JPH05307622A - Waveform equalizer - Google Patents

Abstract

PURPOSE:To provide an inexpensive waveform equalizer capable of being applied also to a case having plural elements exerting influence upon the characteristics of a transmission reproducing system and appropriate for reducing the bad influence of these elements. CONSTITUTION:In the waveform equalizer 30 provided with a neural network 20 consisting of plural conversion units having non-linear I/O characteristics and connected by individual connection factors and plural signal delay circuits 41 to 44 for delaying signals, I/O signals from the circuits 41 to 44 to which signals deteriorated by a transmission/reproducing system are inputted are inputted to the network 20 and signals corresponding to the state of elements exerting influence upon the characteristics of the transmission/reproducing system are also inputted to the network 20. Since plural signal deterioration factors can be removed by the single neural network 20, the deterioration of signal quality can be extremely effectively improved at a low cost.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a waveform equalizer for reducing signal deterioration caused by signal recording / reproduction and transmission.

[0002]

2. Description of the Related Art An example of a conventional magneto-optical disk device will be described with reference to FIG. FIG. 4 is a diagram showing an example of a conventional waveform equalizer. The waveform equalizer shown in the figure is composed of a signal delay circuit including four delay elements 41 to 44 and a neural network (hereinafter referred to as NN) 40. The delay element has a delay time set individually, and the NN 40 includes an input layer including a buffer amplifier and the like, an intermediate layer, and an output layer. The neural network shown in FIG. 4 is described in detail in Japanese Patent Application No. 31952 of 1991 previously filed by the present applicant, and the outline thereof will be described below with reference to FIGS. 4 and 5.

FIG. 4 shows a conceptual diagram of a waveform equalizer including a signal delay circuit composed of four delay elements 41 to 44 in addition to the neural network 40 composed of hardware. The waveform equalizer is, for example, a circuit for processing a rectangular-wave-shaped input signal si having a waveform distortion by NN so as to obtain an output so as close to a desired rectangular wave as possible. The waveform equalizer can be realized without using the NN, but it has been experimentally confirmed that a very good waveform equalizer can be realized by using the NN.

In FIG. 4, the NN 40 includes an input layer composed of buffer amplifiers 51 to 55 and a neuron 61.
˜63 and an output layer composed of the neuron 71, the input / output signals to / from the delay elements 41 to 44 are also input to the buffer amplifiers 51 to 55, respectively. The output of is passed through three individual resistors having a resistance value determined by the synapse coupling coefficient, and is input to the neurons 61 to 63 in the intermediate layer.
In the figure, a small circle indicates the existence of this synaptic coupling coefficient,
The symbol W indicates the value of this synaptic coupling coefficient.

FIG. 5 is a diagram showing a unit connection of the neural network shown in FIG. In FIG. 4, the unit U
Is a conversion system having a nonlinear input / output characteristic that is a model of a brain neuron, and the neural network 40 is a combination of units connected by a variable synapse coefficient (weight) W. However, the unit (U) is a unit of the input layer and is composed of a buffer amplifier or the like. In FIG. 5, U is a nonlinear input / output characteristic, for example, a nonlinear function F (X) that outputs an output value Y = F (X) for an input value X.
Is a conversion system configured by. As the input value, the sum of the results obtained by multiplying the output value of the previous layer (or the input value to the neural network) by the independent variable synapse coupling coefficient W is input, and this input value is nonlinear. It is converted and output to the subsequent layer.

More specifically, FIG. 5 is a diagram showing the i-th unit connection of the n-th layer forming the layered neural network, and the input value X to the i-th unit of the n-th layer is shown.
(N, i) is 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 of each unit of the (n-1) -th layer which is the previous layer. Weight W (n, i, j)
It is configured to be the sum of those multiplied by. Then, 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.

The variable weight W (n, i, j) represents the coupling strength of the units (U), and the weight W (n, i, j) is used in the simulation and learning of the present NN. ) Is changed and converged to construct the desired neural net. Here, the synapse coupling coefficient (weight) is predetermined by learning of the back propagation method with the ideal signal as a teacher signal and the degraded input signal si as a student signal. The deterioration of the input signal si is, for example, deterioration in a recording / reproducing system or deterioration occurring when passing through another transmission system.

Since the synapse coupling coefficient has a positive and a negative, each neuron is configured with hardware so as to correspond to the positive and negative synaptic coupling coefficients, respectively. Since this NN has a small number of neurons, it can be easily realized as hardware. FIG. 6 is a diagram showing an example of the hardware NN. By the way, the determination of the synapse coupling coefficient W is performed by computer simulation and learning. At this time, the delay element and the neural network shown in FIG. 4 are configured by software.

In learning, first, a deteriorated signal (student signal) to be subjected to waveform equalization is taken into a digital memory, transferred to a computer for simulation, and output processed by a neural network simulation program in the computer. In order to reduce the difference between the signal (output signal) and the desired output value (teacher signal), the synapse coupling coefficient (weight), which is the strength of coupling between neural networks, is changed by software and converged. The result is configured in hardware to obtain a circuit that executes desired signal processing (feedforward processing) at high speed. However, in the above method, the synapse coupling coefficient is calculated by processing the student signal first captured in the digital memory so as to approach the teacher signal, and the synapse coupling obtained when the signal waveform of the student signal is different. The value of the coefficient will be different.

Generally, a recording / reproducing system or a transmission system of equipment is
There are various instability factors and often deviate from the ideal state. For example, in the case of an optical disc device, the reproduction waveform may differ due to the optical axis of the optical pickup deviating from the vertical state (hereinafter referred to as tilt), the aberration of the pickup itself, and the residual error of the servo system. Is deviated from the ideal state. Further, the student signal is different because the principle of recording / reproducing is different depending on the type of optical disk medium, for example, ROM (read only) and RAM (recording / reproducing). Therefore, it is impossible to uniquely determine the optimum synaptic coupling coefficient for all the unstable elements.

[0011]

By providing a waveform equalizer which is applicable to a case where there are a plurality of factors that affect the characteristics of a transmission / reproduction system and is suitable for reducing the adverse effects of the above-mentioned factors and which is inexpensive. is there.

[0012]

A neural network including a plurality of conversion units having nonlinear input / output characteristics and coupled by individual coupling coefficients obtained by pre-learning, and a signal delay circuit for delaying a signal. In the waveform equalizer including, the input / output signal of the signal delay circuit, which receives the signal deteriorated in the transmission / reproduction system, is input to the neural network, and the factor affecting the characteristics of the transmission / reproduction system. It is a waveform equalization device that inputs a signal according to a state to the neural network and obtains a waveform equalized signal from the output of the neural network. In the waveform equalizer, a signal identification circuit that identifies a plurality of equalization target signals, and the plurality of equalization target signals are switched and input to the signal delay circuit according to the output of the signal identification circuit. And a means for supplying a DC signal to one of the units of the neural network according to the output of the signal identifying circuit. In addition, an input layer composed of a plurality of units, an intermediate layer composed of a plurality of conversion units each having a nonlinear input / output characteristic and each unit of front and rear layers and combined with individual coupling coefficients obtained by pre-learning, In a waveform equalizer having a signal delay circuit that delays a signal using a neural network composed of an output layer that outputs a signal that is output, the signal that has deteriorated in the transmission and reproduction system is input to the input layer of the neural network. It is a waveform equalization device in which a neuron, which is input and is coupled to each unit of the intermediate layer and the output layer of the neural network by individual coupling coefficient, is provided outside the neural network.

[0013]

According to the present invention, the signal corresponding to the state of the element affecting the characteristics of the transmission / reproduction system is used as one of the student signals together with the main signal to be equalized to configure the waveform equalizer. The waveform equalizer is designed to be input to the NN, and at the time of simulation and learning of the NN, the main signal is changed by changing the amount of elements that affect the characteristics of the transmission / reproduction system within a range that can actually occur. Captured as a student signal,
At the same time, the quantities of the elements are also stored in the digital memory as student signals. Then, a simulation is performed on the computer using both of the student signals to obtain the optimum value of the synapse coupling coefficient of the neural network in the configuration, so that it is optimal even for the variation of the element. A waveform equalizer for equalization is obtained.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the waveform equalizer of the present invention will be described below with reference to FIG. FIG. 1 is a diagram showing a first embodiment of a waveform equalizing device of the present invention, which is applied to a read-only optical disk device. As shown in FIG. 1, in the first embodiment of the waveform equalizer of the present invention, the reflected light from the optical disc 3 is condensed by the condenser lens 9 and is incident on the photodetector 10. A reproduction signal (hereinafter, referred to as si) obtained from the output of the photodetector and a tilt signal (hereinafter, referred to as simplification) detected and amplified by the tilt sensor LC that detects the tilt of the optical pickup 1 and the optical disc 3.
It is written as sc. ) Is input to the signal delay circuit constituting the waveform equalizer 30 and the input layer of the neural network 20, respectively.

In FIG. 1, the optical pickup 1 is used for recording information on the optical disc 3 and reading information from the optical disc 3. Laser light generated from the light source 4 in the optical pickup 1 is collimated into a parallel light flux through a collimator lens 5, and is converged through a beam splitter 6, a quarter wavelength plate 7 that changes the polarization state, and an objective lens 8. A beam spot P is formed on the reflecting surface of the disk 3. The reflected light from the optical disc 3 is the objective lens 8,
It is incident on the beam splitter 6 through the quarter wave plate 7,
The light is changed in direction and separated from the incident light, and is incident on the condenser lens 9. The light flux from this condenser lens 9 is detected by the photodetector 1.
The signal is incident on 0, the output signal of the photodetector 10 is applied to the signal detection circuit 12, and the reproduction signal si is output from the signal detection circuit 12 and input to the waveform equalizer 30.

In the vicinity of the objective lens 8, a tracking actuator TA for causing the beam spot P to follow a track on the disk 3, and the objective lens is controlled so that the beam spot P forms an image on the disk surface. Focusing actuator FA, optical disk 3
A tilt sensor LC for detecting the angle formed by the optical axis of the laser beam applied to the disk and the disk surface is provided. The output of the tilt sensor LC is applied to the delay element 11, and the output sc of the delay element 11 is input to the NN 20.

The waveform equalizer 30 includes delay elements 41-41.
The signal delay circuit is composed of 44 and the NN 20. The waveform equalizer 30 differs from the waveform equalizer shown in FIG. 4 in the structure of the neural network. In the NN shown in FIG. 4, the input layer is composed of units of 51 to 55, but in the NN 30 shown in FIG. 1, the input layer is composed of units of 51 to 56, and the units 51 to 55 have waveforms. A signal corresponding to the input signal si of the equalizer is input, but the tilt signal sc is input to the unit 56, and the unit 56 and the units 61 and 6 of the intermediate layer are input.
2 and 63 are coupled with each other by a synaptic coupling coefficient. Each unit of the input layer and the intermediate layer is coupled with an appropriate synaptic coupling coefficient, but the small circles indicating them are not shown in the drawing. Similarly, each unit of the intermediate layer and the output layer is also connected with an appropriate synaptic connection coefficient (weight) W.

The unit U and the synaptic coupling coefficient W are
As described with reference to FIG. 4, the unit U is a conversion system having a nonlinear input / output characteristic in which a brain neuron is modeled as an engine, and the neural network 30 is a combination of many units connected with variable weights W. It is the body. The unit (U) indicates a unit in the input layer, and a buffer amplifier or the like corresponds thereto. The unit U has a non-linear input / output characteristic, for example, an output value Y = F for an input value X.
It is a conversion system constituted by a non-linear function F (X) that outputs (X). The sum of the results 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 to the unit U, and the input value is subjected to the nonlinear conversion. It is output to the subsequent layer.

Here, the synapse coupling coefficient is determined in advance by computer simulation and learning of the back propagation method using a desired signal as a teacher signal. In the present invention, the simulation and learning for determining the hardware of the NN are performed by the signal delay circuit constructed by software and the NN constructed by software in the computer outside the magneto-optical disk device. Input / output signals to each delay element are also simultaneously input to each unit of the NN input layer, and a predetermined amount of tilt is applied between the disc and the disc device.

Here, a rectangular signal recorded on the disk, that is, a signal obtained by removing high-frequency components from the original signal by a low-pass filter is used as a teacher signal. Student signals are the signals si and sc. Then, the output signal so of the NN is compared with the teacher signal, and each synapse coupling coefficient in the NN is corrected based on the deviation amount. The process is performed on a number of data to determine synaptic coupling coefficients. On the basis of the determined synapse coupling coefficient, a hardware NN is provided in the optical disc device. In the magneto-optical disk device of the present invention, the NN itself incorporated in the magneto-optical disk device does not have a learning function in order to make the NN compatible with high-speed processing.

Now, the main points of the design of the NN 30 will be described. In general, a back-propagation NN design requires a teacher signal and a student signal. When NN is applied to a signal transmission system such as an optical disk, the teacher signal is an original signal whose high frequency range is cut off, and the student signal is a signal deteriorated in the transmission system. The tilt amount is also captured as one of the student signals. By the way, if there is a tilt in the positional relationship between the optical disc and the optical pickup, a coma aberration occurs in the optical system, the shape of the beam spot is distorted by the coma aberration, the spot size increases, and the central intensity of the spot decreases. , Etc., especially when there is a tangential tilt due to the surface wobbling component in the tangential direction of the track, intersymbol interference increases, the quality of the reproduced signal deteriorates, and in extreme cases. , A read error occurs.

Therefore, here, the tilt means the tangential tilt. In learning the NN, the tangential tilt amount is used as one of student signals. As a result, in an actual disk device, the influence of the tangential tilt is reduced by the NN, intersymbol interference and the like are reduced, and a powerful waveform equalized output is obtained. Further, the frequency of the tangential tilt is considerably higher than the radial tilt caused by the surface wobbling component in the radial direction of the disc, and it has been impossible to correct the tangential tilt in the past. However, according to the waveform equalizer of the present invention, since waveform equalization is performed without moving a mechanical part having a large mass, a strongly waveform equalized output can be obtained over a wide frequency band.

Next, a second embodiment of the waveform equalizer of the present invention will be described with reference to FIG. The second embodiment of the waveform equalizer of the present invention is an application of the waveform equalizer of the present invention to a magneto-optical device. FIG. 2 is a diagram showing a second embodiment of the waveform equalizer of the present invention. In FIG. 2, the magneto-optical disk 3
In A, a large number of circular information tracks are provided.
In this track, a read-only ROM section and a recording / playback RAM section are mixed. The ROM section and RA
The recording method and the reproducing method of the M portion are different, and therefore, the magneto-optical disk 3A and the disk device shown in FIG. 2 are different from the read-only optical disk 3 and the optical disk device shown in FIG. Since the shown magneto-optical device has a general configuration, detailed description thereof will be omitted.

In the figure, recording of information on a magneto-optical disk medium 3A having a number of information tracks provided in a circular pattern is performed by a magneto-optical recording method using an optical pickup 1A. Further, the reading of the information signal from the magneto-optical disk medium is also performed by using the optical pickup 1A. The optical pickup 1A includes a light source 4A such as a semiconductor laser, a collimator lens 5 that converts a light beam from the light source 4A into a parallel light beam, an irradiation light beam from the light source 4 such as the semiconductor laser, and a reflected light beam from the optical disc medium 3. The beam splitter 6 for separating is provided.

In FIG. 2, P is a light spot condensed by the objective lens 8, and the optical pickup 1A is also provided with a tracking actuator TA and a focusing actuator FA. The reflected light from the magneto-optical disk medium 3A is reflected by the beam splitter 6
It is reflected by and enters the beam splitter 15. The light beam reflected by the beam splitter 15 enters a condenser lens 16, is condensed, and enters a photodetector 17. A tracking error signal and a focusing error signal are detected from the output S3 of the photodetector.

The light flux that has passed through the beam splitter 15 is incident on the half-wave plate 19, the plane of polarization is changed, and is incident on an analyzer 23 such as a polarization prism. The light flux is split into two by the polarization prism 23, and the condenser lens 2
Approximately equal amounts of light flux enter the condenser lens 24 and the condenser lens 24. The light flux that has passed through the condenser lens 26 enters a photodetector 27,
The information signal S1 is output to the output terminal of the photodetector 27. The light flux passing through the condenser lens 24 is incident on the photodetector 25, and the information signal S2 is output to the output terminal of the photodetector 25.

The two outputs of the optical pickup 1A, that is, the information signal S1 and the information signal S2 are both outputs according to the information recorded on the magneto-optical disk medium 3A. These two signals are input to the differential amplifier 28, and the differential output thereof is taken out as the magneto-optical signal S4. The differential output S4 is used as a reproduction information signal from the RAM section of the optical disc 3A. Also, the two signals S
1 and S2 are input to the summing amplifier 29, and the output of the summing amplifier 29 is taken out as the pre-pit signal S5.
The output S5 of this summing amplifier is the magneto-optical disk 3A.
It is used as a reproduction information signal from the ROM section of the.

The signals S4 and S5 are input to the switch 32. The magneto-optical / pre-pit detection circuit (signal discrimination circuit) 31 discriminates whether the signal currently read by the optical pickup 1A is the signal of the ROM section or the signal of the RAM section, and its output S8 is the switches 32, 33. Is being applied to. The output signal sia of the switch 32 becomes the signal S4 or S5 according to the signal S8 and is applied to the waveform equalizer 30. The waveform equalizer 30 shown in FIG. 2 is similar to the waveform equalizer shown in FIG. 1, and is composed of a signal delay circuit and an NN 20 constructed by hardware. The signal sia corresponds to the signal si shown in FIG.

In FIG. 2, the switch 33 has a terminal 3
The DC signal S6 is applied from No. 4 and the signal sca supplied from the switch 33 to the NN 20 is zero or S6 according to the signal S8. And the signal sca
Corresponds to the signal sc shown in FIG. In the waveform equalizer 30, the DC signal sca applied to the neural network depends on whether the waveform equalization target is the signal S4 or the signal S5.
However, the characteristics of the NN can be changed by switching to zero or S6. The output signal S7 of the NN 20 is strongly waveform-equalized according to the waveform equalization target.

As described above, according to the second embodiment of the waveform equalizing apparatus of the present invention, when reproducing a disc having a different reproducing principle, the kind of the signal to be reproduced is identified and a part of the input layer of the NN is detected. A DC signal corresponding to the type of the disk is supplied to switch the operation of the NN. For example, the ROM of this example
When the reproduction principle is different between the NN section and the RAM section, the state of intersymbol interference also differs, and the optimum synapse coupling coefficient between NN neurons also differs. In order to solve this problem, there is also a method of using different NNs for each waveform equalization target, but in this embodiment, a single NN obtains good results.

Next, the simulation and the method of determining the synaptic coupling coefficient in this example will be briefly described. First, N
The synaptic coupling coefficient of N is determined by the method of the first embodiment.
However, in the second embodiment, the student signals are only student signals, and there are two types of student signals. Learning is done in two stages. In the simulation and learning, the pre-pit reproduction signal S5 and the magneto-optical reproduction signal S4 are used as student signals.
One of them, for example, the pre-pit reproduction signal is used as a basic student signal. As the teacher signal, a recording original signal blunted by a low-pass filter is used.

As a first stage of learning, the DC signal S6
Is set to zero, the basic student signal is learned by the back propagation method, and the synapse coupling coefficient for the pre-pit reproduction signal is determined. Next, in the second step, the student signal is converted to a magneto-optical signal, the DC signal S6 is appropriately supplied, and the coupling coefficient between the NN unit 56 shown in FIG. 1 and the intermediate layer is determined. In this case, the synapse coupling coefficient determined in the first step is not changed as it is. However, if the coupling coefficient does not converge in the learning in the second step, the value of the DC signal S6 is changed and the synapse coupling coefficient is changed again. Learning is done. Alternatively, the basic student signal is switched to the magneto-optical reproduction signal S4 and the process is restarted from the first stage.

NN actually incorporated in the optical disk device
Is a hardware implementation of the NN having the above-described structure. The input signal is not limited to the combination of the pre-pit reproduction signal and the magneto-optical reproduction signal of this example, and other combinations and combinations of more signals can be used for a plurality of neurons and DCs corresponding thereto. It suffices to prepare a signal and use it by switching with a switching means such as a switch.

Next, a third embodiment of the waveform equalizer of the present invention will be described with reference to FIG. FIG. 3 is a diagram showing a third embodiment of the waveform equalizer of the present invention. In this example, signals other than the main signal SM that is the target of waveform equalization, that is, a neuron to which the auxiliary signal SS that assists waveform equalization is supplied,
The external neuron is arranged outside the neural net 40 to which a signal corresponding to the main signal SM is input, and the external neuron and the neurons of the intermediate layer and the output layer of the NN 40 are connected by synapses having different synapse coupling coefficients. To be done. Due to the structure, NN40 is controlled by the external neuron. In FIG. 3, a main signal SM is a reproduction signal to be subjected to waveform equalization, an auxiliary signal SS is a tangential tilt signal in the first embodiment, and a DC signal in the second embodiment is waveform equalization. It is a signal that assists. The external neuron 58 has a structure different from the neural nets shown in the first and second embodiments in that it has a synapse directly connected to the output layer without passing through the intermediate layer.

At the time of learning, as in the first embodiment, the main signal and the auxiliary signal are learned at the same time to determine the synapse coupling coefficient, and as in the second embodiment, first, NN synapses with respect to the main signal. In some cases, the coupling coefficient is determined, and after that, the main signal and the auxiliary signal are input to perform learning to determine the synaptic coupling coefficient from the neuron of the auxiliary signal input unit. In this way, by adopting the structure in which the external neuron 58 is added for the auxiliary signal SS in addition to the original neural net 40, it may be possible to provide a waveform equalizing device with further excellent performance. Although the waveform equalizer of the present invention has been described mainly with respect to the application example to the optical disk device, the present invention can be implemented by combining the first embodiment to the third embodiment. Of course, it can be applied to other transmission systems.

[0036]

According to the waveform equalizer of the present invention, it is possible to remove a plurality of signal deterioration factors in a recording / reproducing system of a signal or a transmission system, which could not be removed by a single NN in the past. Deterioration can be remarkably improved at low cost.

[Brief description of drawings]

FIG. 1 is a diagram showing a first embodiment of a waveform equalizer of the present invention.

FIG. 2 is a diagram showing a second embodiment of the waveform equalizer of the present invention.

FIG. 3 is a diagram showing a third embodiment of the waveform equalizer of the present invention.

FIG. 4 is a diagram showing an example of a conventional waveform equalizer.

5 is a diagram showing unit connection of the neural network shown in FIG. 4. FIG.

FIG. 6 is a diagram showing a hardware neural network.

[Explanation of symbols]

 1, 1A ... Optical pickup 3 ... Magneto-optical disk 4 ... Laser light source 5 ... Collimating lens 6, 15, 23 ... Beam splitter 7 ... Quarter wave plate 8 ... Objective lens 9, 16, 24, 26 ... Condensing lens 10 , 17, 25, 27 ... Photodetector 12 ... Signal detector 19 ... 1/2 wavelength plate 23 ... Polarizing prism (analyzer) 20, 40 ... Neural net 30 ... Waveform equalizer 31 ... Signal identification circuit 32, 33 ... switch (switching means) 11, 41, 42, 43, 44 ... delay element LC ... tilt sensor P ... optical spot

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Office reference number FI technical display location H04B 3/14 8226-5K

Claims (3)

[Claims] 1. A waveform provided with a neural network having a plurality of conversion units which have nonlinear input / output characteristics and are coupled by individual coupling coefficients obtained by pre-learning, and a signal delay circuit for delaying a signal. In the equalizer, input / output signals of the signal delay circuit into which a signal deteriorated in the transmission / reproduction system is input to the neural network, and depending on the states of the elements that affect the characteristics of the transmission / reproduction system. A waveform equalizer, wherein a signal is input to the neural network and a waveform equalized signal is obtained from an output of the neural network. 2. A signal identification circuit for identifying a plurality of equalization target signals, and switching means for switching the plurality of equalization target signals according to the outputs of the signal identification circuits and inputting them to the signal delay circuit. 2. The waveform equalizer according to claim 1, wherein one of the units of the neural network includes a switching unit that supplies a DC signal in accordance with the output of the signal identification circuit. 3. An intermediate unit comprising an input layer composed of a plurality of units, and a plurality of conversion units having nonlinear input / output characteristics and connected to each unit of the front and rear layers by individual coupling coefficients obtained by pre-learning. In a waveform equalizer having a signal delay circuit for delaying a signal using a neural net composed of a layer and an output layer for outputting a converted signal, the neural network detects a signal deteriorated in a transmission reproduction system. A waveform equalizing device, wherein a neuron connected to each of the units of the intermediate layer and the output layer of the neural network with individual coupling coefficients is provided outside the neural network.