptron 20
Is the number of patterns that the N sample values of the reproduction signal from the magnetic recording medium can take. Then, the perceptron performs learning described below so that when a certain data pattern is input to the input element 22, an output signal is output to the corresponding output element 24, each neuron layer 26,
27 weights are determined.

Each neuron layer 2 of the perceptron 20
Reference numerals 6 _and 27 have weights represented by Vij _and Wij, respectively, and the value Yp Zq of the hidden element 23 and the output element 24 is expressed by the following equation. Yp = [sigma] ([Sigma] Wkp * Xk) ... Zq = [sigma] ([Sigma] Vpq * Yp) ... However, [Sigma] is the sum of k = 0 to N, and 0≤p≤L _, 0.
It is assumed that ≦ q ≦ M. Then, σ (S) is defined by the following equation.

Σ (S) = 1 / (1 + exp (−S)) Now, let us say that the set of learning patterns is X, the output of the perceptron is h (X), and the target function is d (X). It is expressed as follows. _{h (X) = (h 0} (X), h 1 (X), h 2 (X), ... hq (X), ... h M (X)) ... d (X) = (d 0 (X), d ₁ (X), d ₂ (X), ... dq (X), ... d _M (X)), where hq (X) = Zq, and the target function dq ( X) is the learning pattern of X
For the learning pattern of q, "1" is output, and for other learning patterns, "0" is output.

Dq (X) = 1 (in the case of a learning pattern of q) ... dq (X) = 0 (in the case of other learning patterns) ... ′ Further, the state of the hidden element 23 when the learning pattern X is given is Let f (X) be represented by the following formula. _{f (X) = (f 0} (X), f 1 (X), f 2 (X), ... fp (X), ... f L (X)) ... Based on the above, {h (X) - d (X)} as ² is minimized, the procedure for determining a hidden element 23 weight Vpq neuron layer 27 between the output element _24, and the weight Wkp neuron layer 26 between the input element 22 and the hidden element 23 As shown in FIG.

In step 301, the weight V _pq of the neuron layer 27 and the weight W kp of the neuron layer 26 are arbitrarily determined.
Then, in step 302, p = 0,1,2, ..., L, q = 0,1,2, using the equation: Vpq = Vpq−ΔΣδ ₂ q (X) · fp (X)
The weight V between the hidden element 23 and the output element 24 for M
pq is updated, and subsequently in step 303, k = 0,1,2, ..., N, p = 0,1,2, using the equation, Wkp = Wkp−ΔΣδ ₁ p (X) · X.
The weight W between the hidden element 23 and the input element 22 for L
Update kp. However, in the above equation, δ ₁ p (X) and δ ₂ q
(X) is expressed by the following equation, and Σ is the sum of q = 0 to M.

Δ ₁ p (X) = {Σδ ₂ q (X) ・ Vpq} ・ fp (X) ・ (1−fp (X)) ... δ ₂ q (X) = {hq (X) −dq ( X)} hq (X) .multidot. (1-hq (X)) ... In the following step 304, it is judged whether or not the error {h (X) -d (X)} ² is sufficiently small, and when it is not small (NO) Returns to step 302 and repeats steps 302 and 303, and when it is sufficiently small (YES), this routine ends in step 305.

In this way, in the above embodiment, when a pattern having N sample values is input to the perceptron 20, the neuron layer is arranged so that an output is output to the output element 24 corresponding to the pattern one to one. The optimum weights Wkp and Vpq of the neuron layer 26 and the neuron layer 27 are obtained by learning.
The optimum weights Wkp and Vpq of the neuron layers 26 and 27 are determined by learning before shipment of the magnetic recording / reproducing apparatus and stored in the ROM, and the weight data is applied to the neuron layers 26 and 27 from the ROM during operation. To

The above-mentioned neuron layer 26 and neuron layer 2
The learning pattern given to 7 is such that the data from the host computer is decoded at once from the data reproduced according to the magnetization state of the magnetic disk. About this learning Figure 4
Will be explained. For example, suppose that the data given from the host computer is "101011", and this data is "1010" suitable for magnetic recording by the encoder 41.
It is assumed that the data is converted to 0101 ″ and recorded on the magnetic disk 42. In this case, the head 43 from the magnetic disk 42 is recorded.
The reproduced signal by means of the waveform has a shape as shown by the waveform 44.
Conventionally, the waveform 44 is demodulated to obtain data “100101” suitable for magnetic recording, and the decoder 45 is used to obtain parallel conversion data “101011” of serial data given from a host computer. In the present invention,
With respect to this waveform 44, the parallel conversion data “101011” of the serial data given from the upper-level computer at once is obtained without obtaining the data “100101” suitable for magnetic recording by the learning pattern given to the neuron layer 26 and the neuron layer 27. I am trying to ask.

As described above, since the output element 24 of the perceptron 20 outputs the parallel conversion data of the serial data given by the higher-order computer, which corresponds to the input data pattern in a one-to-one manner, the output element 24 is parallel-connected.
If the serial conversion circuit 25 is connected, the serial data given by the host computer can be obtained as demodulated data without using a decoder.

The determination of the weights Wkp and Vpq of the neuron layers 26 and 27 of the perceptron 20 is made not only before the shipment of the magnetic recording / reproducing apparatus but also at the time of operation of the apparatus or at regular intervals. Alternatively, learning may be performed a plurality of times using a plurality of learning patterns, and the neuron layer weight data stored in the RAM or the like may be rewritten with the neuron layer weight data obtained by the learning routine.

[0025]

As described above, according to the present invention, even if the resolution of the reproduced signal from the magnetic recording medium is lowered due to high density recording or the signal is deteriorated due to noise, the data demodulation is performed by the decoder. This is possible without using a magnetic disk drive and greatly contributes to improving the reliability of the magnetic disk device.

[Brief description of drawings]

FIG. 1 is a principle configurational diagram showing a demodulation method of a magnetic recording / reproducing apparatus of the present invention.

FIG. 2 is a configuration diagram showing a configuration of an embodiment of a neural net used in the demodulation method of the magnetic recording / reproducing apparatus of the present invention.

3 is a flowchart showing an operation of a demodulator of the magnetic recording / reproducing apparatus of FIG.

FIG. 4 is an explanatory diagram for explaining the operation of the neural net of the present invention.

FIG. 5 is a demodulation circuit diagram of a conventional magnetic disk device.

6 is a waveform diagram showing operation waveforms of respective parts of the conventional circuit of FIG.

[Explanation of symbols]

 20 ... Perceptron 21 ... Delay element sequence 22 ... Input element 23 ... Hidden element 24 ... Output element 25 ... Parallel-serial conversion circuit 26 ... Neuron layer 27 ... Neuron layer

 ─────────────────────────────────────────────────── ─── Continued Front Page (72) Inventor Takenori Oshima 1015 Kamiodanaka, Nakahara-ku, Kawasaki-shi, Kanagawa Fujitsu Limited

Claims (5)

[Claims] 1. A magnetic recording / reproducing apparatus which encodes data from a higher-order computer and stores it in a magnetic storage medium by performing encoding suitable for magnetic recording, from a higher-order computer before encoding from a signal reproduced from the magnetic recording medium. A method for demodulating data, in which N input elements and L hidden elements, M output elements, and a perceptron composed of two stages of neuron layers connecting these elements are used for every N elements from a magnetic recording medium. A learning step (1) of a perceptron for learning to output uncoded data corresponding to a sample value, and a serial reproduction signal from the magnetic recording medium, which is the same as or shorter than the data bit period connected in series. Delay amount T
A data conversion step (2) of inputting to N-1 delay element trains having d and converting the reproduction signal into N kinds of signals which are shifted by a time interval of a delay amount Td, and delaying the N kinds of signals. Sample period T that is N times the amount Td
Sampling step (3) for parallel sampling with s to obtain a parallel signal, and the obtained N sample values are input to the perceptron, the values of M output elements are compared with a threshold value, and N A demodulation method for a magnetic recording / reproducing apparatus, comprising: a data demodulation step (4) for determining a state of an input data string consisting of samples and outputting demodulated data corresponding to the state. 2. The number M of output elements of the perceptron is set equal to the number N of input elements, and an output corresponding to the sample value input to the input element in a one-to-one correspondence is output from this output element. 2. The demodulation method of the magnetic recording / reproducing apparatus according to claim 1, wherein N demodulated data strings are parallel-serial converted at the sample period Ts and output. 3. The number M of output elements of the perceptron
Is the code rate number C of the code using the number N of input elements
There is only a few that _r times, performed the decoding and demodulation of the reproduced signal at the same time, this demodulation method according paragraph 1, characterized in that the parallel ─ are serially converted output. 4. The demodulation is performed by writing the optimum weight of the neuron layer into the storage means by learning before shipment of the magnetic recording / reproducing apparatus, and by giving the weight data from the storage means to the neuron layer during operation. The demodulation method of the magnetic recording / reproducing apparatus according to claim 1. 5. When the apparatus is in operation or periodically at time intervals, learning is performed a plurality of times with a plurality of learning patterns, and the weight data of the neuron layer obtained by this learning routine is stored in the storage means. 5. The weight data of the neuron layer is rewritten.
13. A demodulation method for a magnetic recording / reproducing apparatus according to any one of 1.