JPWO2020166005A1 - Machine learning device, signal specification identification device, machine learning method, control circuit and storage medium - Google Patents

Abstract

The machine learning device (1) is a time-series feature amount calculation unit (11) that calculates a time-series feature amount which is a time-series variation value of the feature amount of the learning signal from the in-phase signal and the orthogonal signal included in the learning signal. ), And a signal specification learning unit (13) that learns the signal specifications according to the training data set created based on the respective waveforms of the in-phase signal and the orthogonal signal and the time-series feature amount. It is characterized by.

Description

The present invention relates to a machine learning device for identifying signal specifications of a received signal, a signal specification identification device, and a machine learning method.

As a method of identifying signal specifications such as a modulation method of a received signal, a method using machine learning has been proposed. In Non-Patent Document 1, the in-phase signal and the orthogonal signal included in the received signal are learned by a convolutional neural network, and the learning result is used to input the common-mode signal and the orthogonal signal. Techniques for identifying modulation schemes are disclosed.

Timothy J. O'Shea, Johnathan Corgan, and T. Charles Clancy “Convolutional Radio Modulation Recognition Networks”, CoRR, vol.abs / 1602, no.04105, pp.1-15, Mar.2016.

However, according to the above-mentioned conventional technique, there is a problem that the amount of data set required for learning increases.

The present invention has been made in view of the above, and an object of the present invention is to obtain a machine learning device capable of reducing the amount of data set required for learning.

In order to solve the above-mentioned problems and achieve the object, the machine learning device according to the present invention is a time-series variation value of the feature amount of the learning signal from the in-phase signal and the orthogonal signal included in the learning signal. A signal for learning signal specifications according to a time-series feature amount calculation unit that calculates a time-series feature amount, a training data set created based on each waveform of an in-phase signal and an orthogonal signal, and a time-series feature amount. It is characterized by having a specification learning department.

The machine learning device according to the present invention has an effect that the number of data sets required for learning can be reduced.

The figure which shows the functional structure of the signal specification identification system which concerns on Embodiment 1 of this invention. The figure which shows the structural example of the time-series feature amount calculation part shown in FIG. The figure which shows the structural example of the quasi-synchronous detection part shown in FIG. The figure which shows the dedicated hardware configuration for realizing the function of the signal specification identification system shown in FIG. The figure which shows the hardware configuration which executes the software which realizes the function of the signal specification identification system shown in FIG. A flowchart showing the operation of the machine learning device shown in FIG. A flowchart showing the operation of the signal specification identification device shown in FIG. The figure which shows the functional structure of the signal specification identification system which concerns on Embodiment 2 of this invention. The figure which shows the functional structure of the signal specification identification system which concerns on Embodiment 3 of this invention. The figure which shows the functional structure of the noise removing part shown in FIG. The figure which shows the structural example of the noise removing part shown in FIG. The figure which shows the functional structure of the signal specification identification system which concerns on Embodiment 4 of this invention. A flowchart showing the operation of the machine learning device shown in FIG. A flowchart showing the operation of the signal specification identification device shown in FIG.

Hereinafter, the machine learning device, the signal specification identification device, and the machine learning method according to the embodiment of the present invention will be described in detail with reference to the drawings. The present invention is not limited to this embodiment.

Embodiment 1.
FIG. 1 is a diagram showing a functional configuration of the signal specification identification system 100 according to the first embodiment of the present invention. The signal specification identification system 100 includes a machine learning device 1, a learning parameter storage unit 2 that stores learning parameters that are learning results of the machine learning device 1, and a signal that identifies signal specifications of a received signal using the learning results. It has a specification identification device 3 and a quasi-synchronous detection unit 4 that obtains in-phase signals and orthogonal signals included in the received signal. The signal specifications are radio transmission methods such as AM (Amplitude Modulation) / FM (Frequency Modulation), FSK (Frequency Shift Keying), and QPSK (Quadrature Phase Shift Keying). Modulation methods such as transmodulation), DSSS (Direct Sequence Spread Spectrum), communication standards including secondary modulation such as OFDM (Orthogonal Frequency Division Multiplexing). Examples of communication standards include IEEE (Institute of Electrical and Electronics Engineers) 802.11a, IEEE802.11g, and IEEE802.1.6.

The machine learning device 1 learns learning parameters for identifying signal specifications using a learning data set. The learning data set includes a learning signal and a label indicating the signal specifications associated with the learning signal. The plurality of learning signals are the IQ signal, which is a combination of the in-phase signal and the orthogonal signal obtained as a result of performing forward synchronous detection on the actually acquired received waveform, the in-phase signal simulated by the computer simulation, and the orthogonal signal. It is a signal or the like and is prepared in advance. The learning signal included in the learning data set is input to each of the time series feature amount calculation unit 11 and the coupling unit 12, and the label included in the learning data set is input to the signal specification learning unit 13.

The machine learning device 1 has a time-series feature amount calculation unit 11, a coupling unit 12, and a signal specification learning unit 13. The time-series feature amount calculation unit 11 calculates the time-series feature amount, which is the time-series variation value of the feature amount of the learning signal, from the in-phase signal and the orthogonal signal included in the learning signal. The time-series feature amount calculation unit 11 inputs the calculated time-series feature amount to the coupling unit 12. The coupling unit 12 combines the in-phase signal and the orthogonal signal included in the learning signal with the time-series feature amount to generate a training data set used by the signal specification learning unit 13 for machine learning. The coupling unit 12 inputs the training data set, which is the data after coupling, to the signal specification learning unit 13. The signal specification learning unit 13 learns the signal specifications according to the training data set created based on the respective waveforms of the in-phase signal and the orthogonal signal and the time-series feature amount. The learning result of the signal specification learning unit 13 shows the relationship between the training data set which is input data and the signal specification, and is a learning parameter for identifying the signal specification from the training data set. The signal specification learning unit 13 stores the learning parameter, which is the learning result, in the learning parameter storage unit 2.

The learning parameter storage unit 2 is a storage device that stores learning parameters that are learning results of the machine learning device 1. The learning parameters stored in the learning parameter storage unit 2 can be read out from the signal specification identification device 3.

The quasi-synchronous detection unit 4 synchronously demodulates the received signal, reproduces the carrier wave, and compensates for frequency deviation and phase deviation. The quasi-synchronous detection unit 4 inputs the in-phase signal and the orthogonal signal obtained after the quasi-synchronous detection to the signal specification identification device 3.

The signal specification identification device 3 includes a time-series feature amount calculation unit 31, a coupling unit 32, and a signal specification identification unit 33. The in-phase signal and the orthogonal signal output by the quasi-synchronous detection unit 4 are input to the time-series feature amount calculation unit 31 and the coupling unit 32, respectively. The time-series feature amount calculation unit 31 calculates the time-series feature amount, which is the time-series variation value of the feature amount of the received signal, from the in-phase signal and the orthogonal signal included in the received signal. It is desirable that the type of the time-series feature amount calculated by the time-series feature amount calculation unit 31 is the same as the time-series feature amount calculated by the time-series feature amount calculation unit 11. The time-series feature amount calculation unit 31 inputs the calculated time-series feature amount to the coupling unit 32. The coupling unit 32 combines the time-series feature amount input from the time-series feature amount calculation unit 31 with the in-phase signal and the orthogonal signal to generate input data to the signal specification identification unit 33. The signal specification identification unit 33 identifies the signal specifications applied to the received signal based on the input data from the coupling unit 32 by using the learning parameters stored in the learning parameter storage unit 2. The signal specification identification unit 33 outputs the identification result.

The above configuration is an example. For example, in FIG. 1, the learning parameter storage unit 2 is provided outside the machine learning device 1 and the signal specification identification device 3, but may be provided inside the machine learning device 1 or a signal. It may be provided inside the specification identification device 3. Further, although the machine learning device 1 and the signal specification identification device 3 are different devices in FIG. 1, the function of the signal specification identification system 100 can be realized on one device. In this case, the time-series feature amount calculation unit 11 and the time-series feature amount calculation unit 31, the coupling unit 12 and the coupling unit 32, and the signal specification learning unit 13 and the signal specification identification unit 33 are each provided by the same hardware. It can also be realized.

FIG. 2 is a diagram showing a configuration example of the time-series feature amount calculation units 11 and 31 shown in FIG. The time-series feature amount calculation units 11 and 31 include an amplitude calculation unit 111 and a phase calculation unit 112. An in-phase signal and an orthogonal signal are input to each of the amplitude calculation unit 111 and the phase calculation unit 112. The amplitude calculation unit 111 calculates the amplitude of the received signal. When the amplitude calculation unit 111 is an in-phase signal I and an orthogonal signal Q, the amplitude A can be expressed by the following mathematical formula (1).

The phase calculation unit 112 can calculate the phase of the received signal. When the in-phase signal I and the orthogonal signal Q are used, the phase φ can be expressed by the following mathematical formula (2).

In addition to the above, the time-series feature amount calculation units 11 and 31 can also calculate the amplitude and phase of the in-phase signal I and the orthogonal signal Q squared to the complex number. Further, the feature amounts calculated by the time-series feature amount calculation units 11 and 31 are not limited to the amplitude and the phase.

FIG. 3 is a diagram showing a configuration example of the quasi-synchronous detection unit 4 shown in FIG. The quasi-synchronous detection unit 4 includes an oscillator 41 that generates an oscillation signal for frequency conversion of a received signal, a π / 2 phase shifter 42 that shifts the phase of the oscillation signal output from the oscillator 41 by π / 2. A multiplier 43 that multiplies the received signal and the oscillated signal, a multiplier 44 that multiplies the received signal and the oscillated signal with a π-phase shift of half, and unnecessary waves and unnecessary waves from the signals output from the multipliers 43 and 44. It has LPFs (Low Pass Filters) 45 and 46 for removing noise.

FIG. 4 is a diagram showing a dedicated hardware configuration for realizing the function of the signal specification identification system 100 shown in FIG. The processing circuit 90 shown in FIG. 4 is, for example, a single circuit, a decoding circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof. Applicable.

FIG. 5 is a diagram showing a hardware configuration for executing software that realizes the functions of the signal specification identification system 100 shown in FIG. The function of the signal specification identification system 100 can be realized by using the quasi-synchronous detection circuit 91, the processor 92, the memory 93, and the storage 94. The quasi-synchronous detection circuit 91 is a circuit having the configuration of the quasi-synchronous detection unit 4 shown in FIG. The processor 92 is composed of, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a microprocessor, a microcontroller, a DSP (Digital Signal Processor), or the like. The memory 93 is, for example, a semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable ROM), or EEPROM (registered trademark) (Electrically EPROM), a magnetic disk, or a flexible disk. , Etc. are applicable.

Each function of the machine learning device 1 and the signal specification identification device 3 shown in FIG. 1 is realized by software, firmware, or a combination of software and firmware. Software and firmware are described as programs, stored in the memory 93, and each function is realized by the processor 92 reading and executing the program stored in the memory 93. Further, when recording information while the program is being executed, the memory 93 may be used to hold the data, or the storage 94 may be used to hold the data. The learning parameter storage unit 2 shown in FIG. 1 holds learning parameters using the storage 94.

FIG. 6 is a flowchart showing the operation of the machine learning device 1 shown in FIG. The machine learning device 1 acquires learning data in which the learning signal included in the learning data set and the label are associated with each other (step S101). The time-series feature amount calculation unit 11 calculates the time-series feature amount from the learning signal after the acquired learning data (step S102).

The connecting unit 12 combines the calculated time-series features and the learning data to generate a training data set (step S103). The signal specification learning unit 13 uses the generated training data set as input data by using a classifier using a neural network, learns the signal specifications indicated by the associated labels, and responds from the input data. The parameters for outputting the signal specifications to be output are learned (step S104). The learning parameter, which is the learning result, is stored in the learning parameter storage unit 2.

FIG. 7 is a flowchart showing the operation of the signal specification identification device 3 shown in FIG. The signal specification identification device 3 acquires the in-phase signal and the orthogonal signal included in the received signal (step S201). The time-series feature amount calculation unit 31 calculates the time-series feature amount from the in-phase signal and the orthogonal signal (step S202). The coupling unit 32 combines the calculated time-series feature amount with the in-phase signal and the orthogonal signal and inputs them to the signal specification identification unit 33 (step S203). The signal specification identification unit 33 identifies the signal specifications of the received signal based on the input data and the learning result of the machine learning device 1 stored in the learning parameter storage unit 2 (step S204).

As described above, in the signal specification identification system 100 according to the first embodiment of the present invention, in addition to the in-phase signal and the orthogonal signal, a time-series feature amount which is a time-series variation value of the signal feature amount is signaled. The signal specifications are learned as an input to the specification learning unit 13. By adopting such a configuration, it is possible to reduce the number of samples required and the number of data sets required at the time of learning, as compared with learning the signal specifications by inputting only the in-phase signal and the orthogonal signal. ..

Embodiment 2.
FIG. 8 is a diagram showing a functional configuration of the signal specification identification system 200 according to the second embodiment of the present invention. Hereinafter, the parts different from the signal specification identification system 100 according to the first embodiment will be mainly described, and detailed description of the parts similar to the signal specification identification system 100 will be omitted. The signal specification identification system 200 includes a machine learning device 1a, a learning parameter storage unit 2, a signal specification identification device 3a, and a quasi-synchronous detection unit 4.

The machine learning device 1a includes a time-series feature amount calculation unit 11, a coupling unit 12, and a signal specification learning unit 13a. The machine learning device 1a acquires learning data including a learning signal, supplementary information, and a label from the learning data set. The supplementary information is information that can be calculated from the learning signal and is acquired in advance. The supplementary information is input to the signal specification learning unit 13a. The signal specification learning unit 13a learns the signal specifications according to the training data set created based on the respective waveforms of the in-phase signal and the orthogonal signal, the time-series feature amount, and the supplementary information.

The signal specification identification device 3a includes a time-series feature amount calculation unit 31, a coupling unit 32, a signal specification identification unit 33a, and a supplementary information calculation unit 34. The quasi-synchronous detection unit 4 inputs the in-phase signal and the orthogonal signal included in the received signal to the supplementary information calculation unit 34. The supplementary information calculation unit 34 calculates the same type of supplementary information as the supplementary information input to the signal specification learning unit 13a from the received signal. The supplementary information is information that supplements the characteristics of the signal, and may include an estimated SNR (Signal to Noise Ratio) and an estimated frequency error. In the calculation in the supplementary information calculation unit 34, the in-phase signal and the orthogonal signal are FFT (Fast Fourier Transform) and the ratio of the maximum value to the minimum value in the frequency domain is used as the estimated SNR, or the deviation of the center frequency from the frequency band is estimated. In addition, it is conceivable to estimate the SNR and frequency band using a neural network, but the present invention is not limited to this.

The signal specification identification unit 33a identifies the signal specifications by using the supplementary information output by the supplementary information calculation unit 34 as input data in addition to the coupling data output by the coupling unit 32. As in the first embodiment, the configuration shown in FIG. 8 is an example, and the learning parameter storage unit 2 is provided outside the machine learning device 1a and the signal specification identification device 3a in FIG. May be provided inside the machine learning device 1a, or may be provided inside the signal specification identification device 3a. Further, in FIG. 8, although the machine learning device 1a and the signal specification identification device 3a are different devices, the function of the signal specification identification system 200 can be realized on one device. In this case, the time-series feature amount calculation unit 11 and the time-series feature amount calculation unit 31, the coupling unit 12 and the coupling unit 32, and the signal specification learning unit 13a and the signal specification identification unit 33a are made of the same hardware. It can also be realized.

As described above, according to the signal specification identification system 200 according to the second embodiment of the present invention, supplementary information is added to the input data of the signal specification learning unit 13a and the signal specification identification unit 33a. It is possible to further reduce the number of data sets required for training.

Embodiment 3.
FIG. 9 is a diagram showing a functional configuration of the signal specification identification system 300 according to the third embodiment of the present invention. The signal specification identification system 300 includes a machine learning device 1b, a learning parameter storage unit 2, a signal specification identification device 3b, a quasi-synchronous detection unit 4, and a noise removal learning parameter storage unit 5.

The machine learning device 1b includes a time-series feature amount calculation unit 11, a coupling unit 12, a signal specification learning unit 13, a signal selection unit 14, and a noise removing unit 15. The signal selection unit 14 and the noise removal unit 15 are arranged in front of the time-series feature amount calculation unit 11.

The signal selection unit 14 selects a noise removal learning signal to be used for learning a noise removal signal obtained by removing a noise component from the learning signal from the learning signals included in the learning data set. The signal selection unit 14 inputs the selection result and the learning signal to the noise removal unit 15. The signal selection unit 14 can select the noise removal learning signal based on the signal-to-noise ratio of the learning signal. For example, the signal selection unit 14 selects a learning signal having a signal-to-noise ratio equal to or greater than a threshold value as a noise removal learning signal. Alternatively, when the learning signal included in the learning data set is previously given a flag indicating whether or not it is a noise removal learning signal, the signal selection unit 14 is used for noise reduction learning based on the flag. The signal can be selected.

The noise removing unit 15 learns the noise removing signal obtained by removing the noise component from the learning signal, and removes the noise component from the learning signal by using the learning result of the noise removing signal. Specifically, the noise reduction unit 15 refers to the learning signal selected as the noise reduction learning signal based on the input learning signal and the selection result of the signal selection unit 14. To learn the denoising signal using. The noise removing unit 15 stores the learning result in the noise removing learning parameter storage unit 5. The noise removing unit 15 removes a noise component from the learning signal by using the learning result for the learning signal that is not selected as the noise removing learning signal. The noise removal unit 15 inputs a noise removal signal, which is a learning signal from which noise components have been removed, to the time series feature amount calculation unit 11 and the coupling unit 12.

The time-series feature amount calculation unit 11, the coupling unit 12, and the signal specification learning unit 13 are the same as those in the first embodiment except that the processing is performed using the signal after noise removal, and thus detailed description thereof will be omitted. ..

FIG. 10 is a diagram showing a functional configuration of the noise removing unit 15 shown in FIG. The noise removing unit 15 includes a learning noise adding unit 151, a normalizing processing unit 152, a noise removing device learning unit 153, a noise removing learning parameter storage unit 154, a normalizing processing unit 155, and a noise removing device 156. Has.

A noise removal learning signal is input to the learning noise addition unit 151. Among the learning signals, the one having a low noise component contained is selected as the noise elimination learning signal. The learning noise addition unit 151 adds a noise component to the input noise removal learning signal. The learning noise addition unit 151 inputs a signal to which a noise component is added to the normalization processing unit 152. A signal to which a noise component is added and a noise removal learning signal before the noise component is added are input to the normalization processing unit 152. The normalization processing unit 152 inputs the result of normalization processing to each of the signal to which the noise component is added and the noise removal learning signal to the noise removal device learning unit 153.

The noise remover learning unit 153 includes a self-encoder, and a normalized signal after noise addition is input to the self-encoder, and the self-encoder is for noise removal based on the signal after noise addition. According to the data set, the normalized noise removal learning signal before noise addition is learned as the noise reduction signal. In other words, the noise remover learning unit 153 learns the parameters so as to output the noise removal learning signal before the noise addition when the noise removal learning signal after the noise addition is input. The noise remover learning unit 153 stores the learning result in the noise removal learning parameter storage unit 154, outputs the learning result, and stores the learning result in the noise removal learning parameter storage unit 5.

An autoencoder is a neural network in which the dimensions of the input and output are the same, and the method of estimating the one with noise added to the input before the noise is added is the noise reduction autoencoder. Known as (Denoising Autoencoder).

When an input signal other than the noise removal learning signal is input, the normalization processing unit 155 performs normalization processing, and the normalization processing unit 155 inputs the processed signal to the noise removal device 156. The noise remover 156 performs noise removal using the learning parameters stored in the noise removal learning parameter storage unit 154, and outputs a signal after noise removal.

Returning to the description of FIG. The signal specification identification device 3b includes a time-series feature amount calculation unit 31, a coupling unit 32, a signal specification identification unit 33, and a noise removal unit 35. The noise removing unit 35 is arranged in front of the time-series feature amount calculating unit 31. A common-mode signal and an orthogonal signal are input from the quasi-synchronous detection unit 4 to the noise removal unit 35. Further, the noise removing unit 35 removes a noise component from the input signal by using the learning result stored in the noise removing learning parameter storage unit 5. The noise removing unit 35 inputs the signal after removing the noise component to each of the time series feature amount calculating unit 31 and the coupling unit 32. Since the time-series feature amount calculation unit 31, the coupling unit 32, and the signal specification identification unit 33 are the same as those in the first embodiment except that the signal after removing the noise component is processed, detailed description thereof is omitted here. To do.

FIG. 11 is a diagram showing a configuration example of the noise removing unit 35 shown in FIG. The noise removing unit 35 includes a normalization processing unit 351 and a noise removing device 352. The normalization processing unit 351 performs normalization processing on the input signal, and inputs the signal after the normalization processing to the noise remover 352. The noise remover 352 performs noise removal processing on the signal after the normalization processing using the learning result stored in the noise removal learning parameter storage unit 5, and outputs the signal after noise removal.

The configurations shown in FIGS. 9 to 11 are examples. Although the learning parameter storage unit 2 is provided outside the machine learning device 1b and the signal specification identification device 3b in FIG. 9, it may be provided inside the machine learning device 1b or the signal specifications. It may be provided inside the identification device 3b. Also, regarding the noise removal learning parameter storage unit 5, if the noise removal learning parameter storage unit 154 inside the machine learning device 1b can be read from the signal specification identification device 3b, the noise removal learning parameter storage unit 5 is omitted. You may. Further, the noise removal learning parameter storage unit 5 may be provided inside the signal specification identification device 3b. Further, in FIG. 9, although the machine learning device 1b and the signal specification identification device 3b are different devices, the function of the signal specification identification system 300 can be realized on one device. In this case, the time-series feature amount calculation unit 11 and the time-series feature amount calculation unit 31, the coupling unit 12 and the coupling unit 32, the signal specification learning unit 13 and the signal specification identification unit 33, the noise removal unit 15 and the noise removal unit 35. Can also be realized by the same hardware.

As described above, in the signal specification identification system 300 according to the third embodiment of the present invention, when a signal including a noise component is input by using the noise removal learning signal, the noise component is removed. Train the self-encoder so that the signal is output. Then, by calculating the time-series feature amount from the result of noise removal, it is possible to improve the identification accuracy of the signal specifications even for the received signal having a low signal-to-noise ratio.

Embodiment 4.
In the third embodiment, noise removal by non-linear processing is used by using a neural network, but when the number of labels to be identified by the system is large, or when the noise removal unit cannot sufficiently learn with a limited amount of data set. There is. The present embodiment is for removing noise without increasing the number of data sets even when the number of labels to be identified by the system is large.

FIG. 12 is a diagram showing a functional configuration of the signal specification identification system 400 according to the fourth embodiment of the present invention. The signal specification identification system 400 includes a machine learning device 1c, a learning parameter storage unit 2, a signal specification identification device 3c, a quasi-synchronous detection unit 4, and a noise removal learning parameter storage unit 5.

The machine learning device 1c includes a time-series feature amount calculation unit 11, a coupling unit 12, a signal specification learning unit 13, a signal selection unit 14, and a first noise removing unit 15-1 to Nth noise removing unit. It has 15-N and a selection unit 16. In the following description, when it is not necessary to distinguish each of the first noise removing units 15-1 to the Nth noise removing units 15-N, it is simply referred to as a noise removing unit 15. The machine learning device 1c is the same as the machine learning device 1b except that it has a plurality of noise removing units 15 and a selection unit 16 between the plurality of noise removing units 15 and the signal selection unit 14. Hereinafter, the parts different from the machine learning device 1b will be mainly described.

The selection unit 16 processes one or more noises from the plurality of noise removal units 15 based on the label included in the training data set, that is, based on the signal specifications indicated by the label. Select the removal unit 15. Each of the plurality of noise removing units 15 has, for example, the configuration shown in FIG. When a learning signal selected as a noise removal learning signal by the signal selection unit 14 is input to each of the plurality of noise removal units 15, the signal after removing the noise component is learned and the learning result is noise-removed. It is stored in the learning parameter storage unit 5. When a learning signal that has not been selected as a noise removal learning signal is input, each of the plurality of noise removing units 15 performs noise component removing processing and outputs the signal after noise removal.

The signal specification identification device 3c includes a time-series feature amount calculation unit 31, a coupling unit 32, a signal specification identification unit 33, and a first noise removal unit 35-1 to an Nth noise removal unit 35-N. , And a selection unit 36. Hereinafter, when it is not necessary to distinguish each of the first noise removing units 35-1 to the Nth noise removing units 35-N, it is simply referred to as a noise removing unit 35.

The selection unit 36 selects one or more noise removal units 35 for processing the received signal from the plurality of noise removal units 35 based on the identification result of the signal specifications output by the signal specification identification unit 33. .. The selection unit 36 inputs a reception signal to be processed to the selected noise removal unit 35. Each of the plurality of noise removing units 35 has, for example, the configuration shown in FIG. When a received signal is input, each of the plurality of noise removing units 35 removes a noise component from the received signal using the learning result stored in the noise removing learning parameter storage unit 5, and receives the received signal after noise removal. Is output to the time-series feature amount calculation unit 31.

The configuration shown in FIG. 12 is an example. Although the learning parameter storage unit 2 is provided outside the machine learning device 1c and the signal specification identification device 3c in FIG. 12, it may be provided inside the machine learning device 1c or the signal specifications. It may be provided inside the identification device 3c. As for the noise removal learning parameter storage unit 5, if the noise removal learning parameter storage unit 154 inside the machine learning device 1c can be read from the signal specification identification device 3c, the noise removal learning parameter storage unit 5 is omitted. You may. Further, the noise removal learning parameter storage unit 5 may be provided inside the signal specification identification device 3c. Further, in FIG. 12, although the machine learning device 1c and the signal specification identification device 3c are different devices, the function of the signal specification identification system 400 can be realized on one device. In this case, the time-series feature amount calculation unit 11 and the time-series feature amount calculation unit 31, the coupling unit 12 and the coupling unit 32, the signal specification learning unit 13 and the signal specification identification unit 33, and the first noise removing unit 15-1. The Nth noise removing unit 15-N and the first noise removing unit 35-1 to the Nth noise removing unit 35-N can be realized by the same hardware, respectively.

FIG. 13 is a flowchart showing the operation of the machine learning device 1c shown in FIG. In the description of FIG. 13, the same parts as those shown in FIG. 6 are designated by the same reference numerals, and detailed description thereof will be omitted.

The machine learning device 1c acquires learning data (step S101). The signal selection unit 14 determines whether or not the learning signal included in the learning data is a noise removal learning signal (step S301). In the case of a noise reduction learning signal (step S301: Yes), the selection unit 16 selects one or more from a plurality of noise removal units 15 according to the label included in the learning data, and selects the noise. The removing unit 15 is made to learn (step S302). After that, the process returns to the process of step S301.

When it is not a noise reduction learning signal (step S301: No), the selection unit 16 selects one or more from a plurality of noise removal units 15 according to the label included in the training data, and selects the noise removal unit. The unit 15 is made to remove noise (step S303). Hereinafter, since the processing of steps S102 to S104 is the same as that of FIG. 6, the description thereof will be omitted.

FIG. 14 is a flowchart showing the operation of the signal specification identification device 3c shown in FIG. The signal specification identification device 3c acquires an in-phase signal and an orthogonal signal included in the received signal from the quasi-synchronous detection unit 4 (step S401). The selection unit 36 selects one noise removal unit 35 from the plurality of noise removal units 35, and causes the selected noise removal unit 35 to remove noise (step S402). The time-series feature amount calculation unit 31 calculates the time-series feature amount from the received signal after noise removal (step S403). The coupling unit 32 combines the calculated time-series features with the in-phase signal and the orthogonal signal included in the received signal after noise removal to generate input data (step S404). The signal specification identification unit 33 identifies the signal specifications of the input signal from the input data by using the learning result stored in the learning parameter storage unit 2 (step S405). The signal specification identification unit 33 outputs the identification result to the outside of the signal specification identification device 3c and feeds it back to the selection unit 36.

The selection unit 36 selects the noise removal unit 35 that processes the received signal from the plurality of noise removal units 35 according to the identification result, and causes the selected noise removal unit 35 to remove the noise component of the reception signal (step). S406). The received signal after removing the noise component is input to the time-series feature amount calculation unit 31. The time-series feature amount calculation unit 31 calculates the time-series feature amount from the input received signal (step S407). The coupling unit 32 combines the calculated time-series features with the in-phase signal and the orthogonal signal included in the received signal after noise removal to generate input data (step S408). The signal specification identification unit 33 identifies the signal specifications of the received signal from the input data by using the learning result stored in the learning parameter storage unit 2 (step S409). The signal specification identification unit 33 outputs the identification result.

When the modulation method is identified by using the signal specification identification system 400, for example, when the label values are AM, FM, BPSK (Binary Phase Shift Keying), QPSK, and FSK. To do. In this case, the plurality of noise removing units 15, 35 are the first noise removing units 15-1, 35-1 for AM, the second noise removing units 15-2, 35-2 for FM, BPSK and QPSK. It is conceivable to provide four types of noise removing units 15, 35, which are the third noise removing units 15-3, 35-3 for FSK and the fourth noise removing units 15-4, 35-4 for FSK. Further, the fifth noise component is removed from the received signal by using the learning results of the fifth noise removing unit 15-5 and the fifth noise removing unit 15-5, which learn about all the labels regardless of the label value. When the noise removing unit 35-5 of the fifth is provided, in step S402 of FIG. 14, the noise component is removed by using the fifth noise removing unit 35-5, and in the subsequent step S406, selection is made based on the identification result. The noise component can be removed by using the noise removing unit 35.

As described above, according to the signal specification identification system 400 according to the fourth embodiment of the present invention, the number of data sets required for learning of the noise removing unit 15 is increased when the signal specifications are identified. It is possible to improve the identification accuracy by removing noise without any noise.

The configuration shown in the above-described embodiment shows an example of the content of the present invention, can be combined with another known technique, and is one of the configurations without departing from the gist of the present invention. It is also possible to omit or change the part.

Further, in the operation described using the flowchart, the operation of each step can be processed in parallel or the order of processing can be changed as needed.

1,1a, 1b, 1c Machine learning device, 2 Learning parameter storage unit, 3,3a, 3b, 3c Signal specification identification device, 4 Semi-synchronous detection unit, 5,154 Noise removal learning parameter storage unit, 11,31 o'clock Series feature amount calculation unit, 12,32 coupling unit, 13,13a signal specification learning unit, 14 signal selection unit, 15,35 noise removal unit, 15-1,35-1 first noise removal unit, 15-2 , 35-2 2nd noise removing unit, 15-3, 35-3 3rd noise removing unit, 15-4, 35-4 4th noise removing unit, 15-5, 35-5 5th noise Removal unit, 15-N, 35-N Nth noise removal unit, 16,36 selection unit, 33,33a signal specification identification unit, 34 supplementary information calculation unit, 41 oscillator, 42 π / 2 phase shifter, 43 , 44 Multiplier, 45,46 LPF, 90 processing circuit, 91 quasi-synchronous detection circuit, 92 processor, 93 memory, 94 storage, 100,200,300,400 signal specification identification system, 111 amplitude calculator, 112 phase calculation Unit, 151 noise addition unit for learning, 152,155,351 normalization processing unit, 153 noise remover learning unit, 156,352 noise remover.

Timothy J. O'Shea, Johnathan Corgan, and T. Charles Clancy “Convolutional Radio Modulation Recognition Networks”, CoRR, vol.abs / 1602, no.04105, pp.1-15, Jun.2016.

Claims (13)

A time-series feature amount calculation unit that calculates a time-series feature amount that is a time-series variation value of the feature amount of the learning signal from the in-phase signal and the orthogonal signal included in the learning signal.
A signal specification learning unit that learns signal specifications according to a training data set created based on the waveforms of the in-phase signal and the orthogonal signal and the time-series feature amount.
A machine learning device characterized by being equipped with. The learning data set in which the learning signal is associated with the signal specifications further includes supplementary information that can be calculated based on the learning signal.
The signal specification learning unit learns the signal specifications according to a training data set created based on the respective waveforms of the in-phase signal and the orthogonal signal, the time-series feature amount, and the supplementary information. The machine learning device according to claim 1, wherein the machine learning apparatus is to be used. A noise removing unit that learns a noise removing signal obtained by removing a noise component from the learning signal and removes the noise component from the learning signal by using the learning result of the noise removing signal.
With more
The machine learning device according to claim 1 or 2, wherein the time-series feature amount calculation unit calculates the time-series feature amount using the learning signal from which the noise component has been removed. A signal selection unit that selects a noise reduction learning signal used for learning the noise reduction signal from the input learning signals.
With more
The noise removing unit adds a noise component to the learning signal selected as the noise removing learning signal, and according to the noise removing data set based on the noise removing learning signal to which the noise component is added, the noise is said. The machine learning apparatus according to claim 3, wherein the noise removal learning signal before adding a component is learned as a signal after removing the noise component. With the plurality of the noise removing units,
A selection unit that selects one or more noise reduction units that process the learning signal from the plurality of noise reduction units based on the signal specifications of the learning signal.
The machine learning device according to claim 4, wherein the machine learning device is provided. The plurality of the noise removing units include a total learning noise removing unit that performs learning using all the noise removing learning signals regardless of the signal specifications.
The machine learning device according to claim 5, wherein the selection unit selects a total learning noise removing unit and a noise removing unit selected based on the signal specifications of the learning signal. A time-series feature amount calculation unit that calculates a time-series feature amount that is a time-series variation value of the feature amount of the received signal from the in-phase signal and the orthogonal signal included in the received signal.
Waveforms of the in-phase signal and the orthogonal signal, respectively, based on the specification learning result of learning the relationship between the waveforms of the in-phase signal and the orthogonal signal, the time-series feature amount, and the signal specifications. And the signal specification identification unit that identifies the signal specifications using the time-series feature amount,
A signal specification identification device comprising. Supplementary information calculation unit that calculates supplementary information indicating the characteristics of the received signal from the received signal,
With more
The signal specification identification unit is based on a learning result of learning the relationship between the waveforms of the in-phase signal and the orthogonal signal, the time-series feature amount, the supplementary information, and the signal specifications. The signal specification identification device according to claim 7, wherein the signal specifications are further used to identify the signal specifications. A noise removal unit that removes noise components from in-phase signals and orthogonal signals included in the received signal based on the noise learning result of learning the relationship between the input signal and the signal after removing the noise component from the signal. ,
With more
The signal specification identification device according to claim 7 or 8, wherein the time-series feature amount calculation unit calculates the time-series feature amount from the signal after removing the noise component. With the plurality of the noise removing units,
A selection unit that selects the noise removal unit that processes the reception signal based on the signal specifications of the reception signal from the plurality of noise reduction units.
The signal specification identification device according to claim 9, further comprising. The plurality of noise removing units include a total learning noise removing unit that removes noise components using all learning results regardless of the values of the signal specifications.
The signal specification identification according to claim 10, wherein the signal specifications of the received signal from which the noise component has been removed are identified by using the total learning noise removing unit, and then the identification result is fed back to the selection unit. apparatus. A first time-series feature amount calculation unit that calculates a time-series feature amount, which is a time-series variation value of the feature amount of the learning signal, from an in-phase signal and an orthogonal signal included in the learning signal.
A signal specification learning unit that learns signal specifications according to a training data set created based on the waveforms of the in-phase signal and the orthogonal signal included in the learning signal and the time-series feature amount.
A second time-series feature amount calculation unit that calculates the time-series feature amount of the received signal from the in-phase signal and the orthogonal signal included in the received signal, and
The waveforms of the in-phase signal and the orthogonal signal included in the received signal, the time-series feature amount calculated by the second time-series feature amount calculation unit, and the learning result of the signal specification learning unit are used. Based on the signal specification identification unit that outputs the identification result of the signal specifications,
A signal specification identification device comprising. A step of calculating a time-series feature amount, which is a time-series variation value of the feature amount of the learning signal, from an in-phase signal and an orthogonal signal included in the learning signal.
It is characterized by including a step of learning signal specifications according to a training data set created based on the respective waveforms of the in-phase signal and the orthogonal signal included in the learning signal and the time-series feature amount. Machine learning method.

Images