WO2022185505A1

WO2022185505A1 - Learning device for distortion compensation circuit, and distortion compensation circuit

Info

Publication number: WO2022185505A1
Application number: PCT/JP2021/008580
Authority: WO
Inventors: 暢彦安藤
Original assignee: 三菱電機株式会社
Priority date: 2021-03-05
Filing date: 2021-03-05
Publication date: 2022-09-09
Also published as: JP7418619B2; JPWO2022185505A1

Abstract

Provided is a distortion compensation circuit capable of suppressing divergence in an operation performed by means of an iterative operation in iterative learning control processing.　A learning device (12) for a distortion compensation circuit provided with a predistorter (11) into which a transmission signal is input is provided with a behavior model generating unit (121) for generating a behavior model of a power amplifier, a teacher data generating unit (122) including a behavior model unit (1223) for holding the generated behavior model, and an iterative learning control unit (1221) for generating teacher data serving as an input signal to the power amplifier, by performing iterative learning control processing using an error between the transmission signal and an output signal from the behavior model unit, and a coefficient updating unit (123) which is provided with a predistorter (1231) configured from a neural network, and which uses the transmission signal and the generated teacher data to calculate a weighting coefficient for the neural network, wherein the iterative learning control unit includes a gain determination processing unit (12211) for obtaining an updated current error on the basis of a current error in the generated behavior model and at least one previous gain.

Description

Learning device for distortion compensation circuit, and distortion compensation circuit

The present disclosure relates to a learning device for a distortion compensation circuit and a distortion compensation circuit.

A high-frequency circuit, which is a component of a radio transmitter, is composed of devices such as power amplifiers and filters, amplifies the power of a transmission signal, and outputs the power-amplified transmission signal to an antenna. When amplifying the power of a transmission signal, the quality of the transmission signal deteriorates due to distortion in the transmission signal due to the frequency characteristics and nonlinear characteristics of the high frequency circuit. Therefore, in order to prevent deterioration of signal quality, there is a method of using a distortion compensation circuit that compensates for the distortion that occurs in the high-frequency circuit by adding distortion to the transmission signal in the preceding stage of the high-frequency circuit so as to cancel out the distortion that occurs in the high-frequency circuit. be.

Conventionally, there is a method using a neural network as such a distortion compensation circuit. In a distortion compensation circuit using a neural network, the neural network is trained so as to obtain an inverse characteristic in which the relationship between the input and output of the high frequency circuit is opposite to that of the input and output. Learning requires teacher data, and the teacher data can be generated by iterative learning control (ILC) processing described in Non-Patent Document 1, for example.

However, according to the ILC processing described in Non-Patent Document 1, there is a problem that in repeated calculations when generating teacher data, errors do not decrease and the teacher data continues to increase and the calculation diverges.

The present disclosure has been made to solve the above-described problems, and one aspect of the present disclosure aims to provide a learning device for a distortion compensation circuit that can suppress the divergence of calculations in iterative learning control processing. and

One aspect of the distortion compensation circuit learning device according to the present disclosure is a distortion compensation circuit learning device including a predistorter to which a transmission signal is input, the behavior model generating a behavior model of a power amplifier. A generation unit, a behavior model unit that holds the generated behavior model, and an input signal to the power amplifier by repeatedly performing learning control processing using an error between the transmission signal and the output signal of the behavior model unit. and a teacher data generation unit comprising a repeated learning control unit that generates teacher data as and a predistorter composed of a neural network, using the transmission signal and the generated teacher data, the neural network and a coefficient updating unit that calculates the weighting coefficient of the iterative learning control unit, based on the current gain in the generated behavior model and at least one gain up to the previous time, the updated current A gain determination processing unit for obtaining an error is provided.

According to the learning device for the distortion compensation circuit of the present disclosure, it is possible to suppress the divergence of calculations due to repeated calculations in the repeated learning control process.

FIG. 1 is a block diagram showing a configuration example of a distortion compensation circuit. FIG. 2 is a block diagram showing a configuration example of an iterative learning control unit. FIG. 3 is a flow chart showing the operation of the distortion compensation circuit. FIG. 4A is a diagram illustrating a hardware configuration example of a distortion compensation circuit. FIG. 4B is a diagram showing another configuration example of the hardware of the distortion compensation circuit.

Embodiment 1.
Various embodiments according to the present disclosure will be described in detail below with reference to the drawings. FIG. 1 is a block diagram showing a configuration example of a distortion compensation circuit 1 according to Embodiment 1. As shown in FIG. A distortion compensation circuit 1 is a circuit that compensates for distortion of a signal amplified by a power amplifier 3 . As shown in FIG. 1, the distortion compensating circuit 1 comprises a predistorter 11 and a learning section 12 . A learning unit 12 is a learning device for the distortion compensation circuit 1 . The learning unit 12 includes a behavior model generation unit 121 , a teacher data generation unit 122 and a coefficient update unit 123 . The learning unit 12 may be always incorporated in the distortion compensation circuit 1 as part of the distortion compensation circuit 1, or may be incorporated only during learning. By constantly incorporating the learning unit 12 into the distortion compensation circuit 1, distortion can be adaptively compensated when the characteristics of the power amplifier 3 change.

(pre-distorter)
The predistorter 11 is a predistorter that digitally compensates for distortion of the transmission signal s. The predistorter 11 is configured by a neural network, and sets a weighting factor obtained by a factor updater 123 (to be described later) in the neural network of the predistorter 11 . The predistorter 11 performs distortion compensation processing on the transmission signal s according to the weighting factor, and sends the transmission signal after the distortion compensation processing to the digital-to-analog converter (D/A converter) 2 and the behavior model generator 121. Output. The signal output to the D/A converter 2 is converted into an analog signal and input to the power amplifier 3, which amplifies the power of the input analog signal. By subjecting the input signal to the power amplifier 3 to distortion compensation processing, the occurrence of distortion in the output signal of the power amplifier 3 is suppressed.

The power amplifier 3 amplifies the power of the analog signal output from the D/A converter 2 . The A/D converter 4 converts the analog signal output from the power amplifier 3 into a digital signal and outputs the digital signal to the behavior model generator 121 . The analog signal output by the power amplifier 3 converted by the A/D converter 4 may be attenuated by an attenuator (not shown).

(behavior model generator)
The behavior model generator 121 generates a behavior model, which is a model of the behavior of the power amplifier 3, using input/output signals of the power amplifier 3. FIG. In order to allow the power amplifier 3 to be used in a wide band where the memory effect occurs, the behavior model generation unit 121 expresses the behavior model BM of the power amplifier 3 by, for example, a memory polynomial, and calculates the coefficients of the polynomial using a technique such as the least squares method. Calculated by The behavior model generator 121 outputs the generated behavior model BM to the teacher data generator 122 . A detailed description of the memory polynomial is omitted because it is described in, for example, Non-Patent Document 2.

Non-patent document 2: L. Ding, G. T. Zhou, D. R. Morgan, Z. Ma, J. S. Kenney, J. Kim, C. R. Giardina, "A robust digital baseband predistorter constructed using memory polynomials”, IEEE Trans. Commun., vol. 52, no. 1, pp. 159-165, Jan. 2004.

Generating a behavior model using the method of least squares is described, for example, in Non-Patent Document 3, so a detailed description will be omitted.

Non-Patent Document 3: Amandeep Singh Sappal, ``Simplified Memory Polynomial modeling of Power Amplifier,'' 2015 International Conference and Workshop on Computing and Communication (IEMCON), 2015.

(Training data generator)
The teacher data generation unit 122 performs iterative learning control (ILC) processing using the behavior model BM generated by the behavior model generation unit 121, thereby generating an optimal input signal to the power amplifier 3. Generate some training data. As shown in FIG. 1 , the teacher data generation unit 122 includes a behavior model unit 1223 , a synthesis unit 1222 and an iterative learning control (ILC) unit 1221 . The behavior model section 1223 holds the behavior model BM generated by the behavior model generation section 121 .

First, an outline of the teacher data generation unit 122 will be described. Let s be the transmission signal, u _i be the input signal to the behavior model BM in the i-th calculation, z _i be the output signal of the behavior model BM, and N be the number of data points. An input signal u _i+1 to the model BM is calculated according to the following equation (1).

u _i+1 =u _i +G(u _i ) ⁻¹ e _i (1)

Here, G(u _i ) is a diagonal matrix represented by the following equation (2).

G(u _i )=diag{G[u _i (0)], . . . , G[u _i (N−1)]}
(2)

Each diagonal component of the diagonal matrix is represented by the following equation (3). In formula (3), n is an integer from 0 to N−1.

G[u _i (N)]=z _i (n)/u _i (n) (3)

Also, _ei is an error represented by the following equation (4). This error _ei is calculated by the synthesizing unit 1222 .

e _i =s−z _i (4)

Sufficient iterative calculations by the iterative learning control unit 1221 reduce the error between the output signal _zi of the behavior model BM and the transmission signal s. A value finally obtained by repeating the operation a predetermined number of times becomes the teacher data. The teacher data is output to the coefficient updating section 123 .

The iterative learning control unit 1221 performs additional processing as described below in the iterative learning control processing described above. In order to perform such additional processing, the iterative learning control unit 1221 includes a gain determination processing unit 12211, a teacher data calculation unit 12212, and a teacher data suppression unit 12213, as shown in FIG. Processing performed by each functional unit of the iterative learning control unit 1221 will be described below with reference to FIGS. 2 and 3. FIG. 3 is a diagram showing a flow chart of the teacher data generation unit 122 and the iterative learning control unit 1221. As shown in FIG.

In step ST31, the teacher data generator 122 calculates the output signal z _i of the behavior model BM, the diagonal matrix component G[u _i (n)], and the error _ei . More specifically, the behavior model unit 1223 receives the input signal _ui as an input of the behavior model BM and calculates the output signal _zi . Note that the initial value u1 of the input signal can take _any value as long as the behavior model BM is operable. Synthesis section 1222 calculates error _ei from transmission signal s and output signal _zi of behavior model section 1223 . The behavior model unit 1223 or the iterative learning control unit 1221 calculates the component G[u _i (n)] from the input signal u _i and the output signal z _i according to Equation (3). When the iterative learning control unit 1221 calculates the component G[u _i (n)], the output signal z _i of the behavior model unit 1223 is also output to the iterative learning control unit 1221 in addition to the synthesizing unit 1222 . Since the component G[u _i (n)] can be considered as the gain of the power amplifier 3, the component G[u _i (n)] will be referred to as the gain G[u _i (n)] in the following description.

In step ST32, the gain determination processing section 12211 determines whether the current gain G[u _i (n)] is smaller than the previous gain G[u _i-1 (n)]. If YES in step ST32, the process proceeds to step ST33. In step ST33, the gain determination processing section 12211 updates the error in the current cycle according to the formula e _i (n)=e _i (n)×α. α is an arbitrary value that satisfies 0<α<1. By updating the error in this way, the error is adjusted. After the error is updated, the process proceeds to step ST34. Moreover, in the case of NO in step ST32, the process proceeds to step ST34. Here, the current gain G[u _i (n)] is compared with the previous gain G[ _u _i-1 (n)]. , may be compared with the average of the gains of a plurality of past times before the current time. In this way, gain determination processing section 12211 obtains the updated current error based on the current gain and at least one previous gain.

In step ST34, the teacher data calculator 12212 calculates the input signal u _i+1 to the i+1-th behavior model BM according to the above equation (1).

In step ST35, the teacher data suppression section 12213 determines whether the absolute value |u _i+1 | of the input signal is greater than a preset threshold value M _th . If YES in step ST35, the process proceeds to step ST36. In step ST36, teacher data suppression section 12213 sets threshold M _th to the absolute value |u _i+1 | of the input signal. After setting, the process proceeds to step ST37. If NO in step ST35, the process proceeds to step ST37.

In step ST37, if the number of iterations i is greater than the maximum number of times, the process is terminated and teacher data |u _i+1 | is obtained. If the number of iterations is not greater than the maximum number of times, the number of iterations i is incremented by 1 by the iterative learning control section 1221 in step ST38, and the processing of steps ST31 to ST37 is repeated.

In this way, when the current gain G[u _i (n)] is smaller than the previous gain G[u _i-1 (n)], the gain determination processing section 12211 determines the formula e _i (n)=e The error is adjusted by updating the error according to _i (n)×α (steps ST32 and ST33). As a result, even when G[u _i (n)] becomes smaller than the previous value G[u _i-1 (n)], the effect of the error can be reduced, and the teaching data can be prevented from becoming large. . Note that when G[u _i (n)] becomes larger than the previous value G[u _i-1 (n)], G(u _i (n)) ⁻ Since ¹ is included, the size of the training data is suppressed.

If the teacher data u _i-1 is an input signal to the saturation region of the behavior model BM, the teacher data u _i-1 cannot suppress the occurrence of distortion. This is because signals in the saturation region cannot be compensated. Therefore, the error does not become small. Since the error does not become smaller, the teacher data u _i operates to become larger than u _i−1 . If this is repeated, the teacher data will diverge.

As mentioned above, if the gain G[u _i (n)] is less than the gain G[u _i-1 (n)], adjust the error according to the equation e _i (n)=e _i (n)×α can prevent the teacher data from entering the saturation region of the behavior model BM.

Further, as described above, when the absolute _value of the input signal | _u _i+1 | Set to u _i+1 |. As a result, since an upper limit can be set for the teacher data, it is possible to suppress divergence of the teacher data during repeated calculations.

(Coefficient updating part)
The coefficient updating unit 123 includes a predistorter 1231 and a synthesizing unit 1232 . The predistorter 1231 is composed of a neural network like the predistorter 11 . The coefficient updating unit 123 uses the teacher data generated by the teacher data generating unit 122 to perform supervised learning for obtaining weighting coefficients of the neural network operating as the predistorter 11 . More specifically, using the transmission signal s used by the teacher data generation unit 122 as the input signal of the neural network and the teacher data generated by the teacher data generation unit 122 as the output signal of the neural network, the neural network weight Learn to find coefficients. For learning of the neural network, a generally known method such as the Levenberg-Marquardt method (Non-Patent Document 4) or ADAM (Non-Patent Document 5) may be used, so a detailed description of the learning method is omitted. . The weighting coefficient obtained by the coefficient updating unit 123 is output to the predistorter 11 .

Non-Patent Document 4 (Levenberg-Marquardt Method): Hagan, M. T., and M. Menhaj, "Training feed-forward networks with the Marquardt algorithm," IEEE Transactions on Neural Networks, Vol. 5, No. 6, 1999, pp. 989-993, 1994.

Non-Patent Document 5 (ADAM): Kingma, Diederik, and Jimmy Ba. "Adam: A method for stochastic optimization." arXiv preprint arXiv:1412.6980 (2014).

All or part of the functional units of the distortion compensation circuit 1 are realized, for example, by a computer having a processor 41 and a memory 42 as shown in FIG. 4A. These functional units are realized by reading out and executing the programs stored in the memory 42 by the processor 41 . Programs may be implemented as software, firmware, or a combination of software and firmware. Examples of the memory 42 include non-volatile or volatile semiconductors such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically-EPROM), etc. Memory, magnetic disk, flexible disk, optical disk, compact disk, mini disk, DVD are included.

As another example, all or part of the functional units of the distortion compensation circuit 1 may be realized by a processing circuit 43 as shown in FIG. 4B instead of the processor 41 and memory 42. In this case, the processing circuit 43 is, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof. is.

Since the distortion compensation circuit 1 is configured as described above, it is possible to prevent the teacher data from diverging in repeated calculations in the ILC process, and to create highly accurate teacher data.

Note that there are no restrictions on the configuration of the neural network in the first embodiment. The expression used to express the behavior model may not be a memory polynomial, but may be a Volterra series.

<Appendix>
Some of the various aspects of the embodiments described above are summarized below.

(Appendix 1)
A learning device (12) for a distortion compensation circuit of appendix 1 is a learning device for a distortion compensation circuit comprising a predistorter (11) to which a transmission signal is input, and generates a behavior model of a power amplifier. A behavior model generation unit (121), a behavior model unit (1223) holding the generated behavior model, and an error between the transmission signal and the output signal of the behavior model unit are used to repeatedly perform learning control processing. , a repetitive learning control unit (1221) that generates teacher data as an input signal to the power amplifier, and a teacher data generation unit (122), and a predistorter (1231) composed of a neural network, a coefficient updating unit (123) that calculates weighting coefficients of the neural network using the transmission signal and the generated teacher data, and the iterative learning control unit updates the current A gain determination processing unit (12211) is provided for obtaining the updated current error based on the gain and at least one gain up to the previous time.

(Appendix 2)
The learning device for the distortion compensation circuit of Supplementary Note 2 is the learning device for the distortion compensation circuit of Supplementary Note 1, wherein the gain determination processing unit determines that the current gain in the generated behavior model is higher than the previous gain. is also small, the updated current error is obtained by multiplying the current error by an arbitrary value greater than 0 and less than 1.

(Appendix 3)
The learning device for the distortion compensating circuit of Supplementary Note 3 is the learning device for the distortion compensating circuit of Supplementary Note 1, wherein the gain determination processing unit determines whether the current gain in the generated behavior model has been obtained a plurality of times in the past. If it is less than the average gain, the updated current error is obtained by multiplying the current error by any value greater than 0 and less than 1.

(Appendix 4)
The learning device for the distortion compensation circuit of appendix 4 is the learning device for the distortion compensation circuit of any one of appendices 1 to 3, wherein the iterative learning control unit uses the updated current error It further comprises a teacher data calculation unit (12212) for calculating the next teacher data according to the formula (1).

(Appendix 5)
The learning device for the distortion compensation circuit of Supplementary Note 5 is the learning device for the distortion compensation circuit of any one of Supplements 1 to 4, wherein the iterative learning control unit calculates the absolute value of the next teacher data calculated. It further comprises a teacher data suppression unit (12213) which, when the value is greater than a preset threshold value, sets the threshold value as the absolute value of the calculated next teacher data.

(Appendix 6)
The distortion compensation circuit (1) of Appendix 6 includes a predistorter (11) configured by a neural network and to which a transmission signal is input, and the neural network of the predistorter is any one of Appendixes 1 to 5. The predistorter (11) has the weighting factor obtained by the learning device for the distortion compensation circuit, performs distortion compensation processing on the transmission signal according to the weighting factor, and after distortion compensation processing output the transmission signal of

It should be noted that the embodiments can be combined, and each embodiment can be modified or omitted as appropriate.

A distortion compensation circuit according to the present disclosure can be used as a distortion compensation circuit that imparts distortion to a transmission signal that is input to a radio frequency circuit of a wireless transmitter.

1 distortion compensation circuit, 2 D/A converter, 3 power amplifier, 4 A/D converter, 11 predistorter, 12 learning unit (learning device), 121 behavior model generation unit, 122 teacher data generation unit, 123 coefficient Update unit 1221 Iterative learning control unit 1222 Synthesis unit 1223 Behavior model unit 1231 Predistorter 1232 Synthesis unit 12211 Gain determination processing unit 12212 Teacher data calculation unit 12213 Teacher data suppression unit BM Behavior model.

Claims

A learning device for a distortion compensation circuit comprising a predistorter to which a transmission signal is input,
a behavior model generator that generates a behavior model of the power amplifier;
a behavior model unit that holds the generated behavior model; and teacher data as an input signal to the power amplifier by repeatedly performing learning control processing using an error between the transmission signal and the output signal of the behavior model unit. a teacher data generation unit comprising an iterative learning control unit that generates
a coefficient updating unit that includes a predistorter configured with a neural network and calculates a weighting coefficient of the neural network using the transmission signal and the generated teacher data;
with
The iterative learning control unit comprises a gain determination processing unit that obtains an updated current error based on the current gain and at least one previous gain in the generated behavior model.
A learning device for distortion compensation circuits.
If the current gain in the generated behavior model is smaller than the previous gain, the gain determination processing unit multiplies the current error by an arbitrary value greater than 0 and less than 1. 2. A learning device for a distortion compensation circuit according to claim 1, wherein the current error is determined.
When the current gain in the generated behavior model is smaller than the average of the gains obtained a plurality of times in the past, the gain determination processing unit multiplies the current error by an arbitrary value greater than 0 and less than 1. 2. A learning device for a distortion compensation circuit according to claim 1, wherein said updated current error is obtained.
4. The distortion according to any one of claims 1 to 3, wherein the iterative learning control unit further comprises a teacher data calculation unit that calculates next teacher data using the updated current error by the following equation: A learning device for the compensator.

u i+1 = u i +G(u i ) −1 e i

here,

G(u i )=diag{G[u i (0)], . . . , G[u i (N−1)]}
G[u i (N)]=z i (n)/u i (n)
e i =s−z i

where u is the teacher data and input signal to the behavior model, s is the transmission signal, z is the output signal of the behavior model, i is the iteration number in the iterative operation, and N is the number of data points.
When the absolute value of the calculated next teacher data is greater than a preset threshold, the repeated learning control unit sets the threshold as the absolute value of the calculated next teacher data. 5. A learning device for a distortion compensation circuit according to claim 4, further comprising a data suppressor.
Equipped with a predistorter that is composed of a neural network and receives a transmission signal,
The neural network of the predistorter has the weighting coefficient obtained by the learning device for distortion compensation circuit according to claim 1,
The predistorter performs distortion compensation processing on the transmission signal according to the weighting factor, and outputs the transmission signal after the distortion compensation processing.
Distortion compensation circuit.