JPH0548391A - Adaptive equalizer - Google Patents

Abstract

PURPOSE:To equalize nonlinear distortion as well as linear distortion adaptively by detecting a characteristic of equalization object in terms of an input and an output signal and using the result so as to equalize the output signal. CONSTITUTION:An adaptive equalizer 20 equalizes the output signal of an equalization object 21 by using a characteristic detector 22 used to detect a transfer characteristic of the equalization object 21 such as a transfer function Homega) and the detected transfer function characteristic. And, the detector 22 and an equalizer 23 are neural networks, the detector 22 makes learning at the time of learning by using a transfer function of the equalization object 21 as a teacher signal and the equalizer 23 makes learning by using an input signal to the equalization object 21 as a teacher signal. Through the learning, the detector 21 detects the characteristic of the equalization object 21 from an input signal and an output signal of the equalization object 21. Then a control signal as a weight signal is outputted in response to the detected transfer function to form the equalizer 20 having the adaptive equalizing function and not only linear distortion but also nonlinear equalization is equalized.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an equalizer for compensating for signal deterioration on a transmission line in a communication system,
More specifically, the present invention relates to an adaptive equalizer that detects characteristics of a transmission line in wireless and wired communication and equalizes received signals using the detection result.

With the development of the information-oriented society in recent years, wired and wireless communication networks have been developed both in Japan and abroad, and information exchange by communication technology has become an important foundation of society. In the information society, conventional TV and radio broadcasts,
In addition to voice and image signals from telephones, information that has a great social impact, such as transaction information between companies and online information about banks, is being transmitted using communication networks.
For such information, there is a growing need to establish an accurate transmission means that reduces the error received due to the characteristics of the transmission path.

Particularly in digital mobile communications such as car telephones, a compensation technique for rapidly changing multipath fading has become an important issue. Therefore, there is a demand for establishment of an adaptive equalization technique that adaptively equalizes the deterioration of signals on the transmission line by adapting to the characteristic changes of the transmission line every moment.

[0004]

2. Description of the Related Art Generally, a wireless transmission path is composed of a transmission / reception filter, a modulator / demodulator, and a wireless device including a transmitter / receiver, and a propagation path whose characteristics change depending on weather conditions, buildings between the transmitter and the receiver, and the like. The causes of signal deterioration include linear and non-linear distortion generated from a device as a component, and two-wave interference fading as a kind of linear distortion generated from a propagation path. In mobile radio, distortion generated from a device or the like is stable over time, but propagation path distortion generated in a transmission line between a mobile station and a base station becomes selective fading and fluctuates over time. It is necessary to adaptively equalize the signal distortion on the transmission line by adapting to this fluctuation.

FIG. 62 is a conceptual diagram of mobile communication. In the figure, mobile communication is performed between the mobile station 1 and the base station 2. Fading occurs between the mobile station 1 and the base station 2 between the direct wave propagating directly between the stations and the reflected wave reflected by the obstacle 3 such as a building, mountain, or earth. Especially when a mobile station moves at high speed like car radio,
The amplitude ratio, propagation delay time, and phase difference between the direct wave and the reflected wave change rapidly. Therefore, it becomes necessary to follow this change at high speed to adaptively equalize the signal deterioration.

FIG. 63 is a block diagram of a transversal type equalizer as a conventional example of the equalizer. This transversal type equalizer demodulates a received signal from a transmission line to a demodulator 4
After being demodulated by, the delay line 5 with taps is sequentially set, the output from each delay device 6 is multiplied by the coefficient by the coefficient adjuster 7, and the result is added by the adder 8 to obtain the output. The distortion is removed by adjusting the coefficient of the device 7 according to the degree of the distortion inside the propagation path or the wireless device.

[0007]

However, as shown in FIG.
Although the conventional transversal type equalizer shown in Fig. 1 is effective for linear distortion, it cannot perform proper equalization when nonlinear distortion occurs in the transmission line or when a nonlinear circuit is used in the demodulation system. There was a problem.

Further, as a filter capable of dealing with non-linear distortion, an equalizer using a neural network has been conventionally known, but it takes a long learning time to make adaptive processing difficult, and a change in transmission line characteristic. Couldn't respond to. Therefore, there is a problem in that it is not possible to cope with the case where the transmission path characteristics change at high speed such as multipath fading in mobile communication.

An object of the present invention is to provide a function of adaptively equalizing not only linear distortion but also nonlinear distortion by detecting characteristics of a transmission line and performing equalization using the result. To do.

[0010]

1 to 4 are block diagrams showing the principle of the present invention. These figures are principle block diagrams of an adaptive equalizer that compensates an output signal by detecting characteristics of the transmission line from the output of the transmission line when a known input signal is input.

FIG. 1 is a principle block diagram of the first and third inventions. In the first aspect of the invention, the characteristic detection means 11 is a characteristic of the equalization target 10 that is to be equalized as compensation for deterioration of the output signal, for example, a characteristic of the equalization target 10 from an output signal with respect to a known input signal to the transmission line, for example, Detect the transfer function. The equalizer 12 uses the result detected by the characteristic detector 11 to equalize the unknown input signal 10
Output signal is equalized and the equalized output signal is output.

FIG. 2 is a block diagram showing the principle of the second invention. In the figure, the characteristic detection means 11 detects the characteristic of the equalization target 10 as in the first aspect of the invention. The equalization control means 14 outputs a control signal for equalizing the output signal of the equalization target 10 with respect to the unjoined input signal using the result detected by the characteristic detection means 11. Further, the equalization means 15 equalizes the output signal of the equalization target 10 with respect to the unknown input signal according to the output of the equalization control means 14. The equalization means 15 includes, for example, a plurality of equalization circuits and the like. The equalization control output 14 outputs a weighting signal for each output of the plurality of equalization circuits, so that the outputs of the plurality of equalization circuits are weighted. The sum is output as the equalized output signal. As mentioned above, FIG. 1 also shows the principle of the third invention. In the third invention, the characteristic detecting means 11
Is a neural network as in the first aspect of the invention, and is an equalization target 10 for which equalization for deterioration compensation of an output signal should be performed, for example, an equalization target 10 from an output signal for a known input signal to a transmission line. The characteristic of, for example, the transfer function is detected. However, this characteristic detection means 1
The internal state value of the neural network as 1, for example, the weight of the internal coupling is constant regardless of the signal pattern of the known input signal in the first invention, whereas it is known in the third invention. It is switched according to the signal pattern of the signal. As a result, the learning of the neural network as the characteristic detecting means 11 is efficiently executed. Further, the equalization means 12 equalizes the output signal of the equalization target 10 with respect to the unknown input signal by using the result detected by the characteristic detection means 11 as in the first aspect of the invention, and outputs the equalized output signal. Output.

FIG. 3 is a block diagram showing the principle of the fourth invention. In the figure, the characteristic detecting means 11 is, for example, a neural network as in the third invention and detects the characteristic of the equalization target 10. However, unlike the third invention, the internal state of the neural network is different. The weight as a value is not switched according to the signal pattern of the input signal, and the characteristic detection is performed while the weight of the internal coupling remains constant.

The post-processing means 16 is an equalization target 10 in time series.
The characteristic detection results of the characteristic detection means 11 are accumulated one after another using the output signal output from the device, and the post-processing of the accumulated characteristic detection results, for example, the average value of the transfer function is obtained. It is composed of a neural network. The equalization means 17 uses the output of the post-processing means 16 as a characteristic detection result of the equalization target 10 to equalize the target 10.
For equalizing the output signal of, which is also constituted by, for example, a neural network.

FIG. 4 is a block diagram showing the principle of the fifth invention. The operation of the characteristic detecting means 11 and the equalizing means 12 in the fifth invention is exactly the same as in the first invention, for example.

Equalization error detection and learning pattern storage means 1
Reference numeral 8 compares the known input signal to the equalization target 10 with the output signal of the equalization means 12, and when an equalization error is detected, the output signal, the detection result of the characteristic detection means 11, and the equalization target. 10 known input signals as correct output signals are accumulated as learning patterns.

The learning control means 19 detects the equalization error and learning pattern storage means 18 during a period when the equalization target 10 is not outputting an output signal, for example, during a period when the transmission line as the equalization target 10 is not used. Using the learning pattern accumulated in, the equalizing means 12, for example, a neural network is made to perform learning.

[0018]

In the present invention, for example, when data is transmitted in a packet format on the transmission line, for example, a training sequence or a known bit string called a recognition signal is inserted at the beginning of the packet, which is the equalization target 10,
For example, it is used as a known input signal for detecting the characteristic of the transmission line. Then, from the relationship between the input signal and the output signal of the equalization target 10, the characteristic detecting means 11, for example, a neural network detects the characteristic of the transmission line, for example, the real number part and the imaginary number part of the transfer function.

In the first invention whose principle is shown in FIG. 1, the equalizing means 12 uses the real and imaginary parts of the transfer function detected by the neural network as the characteristic detecting means 11, for example. The output signal of the transmission line is equalized with respect to an unknown input signal. The equalizing means 12 is also composed of, for example, a neural network.

In the second invention whose principle is shown in FIG. 2, control for equalizing the output signal of the equalization target 10 with respect to the unknown input signal by using the output of the neural network as the characteristic detecting means 11 is performed. Signal is output,
This control signal is a weighting signal for each output of a part of the equalizing means 15, for example, each of a plurality of equalizing circuits, and the weighting results of the outputs of the plurality of equalizing circuits are added to obtain an equalized output signal. Is output. This equalization circuit also
For example, each is a neural network.

In the third invention whose principle is shown in FIG. 1, in the neural network as the characteristic detecting means 11, the neural network as the characteristic detecting parameter corresponding to the input signal known in the neural network, that is, the bit pattern of the training sequence is used. The weights are switched, for example, the real and imaginary parts of the transfer function are detected,
The output signal of the transmission line is equalized with respect to an unknown input signal.

In the fourth invention whose principle is shown in FIG. 3, the bit strings of the training sequence are input to the characteristic detecting means 11 with a shift of, for example, each one bit. The detected result for each bit string is accumulated in, for example, a shift register, the accumulated results are post-processed, for example, averaged, and the processed result is used to output the output signal of the equalization target 10 to the unknown input signal. Is done. As a result, in comparison with the case of using the characteristic detection result only once,
It is possible to improve the accuracy of characteristic detection.

In the fifth invention whose principle is shown in FIG. 4, the characteristic of the equalization target 10 is detected by the characteristic detection means 11 as in the first invention, and the unknown result is input using the detection result. The output signal of the equalization target 10 is equalized with respect to the signal.

At this time, the equalization error detection / learning pattern storage means 18 compares the known input signal to the equalization target 10 with the output signal of the equalization means 12, and when an equalization error is detected, The output signal, the detection result of the characteristic detection means 11, and the correct output signal to be output from the equalization target 10 are accumulated as a learning pattern. Then, for example, during a period in which communication is not performed on the transmission line, the equalizing means 12
Learning is carried out by using this learning pattern.

In the above description, the training sequence at the beginning of the packet is used as the known input signal to the equalization target 10. However, if fading is extremely fast, it may occur during reception of one packet. It is also possible that the characteristics of the transmission line change. In order to prepare for such a case, the present invention can continue the characteristic detection even after the reception of the signal corresponding to the training sequence. In order to continue the characteristic detection by setting the equalized output signal of the adaptive equalizer for the unknown input signal as the correct one (known input signal), the characteristic detection is performed in FIG. 1 to FIG. 4 (first to fifth inventions). This is possible by providing a characteristic detection control device in the preceding stage of the means 11.

In the characteristic detection control device, for example, the known input signal and the output signal of the equalization target 10 are sent to the characteristic detection means 11 during the period when the training sequence as the known input signal is input to the equalization target 10. After the supply of the training sequence and the input of the training sequence, the unknown input signal, that is, the output of the equalization means 12 as a signal in place of the known input signal during the period when the transmission data is input to the equalization target 10 and the equalization target 10. The output signal is supplied to cause the characteristic detecting means 11 to continuously perform characteristic detection.

As described above, in the present invention, the transfer characteristic of the equalization target, for example, the transfer function is detected, and the detection result is used to adapt the transfer characteristic which changes every moment to equalize the output signal. It can be carried out.

[0028]

FIG. 5 is a block diagram showing the basic arrangement of an adaptive equalizer according to the first invention. In the figure, the adaptive equalizer 20
Is a characteristic detection device 22 for detecting a transfer characteristic of the equalization target 21, for example, a transfer function H (ω), and an equalization device for equalizing an output signal of the equalization target 21 using the detected transfer function characteristic. It is composed of 23.

In FIG. 5, the characteristic detecting device 22 and the equalizing device 23 are, for example, a neural network, the characteristic detecting device 22 performs learning by using the transfer function of the equalization target 21 as a teacher signal, and the equalizing device 23 Learning is performed by using an input signal to the equalization target 21 as a teacher signal.

FIG. 6 is a block diagram showing the configuration of an embodiment of the adaptive equalizer according to the first invention. 5, the equalization target 21 in FIG. 5 is a modulation system 26, a propagation path 27 and a demodulation system 28.
Composed by. Further, the adaptive equalizer 25 has a characteristic detecting neural network 2 corresponding to the input pattern creating section 24 and the characteristic detecting device 22 of FIG.
9, an equalization neural network 30 corresponding to the equalization device 23, and a threshold processing unit 31. The characteristic detection neural network 29 uses the known digital signal, that is, the input signal and the demodulated signal of the demodulation system 28 corresponding to the input signal, that is, the output signal, to determine the entire modulation system 26, the propagation path 27, and the demodulation system 28 (hereinafter , The whole is referred to as the transmission line) and the real part X of the transfer function of
(Ω) and the imaginary part Y (ω) are detected, and the equalization neural network 30 uses the detection result to detect the demodulation system 28.
Of the demodulated signal output by the above, the threshold processing unit 31 converts the equalized result into digital data of 0 or 1 by an appropriate threshold value, and outputs the digital signal, that is, an equalized output signal.

FIG. 7 is a model of the propagation path 27 in FIG. Generally, in line-of-sight microwave communication, the propagation path is described as a two-wave interference fading model of a direct wave and a reflected wave. At this time, the transfer function of the propagation path is expressed by the following equation using the amplitude ratio ρ of the reflected wave and the direct wave, the propagation delay time τ of the reflected wave and the direct wave, and the initial phase difference φ of the reflected wave and the direct wave. Be done. Ideally, the transfer functions of the modulation system 26 and the demodulation system 28 cancel each other out, so that the transfer function of the propagation path 27 can be regarded as the transfer function of the entire transmission path.

[0032]

[Equation 1]

Thus, from the original ρ, τ, φ, ρ,
It is possible to calculate τ ′ and φ ′ and calculate the real and imaginary parts of the transfer function from Eqs. (4) and (5). It can be seen that the characteristics detected by detecting the transfer function by the characteristic detection neural network 29 are the phase difference between the reflected wave and the direct wave in the transmission path and the amplitude ratio thereof.

FIG. 8 is a configuration block diagram of the modulation system 26 in the embodiment of FIG. In the figure, for example, I, which corresponds to the detection output when the carrier wave component extracted from the received wave during demodulation is used as the reference signal, and a signal obtained by advancing the phase of the carrier wave component by π / 2 are used as the reference signal. The input signals to the two channels of Q, which correspond to the detection output at the time, are quadrature-modulated and transmitted as transmission signals to the propagation path.

FIG. 9 is a block diagram of the demodulation system 28 shown in FIG. This demodulation system restores the received signal from the received signal by detecting means such as synchronous detection or differential detection. The BPF is a bandpass filter, the LPF is a lowpass filter, and the AD is an AD converter. It is 4 times oversampled and converted into a digital signal.

FIG. 10 shows an embodiment format of a packet format of data transmitted on the propagation path. At the beginning of each packet, called a training sequence, a known bit string common to each packet, for example 16 bits, is inserted,
Transmission path characteristics are detected using this bit string.
The information portion is the data that is actually transmitted, and since the present embodiment is intended for digital car telephones, the signals all take a binary value of 0 or 1. The SYNC word and guard are special codes that indicate the beginning and end of a packet.

FIG. 11 shows an embodiment of the characteristic detecting neural network 29 shown in FIG. In the figure, the characteristic detection neural network 29 has three layers in which the layers are completely connected.
It is a hierarchical neural network of layers. The number of input units is 32, of which 16 are actually received signals as output signals of the demodulation system 28 and the remaining 16 are ideal output values of the training sequence of FIG. A bit string is input. The output layer has two units, and outputs the real and imaginary parts of the transfer function of the transmission path. The number of intermediate layer units is 20, which is experimentally determined.

In FIG. 11, the bit sequence of the training sequence at the beginning of the packet and the bit sequence of the received signal are in units of 8 bits, and the input layer of the input layer is input to both I and Q channels via a delay device (shift register) D. The characteristic detection neural network, which is input to each unit, detects the transfer function of the transmission line by comparing a known training sequence with the actual received signal for the input pattern, and detects its real part X (ω) and imaginary number. The part Y (ω) is output.

FIG. 12 is a block diagram of an embodiment of an input pattern creating unit for the characteristic detection neural network. In the figure, the input pattern creation unit receives a training sequence detection unit 36 that detects a training sequence of a packet from a received signal, a training sequence memory 37 that stores a bit sequence of the training sequence, and a training sequence detection unit 36 that receives the output according to the output of the training sequence detection unit 36. The synchronization device 38 is configured to synchronize the signal with the ideal value stored in the training sequence memory 37 and output the synchronized value to the characteristic detection neural network 39.

FIG. 13 is a detailed block diagram of the training sequence detection unit 36 in the input pattern creation unit. In the figure, the training sequence detection unit detects the start and end timings of the training sequence, and the frame synchronizer 36a establishes packet synchronization from the received signal and outputs a reset signal to the counter 36b. The counter 36b uses this reset signal to output a training sequence timing signal indicating the start and end timings of the training sequence of the packet. As a means for establishing frame synchronization, it is possible to transmit a signal of a series having a strong autocorrelation before a packet and use the correlation on the receiving side for detection, but details thereof will be omitted here.

FIG. 14 is a detailed block diagram of the training sequence memory in the input pattern creating section. The training sequence memory 37 stores a counter 37a to which a training sequence timing signal sent from the training sequence detector 36 via the synchronizer 38 is input, and training sequence data. The contents of the training sequence memory 37 are ideal under the control of the counter 37a. It is composed of a memory 37b which is read as a value.

FIG. 15 shows an example of an input signal to the characteristic detection neural network. In the figure, the waveforms of the ideal signals of both the I and Q channels and the demodulated signal actually output from the demodulation system 28 are shown. In the demodulation system 28, one symbol of the original digital signal is oversampled by a factor of 4, and the characteristic detection neural network has two symbols of the original digital signal with respect to the ideal values of the I and Q channels and the received signal. 8 points of data are input, and the total input data is 32.

In order to detect the transfer function of the transmission line by the characteristic detection neural network shown in FIG. 11, it is necessary to learn the neural network in advance. This learning involves the real part X (ω) and the imaginary part Y of the transfer function of the transmission line.
The value of (ω) is used as a teacher signal, for example, by the back propagation method.

Since it is difficult to accurately obtain the actual transfer function of the transmission system, in the present embodiment, the received signal and the corresponding transfer function were obtained by simulation. As the learning data generation method, firstly, a method of creating software for simulating a modulation system, a propagation path, and a demodulation system on a computer and executing the simulation, and secondly,
A method of creating the hardware of the demodulation system and the hardware that simulates the operation of the propagation path model and creating the data by the experiment can be considered, but in the present embodiment, the first computer simulation method was used.

FIG. 16 shows changes in transmission line parameters in the simulation. This figure shows the changes in ρ, τ ′, and φ ′ in Eqs. (4) and (5). When ρ = 0, the effect of the reflected wave can be ignored, so ρ, τ ′, and There are 25 combinations of all φ '. On the other hand, 64 pieces of data were generated as a combination of 3-bit data for each of I and Q. Further, since the signal value one bit before changes the demodulation waveform due to the influence of the reflected wave, all patterns in which the bit is changed are generated.

FIG. 17 is an explanatory diagram of the necessity of learning all combinations of 3 bits each of I and Q. In this embodiment, it is assumed that the delay of the reflected wave with respect to the direct wave is up to 1 bit. FIG. 17 shows digital signals corresponding to the direct wave and the reflected wave when there is a delay of 1 bit. As described above, the received signal of 2 bits for each of the I and Q channels is input after being 4 times oversampled to the characteristic detection neural network. 1
Since it is affected by bits, it is necessary to generate and learn all combinations of 3 bits each for I and Q.

FIG. 18 shows an example of learning data. Since the demodulated signal is oversampled with 4 bits,
The same transfer function value is made to correspond to each cycle point, the same teacher signal is given to four input data, and the total learning data is
25 × 64 × 4 = 6,400. The learning data is shown in FIG.
As shown in FIG. 3, it is designed to be able to cope with the synchronization deviation.

FIG. 20 shows an embodiment of the equalization neural network 30 shown in FIG. In the figure, the equalization neural network is a three-layer hierarchical neural network in which layers are completely connected. The equalization neural network removes the nonlinear distortion in the propagation path and the demodulation system by equalizing the output signal of the demodulation system by the nonlinear processing of the neural network, and equalizes the output signal of the demodulation system to the target value. .. The unit of the input layer of the equalization neural network inputs the result of A / D conversion by oversampling the output signal of the demodulation system that has passed through the low-pass filter at 4 times the signal period. In order to adaptively respond to changes in the propagation path, the values of the real part and the imaginary part of the transfer function as the transmission line characteristics are also input, and the received signal is equalized using the characteristics.

In FIG. 20, the number of units in the input layer of the equalization neural network is 16 for inputting the received signal.
, And 2 for inputting the real and imaginary parts of the transfer function, 18 in total, 2 for the number of output units I and Q, 12 for the number of intermediate layer units determined by experiment Is. Here, the equalization neural network is 4
It outputs the center ideal value from the 2-bit over-sampled 2-bit data, as will be described later. The demodulated signal is input to the equalization neural network while shifting at each sampling time, Since the equalization result is continuously output at every sampling time, the output signal is one bit for each of the I and Q channels.

FIG. 21 is an explanatory diagram of a control method of data input to the equalization neural network. In the figure,
A latch circuit 400 not shown in FIG. 6 is provided between the characteristic detection neural network 29 and the equalization neural network 30, and a latch reset signal is sent from the training sequence detection unit 36 to the latch circuit 400. Is entered. Here, the training sequence detection unit 36 can of course be shared with the training sequence detection unit in the input pattern creation unit described in FIG.

FIG. 22 is a timing chart of an operation example of the equalization neural network in the circuit of FIG. Since this timing chart is the same for both I and Q channels, this timing chart will be described assuming that it is for either one of the channels.

In FIG. 22, the first 8 bits of the received signal indicate the output result of the transmission line for the training sequence of one channel, and the "information" indicates the output result of the transmission line for the subsequent unknown input information. There is.
First, the latch circuit 400 is reset by the reset signal from the training sequence detector 36,
Input of the received signal to the characteristic detection neural network 29 is started. Characteristic detection neural network 29
Outputs a characteristic detection result at the time point when the 4-fold oversampled 8-bit portion, that is, the 2-bit received signal portion corresponding to the training sequence is input, and the result is latched by the latch circuit 400, and at the same time, the equalization neural network. 30, the received signal corresponding to the training sequence is equalized, and the equalization of the output signal for the unknown input information is performed after the equalization for the eight bits is completed.

Although the operation of the equalization neural network has been described with reference to FIGS. 21 and 22 using two bits of the training sequence for one channel,
When the characteristic detection is performed by using more bits in the training sequence, the received signal is sent to the equalization neural network 30 in order to compensate the operation delay of the characteristic detection neural network 29 as shown in FIG. Will be input via the delay 401.

FIG. 24 shows an example of learning data of the equalization neural network. FIG. 16A shows transmission path parameters for the learning data of the characteristic detection neural network, that is, data generated for the parameters of FIG. The real part X and the imaginary part Y of the transfer function use theoretical values obtained from the calculation formulas, and do not use actual output values of the characteristic detection neural network. The total number of learning data is 25 × 64 × 4 = 6400 as in the case of the characteristic detection neural network. The teacher signal is
As shown in (b), an intermediate point of the ideal values of 8 points corresponding to the digital data of 2 symbols, that is, the values of the 4th to 5th points are used.
This is because 2 bits are input so that a delay of 1 bit can be dealt with, and the middle thereof is used as a teacher signal. Depending on whether the 4th point or the 5th point is taken, performance is improved. There is no big difference.

FIG. 25 is a flow chart of an embodiment of the output signal equalization processing in the first invention. When the processing is started in the figure, first, in step (S) 40, the training sequence of the packet is detected, and when the training sequence is detected, the bit string of the training sequence and the received signal corresponding thereto are detected in the characteristic detection neural network in S41. It is set in the input unit of the network, and in S42, the real part X and the imaginary part Y of the transfer function as the characteristics of the transmission line are detected.

Next, in S43, the characteristic detection result is set in the input unit of the equalization neural network, and in S44, it is determined whether or not the end of the packet, that is, whether or not all the data in one packet has been output. If not, the demodulated signal is set in each input unit of the equalization neural network in S45. S46
The equalization signal is output from the equalization neural network in step S47, the signal is thresholded in step S47 by the threshold value processing unit 31 in FIG. 6, and output as a digital signal.
The process from is repeated. When it is determined in S44 that the packet is the end, the process returns to S40, and the processing after detection of the training sequence for the next packet is repeated.

Next, experimental results of output signal equalization processing using the embodiment of FIG. 6 will be shown. FIG. 26 shows sampling points in the experiment. The output of the equalization neural network is sampled at the sampling points shown in this figure and the threshold
Converted to digital data using 0.5.

FIG. 27 shows the bit error rate as an experimental result.
In the figure, "learning" is the case where the learning data is equalized according to the present embodiment, "unknown" is the case where the unknown data is equalized,
“Unequalization” is a bit error rate when the received signal is sampled at the sampling points shown in FIG. 26 and threshold processing is performed without using the adaptive equalizer of the present invention.

In the above description, the characteristic detection neural network uses the first 4 bits (2 bits for each of the I and Q channels) of the training sequence at the beginning of the packet to measure the characteristics of the transmission path, for example, the transfer function. Was detected, and the equalization neural network used the result to equalize the output signal to the rest of the packet, that is, the information part. However, when fading is fast, even during reception of one packet. There has been a problem that the characteristics of the transmission line change and proper equalization cannot be performed. In order to solve such a problem, FIG. 28 shows that the equalization output signal output by the equalization neural network after the reception of the reception signal corresponding to the training sequence is finished is a correct input signal, that is, a known input signal. Is a block diagram of an embodiment of a feedback adaptive equalizer that continues characteristic detection even after reception of a received signal corresponding to a training sequence is considered.

When FIG. 28 is compared with the embodiment of FIG. 6, it is different in that a characteristic detection control device 48 for controlling characteristic detection by a characteristic detection neural network is provided in place of the input pattern creating section 24.

FIG. 29 is a configuration block diagram of an embodiment of this characteristic detection control device. In the figure, a characteristic detection control device 48 is a reception signal synchronization device 4 for synchronizing a reception signal with an ideal value and giving it to a characteristic detection neural network.
8a, a training sequence detector 48b for detecting a training sequence from the received signal, and an ideal value controller 48c for outputting an ideal value to the characteristic detection neural network.

The received signal synchronizer 48a outputs 1 of the received signal.
An ideal value control device 48c is composed of a reception signal FIFO 48d for delaying the bit and the oversampled signal by 4 bits, and a reception signal switch 48e for switching the output of the reception signal and the reception signal FIFO 48d. Is composed of a training sequence memory 48f storing a training sequence, and an ideal value switch 48g for switching and outputting the output of the training sequence memory 48f and the equalized output signal. The training sequence detector 4
8b and the training sequence memory 48f are
This is the same as that described in FIGS. 13 and 14.

FIG. 30 shows the received signal FIF described in FIG.
It is a configuration block diagram of O48d. In the figure, FI
The FO receives the received signal that has been oversampled 4 times, and is output to the received signal switch 48e via four delay devices having a delay time of 1/4 of 1 bit of the received signal. .. The characteristic detection in FIG. 28 will be described with reference to FIGS. 29 and 30.

As described above, the training sequence detector 48b detects the start and end timings of the training sequence and outputs them to the received signal synchronizer 48a and the ideal value controller 48c. First, when the head of the training sequence is detected, the reception signal switch 48e outputs the input reception signal according to the timing signal, and the ideal value switch 48g outputs the training sequence from the training sequence memory 48f. Can be switched. As a result, the characteristic detection neural network detects the characteristic of the transmission line using the training sequence as a known input signal.

When the training sequence ends, the timing signal is output again, the reception signal switch 48e outputs the output from the reception signal FIFO 48d, and the ideal value switch 48g outputs the equalization output signal output from the equalization neural network. It is switched to output to the characteristic detection neural network 29. Here, the delay time of the equalization operation by the equalization neural network is set to one bit of the signal, and the received signal FIFO is set to the one bit.
The received signal is delayed by 48d, and the delayed received signal is input to the equalization detection neural network 29 at the same time as the output of the equalization neural network, so that the equalized output signal is used instead of the known input signal. Will continue to be detected. Here, it goes without saying that the delay time by the reception signal FIFO of FIG. 30 is adjusted according to the delay time by the equalization neural network.

FIG. 31 is a timing chart of the embodiment of the signal equalization processing in the embodiment of FIG. FIG. 3A shows an example of the received signal. The shaded portion at the beginning, that is, the signals corresponding to the times 0 to 7 correspond to the training sequence, and the signals for the subsequent times correspond to the unknown input signal, that is, the data. The received signal. The received signal is an analog signal, but is shown here as a corresponding digital signal for convenience.

As described above, the received signal is oversampled four times and input to the adaptive equalizer.
In the time chart shown in (b), the time shown in (a) is divided into four parts. For example, time 0 is 0.0-
It is represented by 4 of 0.3. In the characteristic detection neural network, as described in FIG. 11, 2 bits of the received signal for one channel, that is, 8 bits of the oversampled signal are input to the 8 units of the input layer. But those eight inputs are t0
.About.t7, respectively. The shaded signals here represent the signals corresponding to the training sequence, as in FIG.

The training sequence is 8 bits, and the signal corresponding to the training sequence is switched from the signal corresponding to the training sequence to the signal corresponding to the data from time 7.3 to time 8.0. At this time, the training sequence (TS) timing signal is output, and the reception sequence shown in FIG. The signal switch and the ideal value switch are switched, and the received signal delayed by one bit and the equalization output signal output from the equalization neural network are input to the characteristic detection neural network 29. Assuming that the equalization neural network outputs a digital signal corresponding to t4 at the input to the input layer as described with reference to FIG. 24, the equalization output signal can be an ideal value corresponding to the received signal. it can.

As described above, even when the characteristic detection is continued after the reception of the received signal corresponding to the training sequence is completed, the characteristic detection is correctly performed in the normal case where the bit error rate of the equalized output signal is low, No error is accumulated. It is expected that the characteristic detection will not be successful if a bit error occurs, but the advantage of being able to follow the characteristic change of the transmission path is greater.

FIG. 32 is a block diagram showing the basic arrangement of an adaptive equalizer according to the second invention whose principle is shown in FIG.
Comparing this figure with FIG. 5 showing the basic configuration block diagram of the adaptive equalizer according to the first invention, a control device 53 is provided between the characteristic detection device 52 and the equalization device 54, and the control device 5 is provided.
The configuration of FIG. 5 is similar to that of FIG. 5, except that 3 outputs a control signal for controlling the equalizing action of the output signal by the equalizer 54 using the transfer function characteristic detected by the characteristic detector 52. It is the same.

FIG. 33 is a block diagram showing the configuration of an embodiment of the adaptive equalizer according to the second invention. Comparing this figure with FIG. 6 for the first invention, the equalization neural network 6
5 are provided, the outputs of the equalization neural networks are weighted by the multiplier 67 using the weighting signal output by the control neural network 66, added by the adder 68, and input to the threshold processing unit 69. The configuration is the same as that in FIG. 6 except that

Since the propagation path model, the modulation system, the demodulation system, the packet format, the characteristic detection neural network, and the learning data thereof in the second invention are the same as those in the first invention, the description thereof will be omitted.

FIG. 34 shows an embodiment of the control neural network 66 shown in FIG. This drawing is an embodiment corresponding to the case where there are four equalization neural networks 65 in FIG.

In FIG. 34, the unit of the input layer of the control neural network is the real part X and the imaginary part Y of the transfer function as the output of the characteristic detection neural network 64.
Are input, the number of units in the output layer is 4, the number of units in the output layer is 4, and the number of units in the intermediate layer is 3.

FIG. 35 shows an example of learning data for a control neural network. Here, the four equalization neural networks in FIG. 33 have the phase difference 0 between the direct wave and the reflected wave,
π / 2, π, and 3π / 2, respectively. As a result, when the phase difference is 0, the output of the unit 1 of the output layer of the control neural network is 1 and π / 2. For example, the learning is performed so that the output of the unit 2 is 1, the output of the unit 3 is 1 for π, and the output of the unit 4 is 1 for 3π / 2.

That is, as explained using the equations (4) and (5), the characteristics actually detected by the characteristic detection neural network are the phase difference and the amplitude ratio. The output of a plurality of equalization neural networks that perform signal equalization is weighted by the control neural network.

FIG. 36 shows an example of learning data when the amplitude ratio ρ and the phase difference φ are changed. By using a neural network as the control device, an appropriate weighting signal is output by the interpolation function even when X and Y corresponding to other than the actually learned amplitude ratio and phase difference are input. Is expected to For example, when the amplitude ratio is 0.8 and the phase difference is π / 4, it is expected that the output of unit 1 and unit 2 will be 0.5 and the output of unit 3 and unit 4 will be 0.

FIG. 37 shows an embodiment of the equalization neural network 65 in the second invention. In the second invention, the output of the control neural network 66 weights the outputs of the plurality of equalization neural networks 65. The weighting is performed by the real part X of the transfer function as the output of the characteristic detection neural network 64. Since it is performed according to the value of the imaginary part Y, it is not necessary to input the real part X and the imaginary part Y of the transfer function to the equalization neural network, unlike FIG. 20 for the first invention.

FIG. 38 is a block diagram showing the configuration of an embodiment of a feedback equalizer which continues the characteristic detection even after the reception of the signal corresponding to the training sequence in the second invention. Compared with FIG. 33, FIG. 28 is different only in that a characteristic detection control device 70 is provided instead of the input pattern creation unit, and the operation of this characteristic detection control device is shown in FIG.
~ It is exactly the same as that described in Fig. 31.

FIG. 39 is a basic block diagram of the adaptive equalizer according to the third invention. In the figure, the adaptive equalizer 80 indicates a transfer characteristic of the equalization target 79, for example, a transfer function H.
A characteristic detection device 81 for detecting (ω), an equalization device 82 for equalizing the output signal of the equalization target 79 using the detected transfer function characteristic, and a characteristic detection device parameter for managing the characteristic detection parameter. The management unit 83 and a characteristic detection device parameter storage unit 84 that stores the parameters thereof are included.

In FIG. 39, the characteristic detecting device 81 and the equalizing device 82 are, for example, a neural network, the characteristic detecting device 81 performs learning by using the transfer function of the equalization target 79 as a teacher signal during learning, and the equalizing device 82 Learning is performed using the input signal to the equalization target 79 as a teacher signal.

FIG. 40 is a block diagram showing the configuration of an embodiment of the adaptive equalizer according to the third invention. In FIG. 39,
The equalization target 79 is composed of a modulation system 86, a propagation path 87, and a demodulation system 88. Further, the adaptive equalizer 85 has a first
12 for the invention of FIG. 12 described in FIG.
4, a characteristic detection neural network 89 corresponding to the characteristic detection device 81 in FIG. 39, an equalization neural network 90 corresponding to the equalization device 82, a threshold processing unit 91, and a characteristic detection NN corresponding to the characteristic detection device parameter management unit 83.
A (neural network) parameter management unit 92 and a characteristic detection NN parameter storage unit 93 corresponding to the characteristic detection device parameter storage unit 84.

In FIG. 40, the operation of equalizing the received signal is the same as that in FIG. 6 except that the characteristic detection NN 89 detects the characteristic by using the bit pattern of the training sequence as described later. Details of the characteristic detection NN parameter management unit 92 and the characteristic detection NN parameter storage unit 93 will be described later.

The propagation path model, the modulation system, the demodulation system, and the packet format of the third invention are the same as those of the first invention, and the description thereof will be omitted. FIG. 41 shows an embodiment of the characteristic detection neural network 89 shown in FIG. In the figure, the characteristic detection neural network 89 is a three-layer hierarchical neural network in which the layers are completely connected. The number of input units is 16,
Actual received signals as output signals of the demodulation system 88 are input to these units. The output layer has two units, and outputs the real and imaginary parts of the transfer function of the transmission path. The number of intermediate layer units will be experimentally determined, for example.

In FIG. 41, the bit string of the received signal is input to each unit of the input layer via a delay device (shift register) D for both I and Q channels in units of 8 bits, and the input pattern On the other hand, the characteristic detection neural network detects the transfer function of the transmission line by using the internal state value corresponding to the bit pattern of the known training sequence, that is, the connection weight, and the real part X
(Ω) and the imaginary part Y (ω) are output.

The input signal to the characteristic detection neural network is only the demodulated signal in FIG. 15 in the embodiment of the first invention. The signal has 8 points for each of the I and Q channels, and the input data is 16 in total.

FIG. 42 is a configuration block diagram of an embodiment of the characteristic detection neural network parameter management unit 92. In the figure, a characteristic detection neural network parameter management unit 92 includes a training sequence memory 96 that stores a training sequence, a frame synchronization signal, that is, a counter 9 that receives a signal indicating the start of the training sequence and outputs a count value thereafter.
7. According to the output of the count value from the counter, the weight corresponding to the bit string stored in the training sequence memory 96 is extracted from the characteristic detection neural network parameter storage unit 93 and output to the characteristic detection neural network from the weight control unit 98. It is configured.

In the present embodiment, the digital signal and the training sequence are sequentially assigned to each channel like IQIQIQ ... And are quadrature-modulated.
Therefore, in FIG. 42, the underlined bit string 0011 in the training sequence memory 96 has I of 01,
It corresponds to the signal that Q is 01. In this embodiment, I = 00
, Q = 00, I = 00 and Q = 01, ..., 16 detection neural network parameters corresponding to 16 detection patterns of I = 11 and Q = 11 are generated. It is stored in the network parameter storage unit 93.

In FIG. 42, the counter 97 is reset by the frame synchronization signal supplied from the outside, and thereafter the counter value is increased according to the signal frequency, and the value is notified to the weight controller 98. The weight control unit 98 reads the corresponding bit pattern from the training sequence memory 96 according to the counter value, and outputs the weight of the internal connection of the characteristic detection neural network to the characteristic detection neural network as the characteristic detection neural network parameter by the bit pattern. The characteristic detection neural network can detect the transmission path characteristic by using the weight value according to the current input bit pattern (2 bits for each of I and Q).

FIG. 43 is a block diagram showing the detailed arrangement of the weight control unit shown in FIG. In the figure, the weight control unit 98 has a training sequence reading unit 98a that reads the contents of the training sequence memory according to the output of the counter.
According to the output of the training sequence reading unit 98a,
The weight of the characteristic detection neural network is updated using the characteristics detection neural network parameter corresponding to 4 bits of the training sequence from the characteristic detection neural network parameter storage section, that is, the weight reading section 98b for reading the weight, and the output of the weight reading section 98b. The weight update unit 98c is included.

The simulation of the transfer function of the transmission line, the transmission line parameters and the like are the same as those for the first aspect of the invention, and the description thereof will be omitted. FIG. 44 is an example of learning data of the characteristic detection neural network. Compared with FIG. 18 showing the learning data for the first invention, the only difference is that the ideal signal indicating the training sequence is not included as an input signal. The learning data in FIG. 44 is also shown in FIG.
It is taken to cope with the synchronization deviation as in 9.

The equalization neural network, the learning data thereof, and the output signal equalization processing flowchart for the third invention are also the same as those for the first invention, and the description thereof will be omitted. However, in the equalization processing flowchart,
25 differs from the first invention only in that the NN parameter corresponding to the training sequence is set in the characteristic detection NN in place of the training sequence itself in S41 of FIG.

FIG. 45 is a block diagram showing the configuration of an embodiment of a feedback equalizer which continues the characteristic detection even after the reception of the signal corresponding to the training sequence in the third invention. Also in this figure, similar to FIGS. 28 and 38, the characteristic detection control device 99 is provided in place of the input pattern creation unit, which is different from FIG. 40, and the operation of the characteristic detection control device is the same as that described above. It is exactly the same.

FIG. 46 is a block diagram showing the basic configuration of the adaptive equalizer according to the fourth invention whose principle is shown in FIG. Comparing this figure with FIG. 39 showing the basic configuration block diagram of the adaptive equalizer according to the third invention, the characteristic accumulating device 103 and the post-processing unit 1 are provided between the characteristic detecting device 102 and the equalizing device 105.
04, and the characteristic detection device parameter management unit 8
3 is different from that of FIG. 3 in that the storage section 84 does not exist.

The characteristic detecting device 102 shifts the above-mentioned training sequence by 1 bit, for example,
The transfer function characteristic of the equalization target 100 is detected using the output signal output from the equalization target 100 in time series, and the result is output to the characteristic storage device 103. Characteristic storage device 103
Is, for example, a shift register, which accumulates the result of shifting until the characteristic detection result is output from the characteristic detection device 102 a predetermined number of times. Characteristic storage device 103
After the characteristic detection results are accumulated for a predetermined number of times, the post-processing unit 104 performs post-processing, for example, statistical processing on the accumulated detection results, and averages the error of the detection results to obtain stable and highly accurate characteristics. Detection will be performed.

FIG. 47 is a block diagram showing the configuration of an adaptive equalizer according to the fourth embodiment of the present invention. Comparing this figure with FIG. 40 for the third invention, between the characteristic detection neural network 111 and the equalization neural network 114, the characteristic storage device 112 and the post-processing neural network 1 are provided.
The configuration is the same as that in FIG. 40 except that 13 is provided and the characteristic detection NN parameter management unit and the same storage unit do not exist. The equalization neural network 114 then outputs the real part X (ω) and the imaginary part Y (ω) of the transfer function output from the post-processing neural network 113.
Is used to equalize the demodulated signal output from the demodulation system 109.

In the fourth invention as well, as in the third invention, the characteristic detection neural network 111 can perform the characteristic detection using the detection parameter corresponding to the bit pattern of the signal, that is, the weight value. Figure 4
8 is a configuration block diagram of an embodiment of an adaptive equalizer using a characteristic detection parameter corresponding to a bit pattern in the fourth invention. As in FIG. 40 for the third invention, a characteristic detection neural network parameter management unit 116 for outputting a weight value to the characteristic detection neural network 111 and a characteristic detection neural network parameter storage unit 117 are provided.

In FIG. 48, for example, a training sequence is extracted from the beginning of the training sequence of FIG. 42 by 4 bits and a characteristic detection parameter corresponding thereto is used. First, a characteristic detection parameter corresponding to a bit string 0100 (I is 00, Q is 10), 1001, 0011, ...

The differences between the fourth embodiment of the invention and the third embodiment will be described below. FIG. 49 shows an example of the post-processing neural network. The post-processing neural network has a real part X (ω) and an imaginary part Y of the transfer function.
Input layer consisting of 16 input units, the middle layer consisting of 10 units, and the transfer function to which the values at the current time t with respect to (ω) and the previous seven times (up to t-7) are respectively input. 3 is a three-layer structure of an output layer consisting of two units for outputting the real parts X and Y of The post-processing neural network uses the output of the characteristic detection network 111 at the current time t and the past time to output the average value of the real part and the imaginary part of the transfer function to determine the transmission line characteristic at the time t. decide.

In FIG. 49, the values of the real number part and the imaginary number part at time t and the seven times before it are used, but the number of times is determined according to the storage capacity and the calculation time. It is natural that the number is not limited to eight.

FIG. 50 shows an example of learning data of the post-processing neural network. The post-processing neural network receives the time-series pattern of the transmission line characteristic value detected by the characteristic detection neural network as an input, and executes learning by using the theoretical value of the characteristic value as an output value, that is, a teacher signal. As shown in the input data example in the figure, the characteristic detection neural network does not always output the correct characteristic value, and the learning improves the characteristic detection accuracy.

The learning data of the post-processing neural network is obtained by first learning the characteristic detection neural network, inputting the received signal corresponding to the training sequence to the characteristic detection neural network, outputting the characteristic value, and then outputting the characteristic value. The time-series pattern is generated by inputting learning data and using an ideal characteristic value corresponding thereto as a teacher signal.

FIG. 51 is a flow chart of an embodiment of signal equalization processing in the fourth invention. In the figure, first S
At 121, it is determined whether or not it is the training sequence as at S40 in FIG. 25. At the training sequence, the received signal is set to the input unit of the detection neural network at S122, and the characteristic detection neural network is set to the real part of the transfer function at S123. The imaginary part is detected, and the result is stored in the characteristic storage device. And S
At 124, it is determined whether or not a predetermined number of results have been accumulated, and if not accumulated, the processing from S122 is repeated.

If it is determined in S124 that the predetermined number has been accumulated, the post-processing neural network performs post-processing of the accumulated characteristics in S125 to obtain characteristic values, and the characteristics are equalized in the neural network in S126. Set to the input unit of. Hereinafter S127 to S130
Then, equalization of the received signal is performed as in S44 to S47 in FIG.

52 and 53 are configuration block diagrams of an embodiment of a feedback adaptive equalizer which continues characteristic detection even after reception of a signal corresponding to a training sequence in the fourth invention. The input pattern creating section in FIGS. 47 and 48 is replaced with a characteristic detection control device, and the operation of the characteristic detection control device is exactly the same as described above.

FIG. 54 is a block diagram showing the configuration of an embodiment of the statistical processing device which constitutes the characteristic accumulating device and the post-processing unit in the fourth invention (FIG. 46). In the figure, first, the latch circuit 140 is reset by the frame synchronization signal, and then the output of the characteristic detection neural network, that is, the real part and the imaginary part of the transfer function are stored in the characteristic storage device 141. The characteristic storage device 141 is composed of shift registers for storing eight detection results for each of the real number part and the imaginary number part of the transfer function. The counter 142 counts the time until the characteristic detection result is obtained in these shift registers. After counting the time, the statistical processing device 143 is activated by the counter 142, and the processing result of the statistical processing is held in the latch circuit 140 and output to the equalization neural network.

Here, the statistical processing device has the characteristic values X (ω, t-i) and Y (ω, t-i) output from the characteristic storage device.
Is used to perform the statistical processing represented by the following formula. For example λ
When (t) = 1 / n, a simple average is obtained.

[0108]

[Equation 2]

FIG. 55 is a basic configuration block diagram of an adaptive equalizer according to the fifth invention. In the figure, the adaptive equalizer 145 indicates a transfer characteristic of the equalization target 146, for example, a transfer function H.
A characteristic detection device 147 for detecting (ω), an equalization device 148 that equalizes the output signal of the equalization target 146 using the detected transfer function characteristic, a known input signal to the equalization target 146, and the like. Equalization error determination device 149 for stabilizing the equalization error by comparing with the output signal of the equalization device 148, recognition for storing a known bit sequence of the input signal used by the equalization error determination device 149 for determination, for example, a training sequence When the signal storage device 150 and the equalization error determination device 149 determine an equalization error, the output of the equalization target 146 and the characteristic detection device 14
The learning pattern storage device 151 for storing the detection result of No. 7 and the known input signal to the equalization target 146 as the correct output of the equalization device 148 as a learning pattern, and the transmission line as the equalization target are used. When not in use, the learning control device 152 is configured to allow the equalization device 148 to learn the learning pattern stored in the learning pattern storage device 151.

FIG. 56 is a block diagram showing the configuration of an adaptive equalizer according to the fifth embodiment of the present invention. In the figure, the configuration from the input pattern creating section to the threshold value processing section is the same as in FIG. 6 for the first invention. Hereinafter, differences from the first invention will be described as an embodiment of the fifth invention.

In the adaptive equalizer of the fifth aspect of the invention shown in FIG. 56, the processing result by the threshold processing unit 163 of the output of the equalization neural network 162 and the training sequence, that is, the recognition signal storage device 150 stores the result. The recognized signal is compared and the equalization error is determined. In this determination, whether or not the values for both I and Q channels of the training sequence at an arbitrary time and the values for both I and Q channels of the equalized bit pattern are equal in both I channel and Q channel. It is determined that the equalization error occurs when either or both of them are not equal.

FIG. 57 shows an example of equalization error judgment. In the figure, at arbitrary times, the values of both the I and Q channels of the training sequence and the values of the equalized digital signal for the I and Q channels are compared, and the values of the I and Q channels are respectively compared. If they are equal, "0" is output as the determination result indicating that they are correct,
When either one or both are not equal, "1" indicating an error is output as the determination result.

When an equalization error is detected at a certain time, the equalization error determination device 149 inputs the input value to the input unit of the equalization neural network 162, that is, the real part X of the transfer function output by the characteristic detection neural network 161. (Ω), imaginary part Y (ω), and 8 input values for both I and Q channels are collected, and a learning pattern storage device is used as a set of bit pattern I and Q values of the training sequence at the same time. It stores in 151 as a learning pattern. Here, the bit patterns I and Q of the training sequence become the teacher signal given to the output unit of the equalization neural network.

FIG. 58 shows an example of the learning pattern generated by the equalization error judgment. This learning pattern is added as learning data of the equalization neural network 162 in addition to the learning data used for learning the equalization neural network 162 in advance, that is, the learning data of FIG. Learning pattern storage device 15
The learning data of FIG. 24 has already been stored in No. 1 and, by using both of the added data of FIG. Is used to learn the equalization neural network 162.

FIG. 59 is a detailed block diagram of the learning control device in FIG. In the figure, the learning pattern control device 165 reads a learning pattern from the learning pattern storage device and divides the pattern into an input signal and a teacher signal. The input signal presentation device 166 outputs the input signal given from the learning pattern control device 165 to the input unit of the equalization neural network, and the equalization neural network performs forward calculation based on this input signal, and outputs the output signal as an output error. Output to the calculation device 167. The output error calculation device 167 calculates an error between the output of the equalization neural network and the teacher signal given from the learning pattern control device 165, and the error is updated by the weight update device 168.
The weight update device 168 updates the internal engagement weight of the equalization neural network by using, for example, the backpropagation method. These operations are executed for all learning patterns until the error between the teacher signal and the output signal converges below a certain value.

FIG. 60 is a configuration block diagram of an embodiment of a feedback type adaptive equalizer which continues characteristic detection even after receiving a signal corresponding to a training sequence in the fifth invention.
Comparing this figure with the embodiment of FIG. 56, only the point that a characteristic detection control device is provided instead of the input pattern creation unit as in the above, and the operation of the characteristic detection control device is also exactly the same as that described above. Is. Note that the learning patterns stored in the learning pattern storage device in FIG.
Only those that correspond to the training sequence as in 6. This is because the correctness of the equalization result cannot be determined for the received signal after the end of the training sequence.

FIG. 61 is a configuration block diagram of an embodiment in which the adjusting device of the adaptive equalizer is independent. The configuration of FIG.
Which is exactly the same as the embodiment of FIG.
9, recognition signal storage device 150, learning pattern storage device 1
51 and a learning control device 152, an adjusting device for an adaptive equalizer is provided independently of the equalizer, and this adjusting device is used by being connected to the adaptive equalizer.

In the embodiments of the first to fifth inventions described in detail above, the case where the neural network is used as the characteristic detection device, the equalization device, the equalization control device, and the post-processing device has been described. It is also possible to use a device such as a transversal type equalizer. Further, although the case where the hierarchical network is used as the neural network has been described, it is also possible to use a network model having a recursive connection from a unit in the middle layer of the hierarchical network to a unit in the input layer, that is, a recurrent network.

Further, the values of the real number part and the imaginary number part of the transfer function are used as the characteristics of the transmission line, but other characteristic values such as the amplitude frequency characteristic and the envelope characteristic of the transfer function may be used. The amplitude frequency characteristic A (ω) and the envelope characteristic D (ω) in the case of the two-wave model are given as follows.

[0120]

[Equation 3]

In the present embodiment, assuming that the phase difference between the reflected wave and the direct wave is within one symbol of the digital signal, in order to see at least one symbol before, the data for two symbols is four times over. Although the data is sampled and input as 8-bit data to the equalization neural network or the like, it is naturally possible to cope with a larger delay by further increasing the number of input bits.

[0122]

As described in detail above, according to the present invention, for example, a transfer function is detected as the transmission characteristic of the transmission system,
By using the result, the signal can be equalized by flexibly following the temporal change of the transmission characteristic. Further, an equalizer having an adaptive equalization function is output in accordance with the detected transfer function by outputting a control signal as a weighting signal for the outputs of a plurality of equalization circuits, which corresponds to the bit pattern of the training sequence. It is possible to configure an equalizer having a function of detecting a transmission characteristic by using a neural network that can perform learning efficiently by using a detection parameter. Furthermore, when an equalization error is detected, the equalization device learns the correspondence relationship between the received signal including the error, the detected transmission characteristics, and the correct signal, and the equalizer learns the learning pattern. In addition, it is possible to improve the performance of the equalizer, which largely contributes to the improvement of transmission characteristics in the communication system.

[Brief description of drawings]

FIG. 1 is a principle block diagram of first and third inventions.

FIG. 2 is a principle block diagram of a second invention.

FIG. 3 is a principle block diagram of a fourth invention.

FIG. 4 is a principle block diagram of a fifth invention.

FIG. 5 is a block diagram showing a basic configuration of an adaptive equalizer according to the first invention.

FIG. 6 is a block diagram showing a configuration of an embodiment of an adaptive equalizer according to the first invention.

FIG. 7 is a diagram showing a propagation path model in the present invention.

8 is a block diagram showing a configuration of a modulation system in the embodiment of the adaptive equalizer of FIG.

9 is a block diagram showing the configuration of a demodulation system in the embodiment of the adaptive equalizer of FIG.

FIG. 10 is a diagram showing an example format of a packet in the present invention.

11 is a diagram showing a characteristic detection neural network in the embodiment of the adaptive equalizer of FIG.

FIG. 12 is a block diagram showing a configuration of an embodiment of an input pattern creation unit for a characteristic detection neural network.

FIG. 13 is a block diagram showing a detailed configuration of a training sequence detection unit.

FIG. 14 is a block diagram showing a detailed configuration of a training sequence memory.

15 is a diagram showing an example of an input signal to the characteristic detection neural network of FIG.

FIG. 16 is a diagram showing transmission path parameters in a simulation of learning data generation.

FIG. 17 is an explanatory diagram showing the necessity of learning all combinations of 3 bits each of I and Q.

FIG. 18 is a diagram showing an example of learning data of a characteristic detection neural network.

FIG. 19 is a diagram for explaining how to deal with a waveform shift in learning data of a characteristic detection neural network.

20 is a diagram showing an equalization neural network in the embodiment of the adaptive equalizer of FIG.

FIG. 21 is an explanatory diagram of a data input control method for the equalization NN.

FIG. 22 is a timing chart of an operation example of equalization NN.

FIG. 23 is an explanatory diagram of an operation delay compensation method for characteristic detection NN.

FIG. 24 is a diagram showing an example of learning data of an equalization neural network.

FIG. 25 is a flowchart of an embodiment of signal equalization processing in the first invention.

FIG. 26 is a diagram showing sampling points in the signal equalization experiment in the first invention.

FIG. 27 is a diagram showing a result of a signal equalization experiment in the first invention.

FIG. 28 is a block diagram showing a configuration of an embodiment of a feedback adaptive equalizer according to the first invention.

FIG. 29 is a block diagram showing a configuration of a characteristic detection control device.

FIG. 30 is a block diagram showing the structure of a reception signal FIFO.

31 is a timing chart of the signal equalization processing embodiment in the embodiment of FIG. 28. FIG.

FIG. 32 is a block diagram showing a basic configuration of an adaptive equalizer according to the second invention.

FIG. 33 is a block diagram showing the configuration of an embodiment of an adaptive equalizer according to the second invention.

FIG. 34 is a diagram showing a control neural network in the embodiment of the adaptive equalizer of FIG. 33.

FIG. 35 is a diagram showing an example of learning data for a control neural network.

FIG. 36 is a diagram showing an example of learning data of the control neural network when the amplitude ratio and the phase difference are changed.

FIG. 37 is a diagram showing an equalization neural network in the embodiment of the adaptive equalizer of FIG. 33.

FIG. 38 is a block diagram showing the configuration of an embodiment of a feedback type adaptive equalizer according to the second invention.

FIG. 39 is a block diagram showing a basic configuration of an adaptive equalizer according to the third invention.

FIG. 40 is a block diagram showing the configuration of an embodiment of an adaptive equalizer according to the third invention.

41 is a diagram showing a characteristic detection neural network in the embodiment of the adaptive equalizer of FIG. 40. FIG.

FIG. 42 is a block diagram showing the configuration of an embodiment of a characteristic detection neural network parameter management unit.

FIG. 43 is a block diagram showing a detailed configuration of a weight controller.

FIG. 44 is a diagram showing an example of learning data of a characteristic detection neural network.

FIG. 45 is a block diagram showing the configuration of an embodiment of a feedback adaptive equalizer according to the third invention.

FIG. 46 is a block diagram showing a basic configuration of an adaptive equalizer according to a fourth invention.

FIG. 47 is a block diagram showing a configuration of an adaptive equalizer according to an exemplary embodiment of the fourth invention.

FIG. 48 is a configuration block diagram of an embodiment of an adaptive equalizer using a characteristic detection parameter corresponding to a bit pattern in the fourth invention.

FIG. 49 is a diagram showing an example of a post-processing neural network.

FIG. 50 is a diagram showing an example of learning data of a post-processing neural network.

FIG. 51 is a flowchart of an embodiment of signal equalization processing in the fourth invention.

FIG. 52 is a block diagram (part 1) showing the configuration of an embodiment of a feedback type adaptive equalizer according to the fourth invention.

FIG. 53 is a block diagram (No. 2) showing the configuration of an embodiment of a feedback adaptive equalizer according to the fourth invention.

FIG. 54 is a block diagram showing a configuration of an embodiment of a characteristic accumulating device and a statistical processing device in the fourth invention.

FIG. 55 is a block diagram showing a basic configuration of an adaptive equalizer according to the fifth invention.

FIG. 56 is a block diagram showing a configuration of an adaptive equalizer according to an exemplary embodiment of the fifth invention.

FIG. 57 is a diagram illustrating an example of equalization error determination.

FIG. 58 is a diagram showing an example of a learning pattern generated by equalization error determination.

FIG. 59 is a block diagram showing a detailed configuration of a learning control device in the example of FIG. 56.

FIG. 60 is a block diagram showing the configuration of an embodiment of a feedback adaptive equalizer according to the fifth invention.

FIG. 61 is a block diagram showing a configuration of an embodiment in which an adjusting device of an adaptive equalizer is independent.

[Fig. 62] Fig. 62 is a diagram for describing two-wave interference facing in a mobile communication system targeted by the present invention.

FIG. 63 is a diagram showing an example of a transversal type equalizer.

[Explanation of symbols]

10, 21 Equalization target 11 Characteristic detection means 12, 15, 17 Equalization means 14 Equalization control means 16 Post-processing means 18 Equalization error detection and learning pattern storage means 19 Learning control means 20, 25 Adaptive equalizer 22 Characteristics Detection device 23 Equalization device 24 Input pattern creation unit 26 Modulation system 27 Propagation path 28 Demodulation system 29 Characteristic detection neural network 30 Equalization neural network 31 Threshold processing unit 36 Training sequence detection unit 37 Training sequence memory 48 Characteristic detection control device

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Shigemi Nagata 1015 Kamiodanaka, Nakahara-ku, Kawasaki-shi, Kanagawa Fujitsu Limited

Claims (36)

[Claims] 1. A characteristic detecting means for detecting a characteristic of an equalization target (10) from an output signal with respect to a known input signal to the equalization target (10) to be equalized as compensation for deterioration of the output signal. 11) and an equalization means (12) for equalizing an output signal of the equalization target (10) with respect to an unknown input signal by using a result detected by the characteristic detection means (11). A characteristic adaptive equalizer. 2. Before the characteristic detection means (11),
The known input signal and the output signal of the equalization target (10) are supplied to the characteristic detection means (11) during a period in which the known input signal is input to the equalization target (10), During the period in which an unknown input signal is input to the equalization target (10) after the input of the known input signal, the output of the equalization means (12) as a signal replacing the known input signal and the equalization target ( The adaptive equalizer according to claim 1, further comprising a characteristic detection control device which supplies the output signal of 10) and causes the characteristic detection means (11) to perform continuous characteristic detection. 3. The adaptive equalization according to claim 1 or 2, wherein the characteristic detecting means (11) is a neural network or a learning device for acquiring the characteristic detecting function by learning a posted case. vessel. 4. The equalizing means comprises a neural network or a learning device for acquiring the equalizing function by learning a presented case.
Alternatively, the adaptive equalizer described in 2. 5. The characteristic detecting means (11) detects a transfer function of the equalization target (10) or a value of a real part and an imaginary part of the transfer function.
The adaptive equalizer described in 2, 3, or 4. 6. The characteristic detecting means (11) detects values of an amplitude frequency characteristic and an envelope characteristic of a transfer function of the equalization target (10). Adaptive equalizer described. 7. A characteristic detecting means for detecting a characteristic of an equalization target (10) from an output signal corresponding to a known input signal to the equalization target (10) to be equalized as compensation for deterioration of the output signal (10). 11) and an equalization control means for outputting a control signal for equalizing the output signal of the equalization target (10) with respect to the unknown input signal using the result detected by the characteristic detection means (11) ( 14) and equalization means (15) for equalizing an output signal of the equalization target (10) with respect to an unknown input signal according to the output of the equalization control means (14). Adaptive equalizer. 8. A stage before the characteristic detecting means (11),
The known input signal and the output signal of the equalization target (10) are supplied to the characteristic detection means (11) during a period in which the known input signal is input to the equalization target (10), After the input of the known input signal is completed, during the period in which the unknown input signal is input to the equalization target (10), the output of the equalization means (15) as a signal replacing the known input signal and the equalization target ( 10. The adaptive equalizer according to claim 7, further comprising a characteristic detection control device which supplies the output signal of 10) and causes the characteristic detection means (11) to perform continuous characteristic detection. 9. The equalization means (15) is means for weighting and adding outputs of a plurality of equalization circuits, and the equalization control means (14) is adapted to perform the plurality of equalization operations according to the characteristic detection result. 9. The adaptive equalizer according to claim 7, which outputs a weighting signal for the output of the circuit. 10. The adaptation according to claim 7, 8 or 9, characterized in that said characteristic detecting means (11) is a neural network or a learning device for acquiring the characteristic detecting function by learning the presented case. Equalizer. 11. The equalization control means (14) is a neural network or a learning device for acquiring the equalization control function by learning a presented case. Adaptive equalizer. 12. The equalization circuit as a part of the equalization means (15) is a neural network or a learning device for acquiring the equalization function by learning a presented case, respectively. Claims 7 and 8
Alternatively, the adaptive equalizer according to Item 9. 13. The characteristic detecting means (11) detects a transfer function of the equalization target (10) or a value of a real part and an imaginary part of the transfer function.
The adaptive equalizer according to 8, 9, 10, 11 or 12. 14. The characteristic detecting means (11) detects values of an amplitude frequency characteristic and an envelope characteristic of a transfer function of the equalization target (10).
The adaptive equalizer according to 9, 10, 11 or 12. 15. A detection parameter corresponding to a signal pattern of a known input signal to an equalization target (10) to be equalized as compensation for deterioration of the output signal is used to output the signal from the output signal corresponding to the input signal. Characteristic detection means (11) for detecting the characteristic of the equalization target (10), and the output signal of the equalization target (10) for an unknown input signal using the result detected by the characteristic detection means (11) An adaptive equalizer, comprising: an equalizing means (12) for performing equalization. 16. The known input signal is input to the characteristic detecting means (11) before the characteristic detecting means (11) during a period in which the known input signal is input to the equalization target (10). A signal and an output signal of the equalization target (10) are supplied, and as an alternative signal to the known input signal during a period in which an unknown input signal is input to the equalization target (10) after the input of the known input signal is completed. And a characteristic detection control device for supplying the output of the equalization means (12) and the output signal of the equalization target (10) to the characteristic detection means (11) for continuous characteristic detection. 16. The adaptive equalizer according to claim 15, wherein: 17. Adaptation according to claim 15 or 16, characterized in that said characteristic detecting means (11) comprises a neural network or a learning device for acquiring said characteristic detecting function by learning of posted cases. Chemist. 18. Adaptation according to claim 15 or 16, characterized in that said equalizing means (12) comprises a neural network or a learning device for acquiring said equalizing function by learning of posted cases. Chemist. 19. The characteristic detecting means (11) detects a transfer function of the equalization target (10) or a value of a real part and an imaginary part of the transfer function.
The adaptive equalizer according to 5, 16, 17 or 18. 20. The characteristic detecting means (11) detects values of an amplitude frequency characteristic and an envelope characteristic of a transfer function of the equalization target (10).
The adaptive equalizer according to 6, 17 or 18. 21. A characteristic detecting means for detecting a characteristic of an equalization target (10) from an output signal with respect to a known input signal to the equalization target (10) to be equalized as compensation for deterioration of the output signal. 11), post-processing means (16) for accumulating characteristic detection results of the characteristic detection means (11) for a plurality of times, and performing post-processing for the accumulated characteristic detection results, and the post-processing means (16) An equalizer (17) for equalizing an output signal of the equalization target (10) with respect to an unknown input signal by using the result post-processed by the adaptive equalizer. 22. Before the characteristic detecting means (11), the known input signal is input to the characteristic detecting means (11) during a period in which the known input signal is input to the equalization target (10). A signal and an output signal of the equalization target (10) are supplied, and as an alternative signal to the known input signal during a period in which an unknown input signal is input to the equalization target (10) after the input of the known input signal is completed. And a characteristic detection control device for supplying the output of the equalization means (17) and the output signal of the equalization target (10) to the characteristic detection means (11) for continuous characteristic detection. 22. The adaptive equalizer according to claim 21, wherein: 23. The characteristic detecting means (11) detects a characteristic of the equalization target (10) using a detection parameter corresponding to a signal pattern of the known input signal. Alternatively, the adaptive equalizer according to 22. 24. The characteristic detecting means (11) comprises a neural network or a learning device for acquiring the characteristic detecting function by learning a posted case. Adaptive equalizer. 25. The equalizing means (17) comprises a neural network or a learning device for acquiring the equalizing function by learning a posted case, according to claim 21, 22 or 23. Adaptive equalizer. 26. Adaptive equalizer according to claim 21, 22 or 23, characterized in that said post-processing means (16) comprises a statistical processing device. 27. The post-processing means (16) comprises a neural network or a learning device for acquiring the post-processing function by learning a presented case, according to claim 21, 22 or 23. Adaptive equalizer. 28. The characteristic detecting means (11) detects a transfer function of the equalization target (10) or a value of a real part and an imaginary part of the transfer function.
The adaptive equalizer according to 1, 22, 23, 24, 25, 26 or 27. 29. The characteristic detecting means (11) detects values of an amplitude frequency characteristic and an envelope characteristic of a transfer function of the equalization target (10).
The adaptive equalizer according to 2, 23, 24, 25, 26 or 27. 30. A characteristic detecting means for detecting a characteristic of an equalization target (10) from an output signal with respect to a known input signal to the equalization target (10) for which equalization for deterioration compensation of the output signal should be performed. (11) and an equalization means (12) for equalizing an output signal of the equalization target (10) with respect to an unknown input signal by using the result detected by the characteristic detection means (11), Of the input signal and the output signal of the equalizing means (12) are compared, and when an equalization error is detected, the output signal, the detection result of the characteristic detecting means (11), and the equalization target (10). An equalization error detecting and learning pattern accumulating means (18) for accumulating the known input signal as a correct output signal as a learning pattern in the period in which the equalization target (10) is not outputting an output signal. Using learning patterns The equalizing means (1
An adaptive equalizer, comprising: a learning control means (19) for causing 2) to perform learning. 31. A characteristic detecting device constituting the characteristic detecting means (11), an equalizing device constituting the equalizing means (12), the equalization error detecting and learning pattern accumulating means (18).
A recognition signal accumulating device, an equalization error determining device, a learning pattern accumulating device, and the learning control means (19).
31. The adaptive equalizer according to claim 30, further comprising an adaptive equalizer adjusting device including a learning control device constituting the above. 32. The known input is provided to the characteristic detecting means (11) before the characteristic detecting means (11) while the known input signal is being input to the equalization target (10). A signal and an output signal of the equalization target (10) are supplied, and as a signal replacing the known input signal in a period in which an unknown input signal is input to the equalization target (10) after the input of the known input signal is completed. And a characteristic detection control device for supplying the output of the equalization means (12) and the output signal of the equalization target (10) to the characteristic detection means (11) for continuous characteristic detection. The adaptive equalizer according to claim 30 or 31, characterized in that. 33. The characteristic detecting means (11) is a neural network or a learning device for acquiring the characteristic detecting function by learning a presented case. Adaptive equalizer. 34. The equalizing means (12) is a neural network or a learning device for acquiring the equalizing function by learning a presented case, according to claim 30, 31, or 32. Adaptive equalizer. 35. The characteristic detecting means (11) detects a transfer function of the equalization target (10) or a value of a real part and an imaginary part of the transfer function.
The adaptive equalizer according to 0, 31, 32, 33 or 34. 36. The characteristic detecting means (11) detects values of an amplitude frequency characteristic and an envelope characteristic of a transfer function of the equalization target (10).
The adaptive equalizer according to 1, 32, 33 or 34.