JPH0619868A - Waveform processor in neural net and its design method - Google Patents

Abstract

PURPOSE:To provide a highly efficient waveform processor and its design method with easy constitution through the use of a neural net for removing signal deterioration which occurs at the time of transmitting record. CONSTITUTION:The waveform processor 100 using the neural net (N) is provided on a record transmission-side. The waveform processor consists of a conversion means where plural conversion systems consisting of non-linear conversion systems are connected by independently fixed weight and the fixed weight is that which is previously learnt as against the characteristic of a recording transmission system by the neural net. The recording transmission system inputs a deteriorated original signal to the conversion means, and the original signal is waveform-processed so that deterioration by the recording transmission system is cancelled so as to output it to the recording transmission system.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a waveform processing device provided on the recording and transmitting side in order to remove signal deterioration caused during recording and transmission, and particularly, it uses a neural network and has a simple structure and high performance. A high-performance waveform processing device and a design method thereof are provided.

[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. For example, a transversal filter has been used as a waveform equalizer. The transversal type filter delays an input signal with a delay circuit, amplifies it with amplifiers having different amplification coefficients, adds and outputs it, and equalizes the waveform.

However, there are many causes of signal deterioration and it is not possible to quantitatively grasp which factor and how much the signal deterioration has occurred. Therefore, it is often the case that the conventional waveform equalizer cannot deal with the problem. That is, the signal (code)
Since different coefficient values (for example, the amplification coefficient of the amplifier in the above-mentioned transversal filter) are required depending on the degree of interference (degree of signal deterioration) and the cause of deterioration,
It is difficult to uniquely determine this, and the conventional linear waveform equalizer cannot sufficiently cope with this.

Under such a circumstance, 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 high-information-volume communication path system. It is conceivable that the deteriorated digital signals from these transmission systems are waveform-equalized in real time and output, and a patent application has already been filed by the present applicant (Japanese Patent Application No. 3-31952, entitled "waveform by neural network"). Equalizer and its design method ").

A waveform equalizer using this neural network 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 the time is changed in time. A plurality of input signals (values) are input to the neural network and waveform equalization is performed. The neural network is, for example, a layered neural network (for example, three layers including an input layer, an intermediate layer [hidden layer], and an output layer) using a back propagation as a learning algorithm, and is connected with a large number of variable weights. It is a combination of units (body).
The unit is a conversion system with non-linear input / output characteristics,
The unit receives the sum from the front layer obtained by multiplying the output value from the front layer by an independent weight, and the sum is nonlinearly converted and output to the rear layer.

Using 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. 7 (A), this system is composed of a conversion means, ie, a waveform processing device 203, which is a weight learned in advance with respect to the characteristics of the recording and transmitting system by the neural network NN. The present signal deteriorated by is input to the waveform processing device 203, subjected to waveform processing (waveform equalization), and then output. Reference numeral 200 is a recording / transmission device, and 202 is a reproduction / reception device. Then, the learning is as shown in FIG.
The current signal is used as a teacher signal, and the output of the waveform processing device 203 by the neural network NN is used as a student signal, and the learning algorithm of back propagation is executed.

[0007]

The waveform equalizer using such a neural network has extremely high performance for waveform equalization of digital signals. However, there is a limit to the correspondence on the reproducing / receiving device side, the learning efficiency is poor, and an extremely long learning time is required, and the device scale (neural net) is likely to be large. Moreover, since it is necessary to have a waveform equalizer corresponding to the number of reproducing / receiving devices, it is not realistic to make it large-scale. Therefore, the present invention provides a waveform processing apparatus by a new neural network in which waveform processing by a neural network is performed especially on the recording / transmission apparatus side, and a design method thereof.

[0008]

In order to solve the above-mentioned problems, the present invention connects a plurality of conversion systems composed of nonlinear conversion systems with independent fixed weights, and these fixed weights are connected by a neural network. The conversion means is a weight that has been preliminarily learned with respect to the characteristics of the recording / transmission system, and the current signal deteriorated by the recording / transmission system is input to the conversion means to change the original signal so as to cancel the deterioration by the recording / transmission system. A waveform processing device using a neural network, which is characterized in that waveform processing is performed and output to a recording and transmission system.

Further, a plurality of conversion systems composed of non-linear conversion systems are composed of conversion means connected by independent fixed weights, and a current signal deteriorated by the recording transmission system is waveform so as to cancel the deterioration caused by the recording transmission system. A method of designing a waveform processing device for processing and outputting to a recording / transmission system, wherein a plurality of conversion systems which are equivalent to the waveform processing device and which are non-linear conversion systems are connected with independent variable weights. The weight is variably learned according to the characteristics of the recording and transmission system by a neural network, and a fixed weight of the conversion means is designed and determined based on the variably learned weight. A method of designing a waveform processing device using a neural network is provided.

Further, there is provided a means for delaying an original signal which is deteriorated by a recording and transmission system, and a current signal which is temporally moved forward and backward is inputted to the converting means. The present invention provides a designing method for a processing device or a waveform processing device using the neural network described above.

[0011]

According to the above-described neural network waveform processing apparatus and its design method, the current signal is subjected to waveform processing by the waveform processing apparatus, and the waveform-processed signal is recorded and transmitted.
It is deteriorated by the recording and transmission system. The waveform processing and the deterioration cancel each other and are reproduced and received as a signal close to the original current signal.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a waveform processing apparatus using a neural network and a design method thereof according to the present invention will be described in detail below with reference to the drawings. FIG. 1 is an overall configuration diagram of a waveform processing device using a neural network. FIG. 1A is a diagram showing a substantial processing state after design learning, and FIG. 1B is a diagram showing a processing state during design learning. is there. The waveform processing apparatus 100 is a waveform processing apparatus using a neural network as shown in FIG. 2, and is specifically configured as a substantial signal processing apparatus as shown in FIG.

(Basic Concept of Waveform Processing Process by Neural Network N) First, the waveform processing process by the neural network N, which is the basic concept of the waveform processing device by the neural network and its designing method, will be described. FIG. 2 is a conceptual diagram of waveform processing (signal processing) by a neural network (neural network). As shown in the figure,
In a waveform processing process by a neural network, an input signal (original signal described later) is given a predetermined delay time by a delay unit DL connected in series, and a plurality of input signals (values) temporally before and after are given. It is configured to be input to the neural network N.

The neural network N is a layered neural network (for example, three layers of an input layer, an intermediate layer [hidden layer], and an output layer) using a back propagation as a learning algorithm, and a variable weight W. It is a connection (body) of a number of units (neurons, also called synapses) connected by. The unit U is a conversion system having a non-linear input / output characteristic, and the total sum from the front layer obtained by multiplying the output value from the front layer by an independent weight W is input to the unit U, and the sum is nonlinear. It is converted and output to the subsequent layer. In the neural network N, the output signal output with respect to the (known) input signal is compared with the teacher signal, and the variable weight W representing the strength of the coupling between the units is changed / converged, that is, learned and desired. This neural network is constructed to have a waveform processing function.

(Structure of Neural Net N and Unit U) Next, the neural net N and the unit U will be described in detail. As described above, the neural network N shown in FIG. 2 is a layered neural network composed of an input layer, an intermediate layer, an output layer, etc., which are composed of a large number of units (U). The unit U is a conversion system that has a nonlinear input / output characteristic that is an engineering model of a neuron in the brain, and the neural network N is basically a combination (body) of many units connected by variable weights. is there. FIG. 3 is a diagram showing a unit connection of the neural network shown in FIG. The unit U is a conversion system configured by a nonlinear function F (X) that outputs an output value Y = F (X) for an input value X, for example, a nonlinear input / output characteristic. As the input value, the sum obtained by multiplying the output value of the previous layer (or the input value to the neural network) by the independent variable weight W is input, and this input value is nonlinearly converted and It is output to the layer.

More specifically, FIG. 3 is a diagram showing the i-th unit combination of the n-th layer constituting the layered neural network, and the input value X (n to the i-th unit of the n-th layer is shown. , I) is a weight W that represents the strength of the coupling between the j-th unit of the (n−1) -th layer and the i-th unit of the n-th layer in the output value of each unit of the (n−1) -th layer that is the previous layer. (N, i, j)
It is configured to be the sum of those multiplied by. The input value X (n, i) is input to the non-linear function F (X), non-linearly converted to Y (n, i) = F {X (n, i)}, and output to the unit in the next layer. Is configured to.

The variable weight W (n, i, j) represents the coupling strength of the unit (U).
A desired neural network is constructed by changing / converging (n, i, j), that is, learning. As a learning algorithm for constructing a neural network, there is known back propagation. That is,
In the neural network N shown in FIG. 2, when an input value (input pattern) is input to the input layer, this input value is transmitted / processed to the input layer → the intermediate layer → the output layer, and the output value (output pattern) is output from the output layer. ) Is output (feedforward processing). This output value is based on the weight obtained by the learning so far.

On the other hand, a desired output value (teacher signal)
Is given, output layer → middle layer → input layer
It is processed and the weight is learned (feedback process). The weight learning is to change and converge the coupling strength between units, that is, the weight W (n, i, j) so as to reduce the difference between the actual output value and the desired output value.

Next, based on the basic concept of the waveform processing processing by the neural network N described above, a specific configuration of the waveform processing apparatus 100 by the neural network according to the present invention (a hardware-like execution portion by a substantial signal processing apparatus). , Ie, the desired signal processing system with fixed weights),
A specific method of designing a waveform processing device using a neural network (a software execution part by a computer, that is, a pre-learning processing method using variable weights) will be described in detail. FIG. 1 is an overall configuration diagram (a hardware-like execution portion) of a waveform processing device using a neural network. FIG. 1A shows a substantive processing state after design learning, and FIG. 1B shows design learning. 3 is a diagram (software-like execution portion) showing the processing state of FIG.

(Substantial Processing System of Waveform Processing Apparatus 100 by Neural Network N) First, with reference to FIG. 1A, a substantive processing system after design learning will be described. In the figure, 100 is a waveform processing device to which an original signal is input. In this processing system, a plurality of non-linear conversion systems are connected by independent fixed weights, and
The fixed weight is composed of the conversion means, that is, the waveform processing device 100, which is a weight previously learned by the neural network N with respect to the characteristics of the recording and transmission system. Then, the current signal which is deteriorated by the recording and transmitting system mainly composed of the recording and transmitting medium 102 is inputted to the waveform processing device 100,
The original signal is waveform-processed in advance so as to cancel the deterioration caused by the recording / transmission system, and then output to the recording / transmission medium 102 via the recording / transmission device 101. The signal from the recording / transmission medium 102 is reproduced and received by the reproduction / reception device 103 to be an output signal. The recording / transmission device 101 and the reproduction / reception device 103 correspond to the recording / transmission medium 102. When the recording / transmission medium 102 is a magnetic recording medium or an optical recording medium, the recording / transmission device 101 is a magnetic / optical recording / reproducing device.
When 02 is a transmission line, it is a transmission / reception device.

(Structure of Waveform Processing Device 100 by Neural Net N) Next, the neural network N according to the present invention will be described.
A specific configuration (hardware execution part) of the waveform processing apparatus 100 according to the above will be described in detail. FIG. 4 shows an operational amplifier,
Figure 2 is a concrete configuration of a diode pair and a resistor.
4 is a circuit diagram of a waveform processing device equivalent to the neural network N shown in FIG. The waveform processing apparatus 100 is a substantial signal processing apparatus having a fixed weight, which executes only desired signal processing (feedforward processing) at high speed, and is composed of a simple circuit having no learning function. It is a thing.

As shown in FIG. 3, the unit (U) is a general-purpose inverting operational amplifier 1a having a predetermined inverting amplification characteristic.
1b, 2a · 2b, 3a · 3b, 4a · 4b and buffer amplifiers 5, 6, 7, 8, 9 having predetermined amplification characteristics
And a diode pair (10, 12, 13) having non-linear input / output characteristics (junction characteristics). The diode pair is
The diode clip circuit is connected in parallel in the opposite direction between the signal line and the ground line and has a nonlinear input / output characteristic (junction characteristic).

Resistances Rji, Rj (R1 between units (U)
1, R21, R31, R41, R51, R12, R22, R32, R42, R52, R13, R23, R33,
R43, R53, R1, R2, R3) function as a fixed weight that represents the strength of the bond between the units. That is, the weight W (2, i, j) corresponds to the resistance Rji, and the weight W (3,1, j) corresponds to the resistance Rji.
Corresponds to j. The resistance values of the fixed resistors Rji and Rj are values determined by the design method described later. The weight (coefficient) representing the strength of the coupling between the units has positive and negative values as described later, but for the negative weight, the inverting operational amplifier is used alone (for example, resistor R21 and The configuration of the inverting operational amplifier 1a) is inverted, and for positive weighting, two inverting operational amplifiers are used in series (for example, the configuration of the resistor R31, the inverting operational amplifier 1b, and the inverting operational amplifier 1a) to regenerate the signal. It corresponds by reversing. Reference numerals 14, 15, 16 and 17 are delay lines having a predetermined delay time.

The input signal (original signal) is given the predetermined delay time by delay means (delay lines) 14, 15, 16 and 17 and input to the input layer (of the neural network). A plurality of input waveforms (values) temporally before and after were amplified by the respective buffer amplifiers 5, 6, 7, 8, 9 which are units U '(1,1) to U' (1,5) in the input layer. It is later divided and input to the unit of the intermediate layer via a resistor.

As the input value to the unit, the output value of the input layer which is the front layer is the resistance Rji (R11, R21, R31, R41, R51, R12, R22, R32, R42 which functions as an independent fixed weight. , R5
2, R13, R23, R33, R43, R53), and is then weighted via the operational amplifier 1a constituting the units U (2,1), U '(2,2), U (2,3) in the intermediate layer. Input to the inverting input terminals of 1b, 2a, 2b, 3a, and 3b, and they are added and combined. The non-inverting input terminal of the operational amplifier is grounded. The input value amplified (inverted or re-inverted) by the operational amplifier is nonlinearly converted by the diode pair 10 and 12 and output to the output layer which is the rear layer. Since the diode pair has a non-linear input / output characteristic, the non-linear conversion is performed and the output is made.

The output from the intermediate layer is input to the operational amplifier 4a (and 4b), which is the unit U (3,1) of the output layer, and the diode pair 13 through the resistor Rj (R1, R2, R3), Similar to the above-mentioned unit of the intermediate layer, it is non-linearly converted and output as an output signal.

(Neural network waveform processing apparatus 1
00 design method) Next, a method of designing the waveform processing device 100 by a neural network (software-executed portion by a workstation) will be described. The waveform processing apparatus (100) targeted by this design method is the circuit described in FIG. The method of designing a waveform processing device using a neural network is based on the method shown in FIG.
Resistances Rji, Rj (R11, R21, R31, R of the waveform processing apparatus 100
41, R51, R12, R22, R32, R42, R52, R13, R23, R33, R43, R53, R1
, R2, R3) values (fixed weights) are obtained by learning a neural network (neural network N shown in FIG. 2) equivalent to the waveform processing apparatus 100, and weights W (2, i, j), W
It is specifically determined by (3,1, j), that is, by the variable weight that has changed / converged by learning. In addition,
The weight representing the strength of the coupling between the units has positive and negative values as described above, but for the negative weight, the inverting operational amplifier is used alone to invert the signal, and for the positive weight, Two inverting operational amplifiers are used in series to re-invert the signal to correspond.

As shown in FIG. 1B, the reproduction / reception device 1
The output signal d reproduced by 03 is A / D converted and taken into a digital memory (not shown) at a predetermined sampling period. This output signal d is (waveform processing device 1
Neural network N (equivalent to 00) → recording and transmitting device 10
1 → recording / transmission medium 102 → reproducing / receiving device 103, which has been subjected to waveform processing in advance by the neural network N, and further has a recording / transmission system (recording / transmission device 101 / recording / transmission medium 1).
02 / playback / reception device 103, etc.). The data taken in the digital memory is transferred to a work station (not shown) for learning processing, and learning calculation processing described later is performed. The sampling period is
It is equal to the learning calculation processing interval or shorter than the learning calculation processing interval.

Next, a learning calculation process is performed by the workstation. A neural network simulation program having a known back-propagation learning algorithm is prepared in the workstation, and is shown in FIG. 2 equivalent to the concrete (neural net) waveform processing apparatus 100 shown in FIG. Learning calculation processing is executed for the neural network N shown.

That is, as for the captured data which is captured in the digital memory and transferred to the workstation, the original signal a is input to the neural network N equivalent to the waveform processing device 100, and this input signal is input layer → intermediate layer → This is an output signal that has been transmitted / arithmetically processed in the output layer, output from the output layer, and deteriorated by the recording and transmission system. This output signal is based on the weight obtained by the learning so far. The output signal is compared with the teacher signal (teacher data consisting of the original signal a that has been fetched into the workstation in advance), and is transmitted / arithmetically processed in the order of output layer → intermediate layer → input layer for learning. The deterioration characteristics of the recording / transmission system to be learned (learned) include the characteristics of the recording medium, the characteristics of the transmission line, the characteristics of the recording apparatus, the characteristics of the reproducing apparatus, the characteristics of the data itself, etc., depending on the size of the neural network. Is appropriately selected.

Then, the weight of the neural network N that has converged after learning the characteristics of the transmission / reproduction system is the waveform processing device 100.
Resistance of Rji, Rj (R11, R21, R31, R41, R51, R12, R22, R3
2, R42, R52, R13, R23, R33, R43, R53, R1, R2, R3). Finally, as shown in FIG. 4, the waveform processing device 100 is manufactured based on the specifically determined resistance value and the weight representing the strength of the inverting operational amplifier.
Then, as shown in FIG. 1A, the recording transmission medium 102
A current signal that deteriorates due to the above is input to the waveform processing device 100, and the original signal is subjected to waveform processing so as to cancel the deterioration due to the recording and transmission system, and then output to the recording transmission medium 102 through the recording and transmitting device 101. As a result, even if deterioration occurs in the recording and transmission system, it is canceled and a signal close to the original signal is obtained from the reproducing / receiving apparatus 103.

(Specific Example) A magnetic recording medium was used as the recording / transmission medium 102, and a magnetic recording / reproducing apparatus was used as the recording / reproducing apparatuses 101 and 103 for learning design. FIG. 5 shows each waveform before learning, and FIG. 6 shows each waveform after learning. Waveform a
Is the original signal, waveform b is the output of the neural network N equivalent to the waveform processing device 100, and waveform c is the recording and transmission medium 102.
, Waveform d is the output of the recording / reproducing apparatus 103 via the recording / transmission medium 102. At the beginning of learning,
The neural network N is initially set to output an input signal almost as it is.

Before learning as shown in FIG. 5, as shown in FIG. 5C, both peaks are shifted by the intersymbol interference so that the output signal d is deteriorated. On the other hand, after the learning shown in FIG. 6, as shown in FIG. 6B, learning output is performed so that the interval between the output signals b of the neural network N equivalent to the waveform processing apparatus 100 becomes narrower, and as a result, As the output signal d, a waveform close to the current signal was obtained.

The waveform processing apparatus 100 (and the neural network N equivalent thereto) is connected to the recording and transmitting apparatus 101.
It may be provided inside (between the recording transmission medium 102).

[0035]

As described above in detail, since the present invention is a waveform processing apparatus by a neural network in which waveform processing by a neural network is performed on the recording / transmission apparatus side, and a designing method thereof, a reproducing / receiving apparatus as in the prior art is used. There is no need for a number of waveform equalizers. Therefore, a large-scale neural network can be used to provide a high-performance device. Moreover, since the substantial waveform processing device is designed by the neural network, the design is easy.

[Brief description of drawings]

FIG. 1 is a system configuration diagram showing an embodiment of a waveform processing device using a neural network according to the present invention, (A)
FIG. 4 shows an example of a physical signal device that is executed by hardware, and (B) is a diagram showing a method for designing a waveform processing device by a neural network.

FIG. 2 is a basic conceptual diagram showing an embodiment of a waveform processing device using a neural network.

FIG. 3 is a diagram showing a unit connection of the neural network shown in FIG.

4 is a specific circuit diagram of the waveform processing device using the neural network shown in FIG.

FIG. 5 is a diagram showing each signal before learning by a neural network.

FIG. 6 is a diagram showing each signal after learning by a neural network.

FIG. 7 is a block diagram showing a waveform processing device using a conventional neural network.

[Explanation of symbols]

DL delay means, N neural network U unit (non-linear conversion system), W variable weight, 1a • 1b, 2a • 2b, 3a • 3b, 4a • 4b inverting operational amplifier 10, 12, 13 Diode pair 14, 15, 16, 17 Delay line 20 Digital memory, 21 Workstation 100 Waveform processing device, 101 Recording transmission device, 102 Recording transmission medium, 103 Playback receiving device, R11, R21, R31, R41, R51 Resistance to the first unit of the intermediate layer (Fixed weight) R12, R22, R32, R42, R52 Resistance to the second unit in the middle layer (Fixed weight) R13, R23, R33, R43, R53 Resistance to the third unit in the middle layer (Fixed weight) ) R1, R2, R3 Resistance to (the unit of) the final layer (fixed weight) a Original signal, b Output of neural network N equivalent to waveform processing apparatus 100, c Output from recording transmission medium 102, d Recording transmission The output of the recording and reproducing device 103 via the medium 102.

Claims (3)

[Claims] 1. A plurality of conversion systems composed of non-linear conversion systems are connected by independent fixed weights, and these fixed weights are weights learned in advance for the characteristics of the recording transmission system by a neural network. The present invention comprises a conversion means, wherein a current signal deteriorated by the recording and transmission system is inputted to the conversion means, the original signal is subjected to waveform processing so as to cancel the deterioration by the recording and transmission system, and output to the recording and transmission system. A waveform processing device using a neural network. 2. A plurality of conversion systems, each of which is composed of a non-linear conversion system, is composed of independent conversion means connected by fixed weights, and has a waveform for canceling the deterioration of the current signal deteriorated by the recording transmission system. A method of designing a waveform processing device for processing and outputting to a recording / transmission system, wherein a plurality of conversion systems which are equivalent to the waveform processing device and which are non-linear conversion systems are connected with independent variable weights. The weight is variably learned according to the characteristics of the recording and transmission system by a neural network, and a fixed weight of the conversion means is designed and determined based on the variably learned weight. A method for designing a waveform processing device using a neural network. 3. A neural system according to claim 1, further comprising means for delaying an original signal which is deteriorated by a recording and transmission system, and the current signal which is temporally moved forward and backward is inputted to the converting means. A method of designing a waveform processing apparatus using a net or a waveform processing apparatus using a neural network according to claim 2.