signal processing device

Abstract

(summary) (Problem) When an interference signal received simultaneously by N antennas in space contains (N+1) or more transmission signals with unknown transmission paths, it is not possible to determine the multiple transmission signals. (Solution) A signal processing device 100 determines multiple transmission signals from interference signals received simultaneously by N antennas with aligned directivity in outer space, where multiple phase-shift keyed transmission signals with overlapping bands on the same frequency channel that are not frequency-spread are received. The signal processing device 100 is equipped with a sampling unit 10 that samples the interference signals, a division unit 20 that divides the sampled interference signals into time windows of the symbol interval width, and a recurrent neural network with a many-to-many relationship between the number of input time steps and the number of output time steps. The signal processing device 100 receives as input a time window unit of the symbol interval width obtained by the division and outputs an extracted signal in time window unit of the symbol interval width, and a final output unit 40 that determines multiple transmission signals by arranging the extracted signals output from the calculation unit 30 in chronological order. The signal in time window unit of the symbol interval width is input to the calculation unit 30 for each time step of the recurrent neural network. (Selection diagram) Figure 4

Description

[Technical Field]

[0001] The present invention relates to a signal processing device that determines multiple transmission signals contained in an interference signal, from the interference signal received simultaneously by N (N is a natural number) antennas with aligned directionality in space, in which multiple phase shift keying transmission signals with overlapping bands are received on the same frequency channel without frequency spreading. [Background technology]

[0002] Patent Document 1 discloses a satellite array antenna device required for acquiring AIS information.

[0003] Non-Patent Document 1 describes the invention of Patent Document 1. That is, according to Non-Patent Document 1, the multi-antenna according to the invention of Patent Document 1 uses eight multi-antennas to ensure an antenna aperture length of about 5 m, and the sampled data from each antenna is downlinked to the terrestrial system and transmitted to the ground independently without being subjected to beam combining within the satellite. Also, terrestrial digital beamforming is said to perform multi-beam processing to achieve both a spatial diversity effect and wide-area observation. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] International Publication No. 2016 / 084975 [Non-Patent Document 1] IEICE Technical Report, Study on Signal Separation System for Satellite-based Automatic Identification System Experiment 3, SANE2019-117(2020-02) Summary of the Invention [Problem to be solved by the invention]

[0005] However, the signal processing device described in Patent Document 1 forms a desired antenna beam digitally by digital beamforming and separates signals according to the formed antenna beam. Therefore, although it can receive a signal in which N transmission signals are mixed with each other using N antennas and separate it into N transmission signals, it cannot simultaneously receive a signal in which (N+1) or more transmission signals are mixed with each other using N antennas and separate it into the transmission signals before the interference.

[0006] Furthermore, Non-Patent Document 1 considers a neural network method for separating and decoding interfering signals, but does not disclose a specific configuration for simultaneously receiving a signal interfering with (N+1) or more transmission signals using N antennas and separating and decoding the pre-interference transmission signals.

[0007] The present invention aims to provide a signal processing device that can determine multiple transmission signals contained in an interference signal even when the interference signal is received simultaneously by N antennas with aligned directivity and contains (N+1) or more transmission signals. [Means for solving the problem]

[0008] In one embodiment, a signal processing device is a signal processing device that determines a plurality of transmission signals from interference signals received simultaneously by N (N is a natural number) antennas with aligned directivity in space, the interference signals being phase shift keyed transmission signals with overlapping bands on the same frequency channel that are not frequency spread, the signal processing device comprising: a sampling unit that samples the interference signals; a division unit that divides the interference signals sampled by the sampling unit into time windows each having a symbol interval width; a calculation unit that is configured by a recurrent neural network in which the relationship between the number of input time steps and the number of output time steps is many-to-many, and that receives as input signals in units of time windows of the symbol interval width obtained by division by the division unit, and outputs extracted signals in units of time windows of the symbol interval width; and a final output unit that determines the plurality of transmission signals by chronologically arranging the extracted signals in units of time windows of the symbol interval width output from the calculation unit, wherein the signals in units of time windows of the symbol interval width are input to the calculation unit for each time step of the recurrent neural network.

[0009] The number of the plurality of transmission signals included in the interference signal may be (N+1) or more, and the recurrent neural network may have been trained in advance to extract the plurality of transmission signals before interference from an interference signal of (N+1) or more transmission signals.

[0010] The calculation unit may be configured as a mixed expert model including a plurality of recurrent neural networks that have been trained in advance to obtain a plurality of pre-interference transmission signals from an interference signal that is a combination of the plurality of transmission signals.

[0011] In one embodiment, a communication system includes any one of the signal processing devices described above and N spacecraft flying in formation, each spacecraft carrying a different one of the N antennas, and the distance between the two spacecraft farthest from each other among the N spacecraft is shorter than the distance calculated by multiplying the symbol interval of the baseband signal by the speed of light in a vacuum. [Effects of the Invention]

[0012] According to the signal processing device of the present invention, even when an interference signal containing (N+1) or more transmission signals is received simultaneously in space by N antennas with uniform directivity, the interference signal can be determined to include multiple transmission signals that are phase shift keyed and have overlapping bands on the same frequency channel that are not frequency spread. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram schematically illustrating the relationship between N antennas with aligned directivities and (N+1) signal sources. [Figure 2] 2 is a diagram schematically illustrating the configuration of a first receiver provided on a first artificial satellite and a second receiver provided on a second artificial satellite in the signal processing device according to the first embodiment. FIG. [Figure 3] FIG. 1 is a diagram showing a schematic diagram of the structure of a first interference signal simultaneously received by a first antenna and a second interference signal simultaneously received by a second antenna, and a diagram showing a schematic diagram of the configuration of a calculation unit to which the interference signals are input. [Figure 4] 1 is a block diagram schematically illustrating a configuration of a signal processing device according to a first embodiment. [Figure 5] FIG. 2 is a diagram schematically illustrating a configuration of a final output unit. [Figure 6] FIG. 1 is a diagram schematically illustrating three signals with different signal intensities received at the same time. [Figure 7] This is a diagram to explain a situation in which, after receiving a signal of signal strength C, a signal of signal strength A and a signal of signal strength B are received simultaneously, and a signal of signal strength D is received following the signal of signal strength C. [Figure 8] FIG. 10 is a diagram schematically illustrating the configuration of a first receiver provided on a first artificial satellite and a second receiver provided on a second artificial satellite in a signal processing device according to a second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments of the present invention will be described with reference to the drawings. However, the present invention should not be construed as being limited to the following embodiments.

[0015] The signal processing device of the present invention is configured to be capable of determining multiple transmission signals from an interference signal obtained by simultaneously receiving multiple phase shift keying transmission signals with overlapping bands on the same frequency channel that are not frequency spread using N (N is a natural number) antennas with aligned directivity in space. As an example, the number of multiple transmission signals included in the interference signal is (N+1) or more, and as a specific example, N is 2 and the number of multiple transmission signals included in the interference signal is 3.

[0016] The signal processing device of the present invention can be used, for example, to process Automatic Identification System (AIS) signals (hereinafter referred to as AIS signals). For example, the signal processing device of the present invention can extract multiple AIS signals from a mixed signal simultaneously received by an antenna mounted on a satellite. In the case of AIS, a low-cost solution that does not require major changes to the widely used existing transmission side configuration is desirable. AIS does not have channel state information (CSI) for understanding the state of the transmission path, and since there are countless candidate signal sources, it is not possible to use multiple input multiple output (MIMO).

[0017] First Embodiment Fig. 1 is a diagram showing a schematic diagram of the relationship between N antennas with aligned directivity and (N+1) signal sources. Here, the explanation is given assuming that N is 2, but N is not limited to 2. Furthermore, the number of signal sources that transmit transmission signals is not limited to (N+1), and may be a number less than N or a number greater than (N+2).

[0018] In this embodiment, an example in which antennas are mounted on artificial satellites will be described. Referring to Fig. 1, a first antenna 2A is mounted on a first artificial satellite 1A, and a second antenna 2B is mounted on a second artificial satellite 1B. The first artificial satellite 1A and the second artificial satellite 1B fly in formation so that their traveling directions and speeds are substantially the same.

[0019] The directivities of the first antenna 2A and the second antenna 2B are preferably aligned by controlling the attitude of the first satellite 1A and the second satellite 1B. Furthermore, the first antenna 2A and the second antenna 2B are preferably positioned within a range shorter than the distance between the symbol interval of the baseband signal and the speed of light in a vacuum, but longer than the wavelength of the carrier wave. By positioning the first antenna 2A and the second antenna 2B within this range, the transmission delay of the baseband signal can be minimized, and the signal strengths received by the two antennas 2A and 2B from the same signal source can be made approximately equal. For example, in the case of an AIS signal, the transmission rate is 9600 baud. The distance between the first satellite 1A and the second satellite 1B can be maintained so that the first antenna 2A and the second antenna 2B are positioned within a range of several hundred meters, which is shorter than the distance of 31 km between the symbol interval of the baseband signal and the speed of light in a vacuum. It should be noted that due to the phase difference between the signals received by the first antenna 2A and the second antenna 2B, it is not necessary to control the distance between the first satellite 1A and the second satellite 1B with strict accuracy.

[0020] 2 is a diagram schematically illustrating the configuration of a first receiver 6A provided on a first satellite 1A and a second receiver 6B provided on a second satellite 1B. The first receiver 6A includes a first antenna 2A, a first amplifier 3A, a first downconverter 4A, and a first sampling circuit 5A. The second receiver 6B includes a second antenna 2B, a second amplifier 3B, a second downconverter 4B, and a second sampling circuit 5B. It is preferable that the first receiver 6A and the second receiver 6B have equivalent configurations.

[0021] When signals transmitted from multiple signal sources are received on the same frequency channel at the same time, the first antenna 2A receives an interference signal resulting from the overlap of multiple signals. The same is true for the second antenna 2B. Currently, there are four frequency channels for AIS signals, including two frequency channels for satellite observation. In addition, there are I and Q channels that represent orthogonal carrier wave components, but the I and Q channels are collectively referred to as a single channel.

[0022] The first amplifier 3A amplifies the signal received by the first antenna 2A. Similarly, the second amplifier 3B amplifies the signal received by the second antenna 2B.

[0023] The first downconverter 4A performs processes such as carrier removal and frequency conversion on the signal amplified by the first amplifier 3A. Similarly, the second downconverter 4B performs processes such as carrier removal and frequency conversion on the signal amplified by the second amplifier 3B.

[0024] The first sampling circuit 5A samples the signal processed by the first down-converter 4A. Similarly, the second sampling circuit 5B samples the signal processed by the second down-converter 4B.

[0025] For signal processing, the signals sampled by the first sampling circuit 5A and the second sampling circuit 5B may be stored as digital data for a certain period of time and then periodically transmitted to Earth. The transmission speed of one channel of the AIS signal is 9600 baud, with 9600 symbols per second. Because 9600 Hz is a frequency similar to the audible range of audio frequencies, digital sampling is possible, just like with voice. This is similar to the way audio data is exchanged using wireless communication standards such as Bluetooth (registered trademark). Synchronization of sampling points on different satellites can be achieved by, for example, correlating the received power of sampled data. Time synchronization can also be achieved using GNSS signals, such as GPS, transmitted from satellites positioned higher up.

[0026] Even if the oscillators in the first downconverter 4A and the second downconverter 4B are of the same specification, there is a certain tolerance for industrial products, making it difficult to maintain perfect phase matching. However, since there is no need to synchronize the phase with any of the multiple signal sources 7A-7C, it is not necessary to synchronize the phase between the oscillators in the first downconverter 4A and the second downconverter 4B and the signal sources 7A-7C. Therefore, there is no practical problem even if the phases of the oscillators in the first downconverter 4A and the second downconverter 4B do not perfectly match.

[0027] Furthermore, when processing such as carrier removal is performed in the first downconverter 4A, a difference of exp(iθ) occurs in the initial phase between signals transmitted from the same signal source and received by the two antennas 2A and 2B due to minute differences in the propagation path lengths. In receivers that align the antenna directivity and do not share a carrier frequency oscillator, the oscillators in the downconverters are different between the receivers, and the phase difference between the carrier frequency oscillators is further included in θ.

[0028] 1, the first antenna 2A receives a first transmission signal 71A transmitted from a first signal source 7A, a second transmission signal 72A transmitted from a second signal source 7B, and a third transmission signal 73A transmitted from a third signal source 7C. The second antenna 2B receives a first transmission signal 71B transmitted from the first signal source 7A, a second transmission signal 72B transmitted from the second signal source 7B, and a third transmission signal 73B transmitted from the third signal source 7C. For example, the first signal source 7A, the second signal source 7B, and the third signal source 7C are marine ship radio stations, and the first transmission signals 71A and 71B, the second transmission signals 72A and 72B, and the third transmission signals 73A and 73B are AIS signals.

[0029] The pair of first transmission signals 71A and 71B, the pair of second transmission signals 72A and 72B, and the pair of third transmission signals 73A and 73B are all identical signals transmitted from the same signal source, but are assigned different symbols to distinguish between signals received by two different antennas 2A and 2B. After being downconverted by downconverters 4A and 4B, the pairs of transmission signals differ only in initial phase due to a slight difference in the propagation path length.

[0030] The first transmission signals 71A and 71B, the second transmission signals 72A and 72B, and the third transmission signals 73A and 73B are three phase-shift keying transmission signals that overlap in band on the same frequency channel without frequency spreading. GMSK (Gaussian Minimum Shift Keying), which is commonly used for AIS signals, is a type of phase-shift keying.

[0031] 4 is a block diagram schematically showing the configuration of a signal processing device 100 according to the first embodiment of the present invention. The signal processing device 100 includes a sampling unit 10, a division unit 20, a calculation unit 30, and a final output unit 40. For example, the sampling unit 10 can be mounted on an artificial satellite, and the division unit 20, the calculation unit 30, and the final output unit 40 can be installed in a data center or the like on Earth.

[0032] The sampling unit 10 samples interference signals received by antennas (in this embodiment, the first antenna 2A and the second antenna 2B). In this embodiment, the sampling unit 10 is a first sampling circuit 5A and a second sampling circuit 5B. However, the sampling unit 10 may also be configured by a first amplifier 3A, a second amplifier 3B, a first down-converter 4A, a second down-converter 4B, a first sampling circuit 5A, and a second sampling circuit 5B.

[0033] The dividing unit 20 divides the interference signal sampled by the sampling unit 10 into time windows each having a symbol interval width.

[0034] FIG. 3 schematically illustrates the structures of a first interference signal 41 received by the first antenna 2A and a second interference signal 42 received by the second antenna 2B. The first interference signal 41 is a signal resulting from interference between a first transmission signal 71A transmitted from a first signal source 7A, a second transmission signal 72A transmitted from a second signal source 7B, and a third transmission signal 73A transmitted from a third signal source 7C, and is received by the first antenna 2A. The second interference signal 42 is a signal resulting from interference between a first transmission signal 71B transmitted from the first signal source 7A, a second transmission signal 72B transmitted from the second signal source 7B, and a third transmission signal 73B transmitted from the third signal source 7C, and is received by the second antenna 2B. FIG. 3 also schematically illustrates a configuration in which interference signals are input to a calculation unit. Here, the description will be given assuming that the interference signal has a length of three symbols in the time direction. However, in practice, the length of the interference signal in the time direction is not limited to three symbols.

[0035] In this embodiment, the dividing unit 20 divides the first interference signal 41 sampled by the sampling unit 10 into time windows each having a symbol interval width, thereby obtaining three input signals: a first input signal 411, a second input signal 412, and a third input signal 413. Similarly, the dividing unit 20 divides the sampled second interference signal 42 into time windows each having a symbol interval width, thereby obtaining three input signals: a first input signal 421, a second input signal 422, and a third input signal 423. When the sampling unit 10 performs sampling at eight points per symbol, the first input signals 411 and 421, the second input signals 412 and 422, and the third input signals 413 and 423 are each a sequence of complex numbers having a length of one symbol, i.e., eight points.

[0036] The calculation unit 30 is configured with a recurrent neural network, and receives as input a signal in units of a time window of the symbol interval width obtained by division by the division unit 20, and outputs an extracted signal in units of a time window of the symbol interval width. The calculation unit 30 receives as input a signal in units of a time window of the symbol interval width for each time step of the recurrent neural network.

[0037] The calculation unit 30 has intermediate calculation units whose number corresponds to the number of symbols of the interference signal. In this embodiment, since the received signal has three symbols, the calculation unit 30 has three intermediate calculation units: a first intermediate calculation unit 30A, a second intermediate calculation unit 30B, and a third intermediate calculation unit 30C. This expanded state is a recurrent neural network in which the relationship between the number of input time steps and the number of output time steps is many-to-many. In the case of Figure 4, the relationship between the number of input time steps and the number of output time steps is three-to-three.

[0038] In this embodiment, the calculation unit 30 is configured as a so-called Mixture of Experts model of machine learning. Therefore, as shown in FIG. 3 , the calculation unit 30 includes multiple recurrent neural networks 31, 32, 33, and 34 that are pre-trained to obtain multiple pre-interference transmission signals from an interference signal that is a combination of multiple transmission signals. In this embodiment, the first intermediate calculation unit 30A is described as being configured with four recurrent neural networks 31A, 32A, 33A, and 34A. Similarly, the second intermediate calculation unit 30B is configured with four recurrent neural networks 31B, 32B, 33B, and 34B, and the third intermediate calculation unit 30C is configured with four recurrent neural networks 31C, 32C, 33C, and 34C. The first, second, and third intermediate calculation units are diagrammatically illustrated as they are developed in chronological order. However, the number of recurrent neural networks that make up the calculation unit 30 is not limited to four, and in practice, the number is expected to be in the range of several tens to several hundreds.

[0039] Each recurrent neural network 31-34 is internally composed of multiple recurrent neural network cells, and the structures of these networks do not necessarily need to be completely identical. For example, different tasks may be assigned to each of the recurrent neural networks 31-34, and the size of the recurrent neural network may be adjusted depending on the difficulty of the task.

[0040] 3 and 4 also show a recurrent loop 36 of the calculation unit 30. This recurrent loop 36 collectively represents the recurrent loops of the four recurrent neural networks 31, 32, 33, and 34. The second intermediate calculation unit 30B receives internal storage from the first intermediate calculation unit 30A via a recurrent loop 36A. The third intermediate calculation unit 30C receives internal storage from the second intermediate calculation unit 30B via a recurrent loop 36B.

[0041] Each of the recurrent neural networks 31, 32, 33, and 34 in the calculation unit 30 functions as an expert and independently processes the same input. For example, each of the recurrent neural networks 31, 32, 33, and 34 is trained in advance to extract the original transmission signal from a mixed signal of multiple transmission signals with different signal strengths, and then integrated into a mixed expert model. Each of the intermediate calculation units 30A, 30B, and 30C is obtained by expanding the calculation unit 30 three times along the time steps of the recurrent neural network. Each of the recurrent neural networks 31A, 32A, 33A, and 34A in the first intermediate calculation unit 30A performs processing independently, and the same is true for each of the recurrent neural networks in the second and third intermediate calculation units.

[0042] Each of the recurrent neural networks 31 to 34 has been trained in advance according to the number of transmission signals included in the interference signal. That is, when the number of transmission signals included in the interference signal is (N+1) or more, each of the recurrent neural networks 31 to 34 has been trained in advance to extract multiple transmission signals before interference from the interference signal of (N+1) or more transmission signals. In this embodiment, each of the recurrent neural networks 31 to 34 has been trained in advance to extract multiple transmission signals before interference from the interference signal of three transmission signals.

[0043] In the case of AIS signals, the transmission power of the signal source is predetermined, and the satellite orbit does not change significantly, so the received strength of the signal source is limited within a certain range, making it possible to classify the signals by the received strength of the signal source. Classification is performed in advance using a histogram with frequency on the vertical axis and received strength on the horizontal axis. The expected strength when the signal transmitted from each signal source is received with equal signal strength by the first antenna 2A and the second antenna 2B is classified into classes, and interference signals are classified according to the combination of each strength class. An expert model corresponding to each combination of the classified cases is trained. In this case, the binary of the training data signals given to the expert model can be a random pattern.

[0044] As an example, an interference signal 41 received by the first antenna 2A and an interference signal 42 received by the second antenna 2B are input as training data, and three signals 71A, 72A, and 73A contained in interference signal 41 and three signals 71B, 72B, and 73B contained in interference signal 42 are input as true output values, and fitting is performed using the backpropagation method. Alternatively, fitting is performed using the backpropagation method using a training data set in which square waves representing the binary values ​​contained in the interference signals are used as true output values. In this case, the square waves 71A and 71B are completely identical, the square waves 72A and 72B are completely identical, and the square waves 73A and 73B are completely identical. Learning using square waves representing binary values ​​as true output values ​​may contribute to reducing the size of the recurrent neural network. For the backpropagation method, backpropagation through time, a common technique used in training recurrent neural networks, is used. Suitable parameters can be obtained using Back Propagation Through Time and variations of stochastic gradient descent such as RMSprop or Adam.

[0045] The first intermediate arithmetic unit 30A, the second intermediate arithmetic unit 30B, and the third intermediate arithmetic unit 30C shown in FIG. 4 represent three expansions of a recurrent neural network operating in the past, present, and future directions. As described above, a signal in a time window unit of the symbol interval width is input to the arithmetic unit 30 for each time step of the recurrent neural network. Specifically, first input signals 411 and 421 are input to the first intermediate arithmetic unit 30A of the arithmetic unit 30, second input signals 412 and 422 are input to the second intermediate arithmetic unit 30B, and third input signals 413 and 423 are input to the third intermediate arithmetic unit 30C.

[0046] Recurrent neural network 31A of first intermediate arithmetic unit 30A performs a calculation based on input first input signals 411 and 421 and outputs first signal 211. Similarly, recurrent neural network 32A of first intermediate arithmetic unit 30A performs a calculation based on input first input signals 411 and 421 and outputs first signal 212, recurrent neural network 33A performs a calculation based on input first input signals 411 and 421 and outputs first signal 213, and recurrent neural network 34A performs a calculation based on input first input signals 411 and 421 and outputs first signal 214.

[0047] 5 is a diagram schematically illustrating the configuration of final output unit 40 (described later). First signal 211 output from recurrent neural network 31A includes three signals 211a, 211b, and 211c obtained by computing first input signal 411, and three signals 211d, 211e, and 211f obtained by computing first input signal 421. First signal 212 output from recurrent neural network 32A includes three signals 212a, 212b, and 212c obtained by computing first input signal 411, and three signals 212d, 212e, and 212f obtained by computing first input signal 421. The first signal 213 output from the recurrent neural network 33A includes three signals 213a, 213b, and 213c obtained by computing the first input signal 411, and three signals 213d, 213e, and 213f obtained by computing the first input signal 421. The first signal 214 output from the recurrent neural network 34A includes three signals 214a, 214b, and 214c obtained by computing the first input signal 411, and three signals 214d, 214e, and 214f ​​obtained by computing the first input signal 421.

[0048] As shown in FIG. 4, the first intermediate calculation unit 30A outputs the first internal memory to the second intermediate calculation unit 30B through a recurrent loop 36A.

[0049] The recurrent neural network 31B of the second intermediate arithmetic unit 30B performs calculations based on the second input signals 412, 422 input thereto and the first internal memory 36A output from the first intermediate arithmetic unit 30A, and outputs a second signal 221. Similarly, the recurrent neural network 32B of the second intermediate arithmetic unit 30B performs calculations based on the second input signals 412, 422 input thereto and the first internal memory 36A output from the first intermediate arithmetic unit 30A, and outputs a second signal 222; the recurrent neural network 33B performs calculations based on the second input signals 412, 422 input thereto and the first internal memory 36A output from the first intermediate arithmetic unit 30A, and outputs a second signal 223; and the recurrent neural network 34B performs calculations based on the second input signals 412, 422 input thereto and the first internal memory 36A output from the first intermediate arithmetic unit 30A, and outputs a second signal 224.

[0050] 5, second signal 221 output from recurrent neural network 31B includes three signals 221a, 221b, and 221c obtained by computing second input signal 412, and three signals 221d, 221e, and 221f obtained by computing second input signal 422. Second signal 222 output from recurrent neural network 32B includes three signals 222a, 222b, and 222c obtained by computing second input signal 412, and three signals 222d, 222e, and 222f obtained by computing second input signal 422. Second signal 223 output from recurrent neural network 33B includes three signals 223a, 223b, and 223c obtained by computing second input signal 412, and three signals 223d, 223e, and 223f obtained by computing second input signal 422. The second signal 224 output from the recurrent neural network 34B includes three signals 224a, 224b, and 224c that are calculated from the second input signal 412, and three signals 224d, 224e, and 224f that are calculated from the second input signal 422.

[0051] Also, as shown in FIG. 4, the second intermediate arithmetic unit 30B outputs the second internal memory to the third intermediate arithmetic unit 30C through a recurrent loop 36B.

[0052] The recurrent neural network 31C of the third intermediate arithmetic unit 30C performs calculations based on the third input signals 413, 423 input thereto and the second internal memory 36B output from the second intermediate arithmetic unit 30B, and outputs a third signal 231. Similarly, the recurrent neural network 32C of the third intermediate arithmetic unit 30C performs calculations based on the third input signals 413, 423 input thereto and the second internal memory 36B output from the second intermediate arithmetic unit 30B, and outputs a third signal 232; the recurrent neural network 33C performs calculations based on the third input signals 413, 423 input thereto and the second internal memory 36B output from the second intermediate arithmetic unit 30B, and outputs a third signal 233; and the recurrent neural network 34C performs calculations based on the third input signals 413, 423 input thereto and the second internal memory 36B output from the second intermediate arithmetic unit 30B, and outputs a third signal 234.

[0053] 5, the third signal 231 output from the recurrent neural network 31C includes three signals 231a, 231b, and 231c obtained by computing the third input signal 413, and three signals 231d, 231e, and 231f obtained by computing the third input signal 423. The third signal 232 output from the recurrent neural network 32C includes three signals 232a, 232b, and 232c obtained by computing the third input signal 413, and three signals 232d, 232e, and 232f obtained by computing the third input signal 423. The third signal 233 output from the recurrent neural network 33C includes three signals 233a, 233b, and 233c obtained by computing the third input signal 413, and three signals 233d, 233e, and 233f obtained by computing the third input signal 423. The third signal 234 output from the recurrent neural network 34C includes three signals 234a, 234b, and 234c that are calculated from the third input signal 413, and three signals 234d, 234e, and 234f that are calculated from the third input signal 423.

[0054] The calculation unit 30 can be implemented as a complex neural network. The complex-valued recurrent neural network may be a simple RNN, LSTM, GRU, or the like. In the case of a simple RNN, the input, output, and weight matrix parameters are converted from real numbers to complex numbers, and a complex activation function expressed as (z / |z|)sigmoid(|z|+b) or the like for complex number z is used. Here, b is the bias of the real number, (z / |z|) is the phase of complex number z, and |z|+b is the sum of the absolute value component of the complex number and the bias of the real number. Sigmoid is a general sigmoid function that performs nonlinear processing.

[0055] A complex neural network that uses only (z / |z|)sigmoid(|z|+b) as an activation function has the property of preserving phase components, and therefore its greatest feature is that it behaves identically to a given input and another input that is rotationally symmetric on the complex plane. It is desirable that the signal processing device 100 of the present invention can also utilize rotational symmetry on the complex plane.

[0056] Generally, an orthogonal matrix is ​​used as the initial value before learning for the weight matrix parameters that represent the recurrent loop part of a recurrent neural network, but a unitary matrix can be used as the initial value before learning for the complex weight matrix parameters that represent the recurrent loop part of a complex recurrent neural network.

[0057] As with recurrent neural networks, appropriate parameters can be acquired for the complex recurrent neural network of the signal processing device 100 using Back Propagation Through Time, a common algorithm used for learning. Appropriate parameters can be acquired by applying the back propagation method using partial differentiation of the real and imaginary parts, and using Back Propagation Through Time and a gradient method such as RMSprop or Adam.

[0058] The final output unit 40 determines multiple transmission signals included in the interference signal by chronologically arranging the signals output from the calculation unit 30 in time window units of the symbol interval width. In this embodiment, the final output unit 40 determines a first extracted signal 216, which is the final solution, based on first signals 211, 212, 213, and 214 output from the recurrent neural networks 31A, 32A, 33A, and 34A of the first intermediate calculation unit 30A. More specifically, the final output unit 40 references the first signals 211, 212, 213, and 214 and the signal patterns included in the output sequences of the recurrent neural networks 31, 32, 33, and 34, and selects one of the first signals 211, 212, 213, and 214 to be the first extracted signal 216. As shown in FIG. 5, the first extracted signal 216 includes three signals 216a, 216b, and 216c extracted from the first input signal 411, and three signals 216d, 216e, and 216f extracted from the first input signal 421.

[0059] Similarly, the final output unit 40 obtains a second extracted signal 226, which is the final solution, based on second signals 221, 222, 223, and 224 output from recurrent neural networks 31B, 32B, 33B, and 34B of the second intermediate arithmetic unit 30B. More specifically, the final output unit 40 references the second signals 221, 222, 223, and 224 and the signal patterns included in the output sequences of the recurrent neural networks 31, 32, 33, and 34, and selects one of the second signals 221, 222, 223, and 224 as the second extracted signal 226. As shown in FIG. 5 , the second extracted signal 226 includes three signals 226a, 226b, and 226c extracted from the second input signal 412 and three signals 226d, 226e, and 226f extracted from the second input signal 422.

[0060] Similarly, the final output unit 40 obtains a third extracted signal 236, which is the final solution, based on third signals 231, 232, 233, and 234 output from recurrent neural networks 31C, 32C, 33C, and 34C of the third intermediate arithmetic unit 30C. More specifically, the final output unit 40 references the third signals 231, 232, 233, and 234 and the signal patterns included in the output sequences of the recurrent neural networks 31, 32, 33, and 34, and selects one of the third signals 231, 232, 233, and 234 as the third extracted signal 236. As shown in FIG. 5 , the third extracted signal 236 includes three signals 236a, 236b, and 236c extracted from the third input signal 413 and three signals 236d, 236e, and 236f extracted from the third input signal 423.

[0061] As an example, as shown in Figure 6, a case will be described in which three signals, signal 401 with signal strength A, signal 402 with signal strength B, and signal 403 with signal strength C, are received at the same time. It is assumed that signal strength A, signal strength B, and signal strength C are all different signal strengths. Here, the processing of recurrent neural networks 31, 32, 33, and 34 of calculation unit 30 will be explained as an example. If one frame length is, for example, 256 symbols, the intermediate calculation unit of calculation unit 30 will expand 256 times to extract the three signals, signal 401, signal 402, and signal 403.

[0062] It is assumed that recurrent neural network 31 has been trained to be able to analyze an interference signal of three signals, signal 401 of signal strength A, signal 402 of signal strength B, and signal 403 of signal strength C; recurrent neural network 32 has been trained to be able to analyze an interference signal of three signals, signal 401 of signal strength A, signal 402 of signal strength B, and signal 403 of signal strength D; recurrent neural network 33 has been trained to be able to analyze an interference signal of three signals, signal 401 of signal strength A, signal 403 of signal strength C, and signal 403 of signal strength D; and recurrent neural network 34 has been trained to be able to analyze an interference signal of three signals, signal 402 of signal strength B, signal 403 of signal strength C, and signal 403 of signal strength D.

[0063] Each of the recurrent neural networks 31A, 32A, 33A, and 34A outputs a first signal. The first signal 211 with the highest accuracy is output from the recurrent neural network 31A, which has learned about the interference signals of three signals: signal 401 with signal strength A, signal 402 with signal strength B, and signal 403 with signal strength C. In this case, the final output unit 40 selects the first signal 211 with the highest accuracy output from the recurrent neural network 31A as the final solution, first extracted signal 216, from among the first signals 211, 212, 213, and 214 output from the recurrent neural networks 31A, 32A, 33A, and 34A. The same process is repeated for the second intermediate calculation unit 30B and the third intermediate calculation unit 30C. This process is repeated 256 times. Here, the selection criterion can be whether the 256 output sequences of the recurrent neural network 31 contain some of the signal preambles and signal postambles of the signals 401, 402, and 403.

[0064] The final output unit 40 obtains a plurality of pre-interference transmission signals 251A, 251B, 252A, 252B, 253A, and 253B by chronologically arranging the first extracted signal 216, which is the final solution, the second extracted signal 226, which is the final solution, and the third extracted signal 236, which is the final solution (see FIG. 5). The pre-interference transmission signals 251A and 251B obtained by the final output unit 40 correspond to the first transmission signals 71A and 71B transmitted from the first signal source 7A, the transmission signals 252A and 252B correspond to the second transmission signals 72A and 72B transmitted from the second signal source 7B, and the transmission signals 253A and 253B correspond to the third transmission signals 73A and 73B transmitted from the third signal source 7C.

[0065] Even if the combination of interfering signals changes during the reception of one frame length, for example, 256 symbols, it is possible to continuously capture signals of the same intensity class by switching the output of the recurrent neural network. This will be explained below.

[0066] As an example, a case will be described in which an antenna receives a signal as shown in Fig. 7. That is, after the antenna receives a signal 403 with signal strength C, it simultaneously receives a signal 401 with signal strength A and a signal 402 with signal strength B, and immediately after it stops receiving signal 403 with signal strength C, it receives a signal 404 with signal strength D.

[0067] Here again, the processing of recurrent neural networks 31, 32, 33, and 34 of calculation unit 30 will be described as an example. It is assumed that recurrent neural network 31 has been trained to be able to analyze an interference signal of three signals, signal 401 of signal strength A, signal 402 of signal strength B, and signal 403 of signal strength C, recurrent neural network 32 has been trained to be able to analyze an interference signal of three signals, signal 401 of signal strength A, signal 402 of signal strength B, and signal 404 of signal strength D, recurrent neural network 33 has been trained to be able to analyze an interference signal of three signals, signal 401 of signal strength A, signal 403 of signal strength C, and signal 404 of signal strength D, and recurrent neural network 34 has been trained to be able to analyze an interference signal of three signals, signal 402 of signal strength B, signal 403 of signal strength C, and signal 404 of signal strength D.

[0068] The interference signals received in interval T1 in Figure 7 include signal 401 with signal strength A, signal 402 with signal strength B, and signal 403 with signal strength C. The interference signals received in interval T2 include signal 401 with signal strength A, signal 402 with signal strength B, and signal 404 with signal strength D. In the example shown in Figure 7, signal strengths A, B, C, and D are all different signal strengths. If one frame length is, for example, 256 symbols, the intermediate calculation unit of calculation unit 30 performs expansion 256 times across intervals T1 and T2 to extract three signals. Final output unit 40 monitors the output sequences of each recurrent neural network 31-34 throughout intervals T1 and T2 and detects whether any of recurrent neural networks 31-34 has output a signal preamble.

[0069] When final output unit 40 detects that recurrent neural network 31 has output the preambles of signal 401 of signal strength A and signal 402 of signal strength B, it determines that one frame of signal 401 of signal strength A and signal 402 of signal strength B has started. Final output unit 40 identifies the end positions of signal 401 of signal strength A and signal 402 of signal strength B based on the specified length of one signal frame, for example, 256 symbols, and determines that one frame of signal 401 of signal strength A and signal 402 of signal strength B has ended.

[0070] One of the recurrent neural networks 31 to 34 correctly outputs the postambles of signal 401 of signal strength A and signal 402 of signal strength B at the end position. Final output unit 40 detects that recurrent neural network 31 has not correctly output the postambles of signal 401 of signal strength A and signal 402 of signal strength B, and that recurrent neural network 32 has correctly output the postambles of signal 401 of signal strength A and signal 402 of signal strength B. Since the combination of interfering signals has changed while receiving one frame length of signal 401 of signal strength A and signal 402 of signal strength B, it is determined that the output of the recurrent neural network needs to be switched. Next, it is necessary to identify the boundary between intervals T1 and T2 as the position where the output should be switched.

[0071] When the final output unit 40 determines that the interval T1 has started, it adopts the output sequence of the recurrent neural network 31 for the interval T1. When the final output unit 40 determines that the interval T2 has started, it adopts the output sequence of the recurrent neural network 32 for the interval T2.

[0072] The final output unit 40 identifies the positions of the rear end of the section T1 and the front end of the section T2, and switches the recurrent neural network that adopts the output at those points from the recurrent neural network 31 to the recurrent neural network 32.

[0073] The output sequence of recurrent neural network 31 includes the postamble of signal 403 with signal strength C, and therefore identifies the end of section T1. The output sequence of recurrent neural network 32 includes the preamble of the signal with signal strength D, and therefore identifies the start of section T2.

[0074] Although the final output unit 40 can theoretically be implemented using a neural network, reliability is ensured by creating a rule-based program that monitors the output as described above.

[0075] According to the signal processing device 100 of this embodiment, the sampling unit 10 samples an interference signal, which is a combination of multiple non-spread frequency phase-shift keyed transmission signals with overlapping bands on the same frequency channel, received simultaneously in space by N antennas with aligned directivity. The sampled interference signal is then divided into time windows each having a symbol interval width by a dividing unit 20, and input to a calculation unit 30 for each time step of a recurrent neural network. The calculation unit 30 is configured as a recurrent neural network and outputs an extracted signal for each time window having a symbol interval width. A final output unit 40 chronologically arranges the extracted signals output from the calculation unit 30 for each time window having a symbol interval width to determine the multiple transmission signals before interference. For example, by training the calculation unit 30, which is configured as a recurrent neural network, to extract the original transmission signal from an interference signal of (N+1) or more transmission signals, the multiple transmission signals contained in the interference signal can be determined even when the interference signal received simultaneously by N antennas contains (N+1) or more transmission signals.

[0076] Furthermore, conventional signal processing devices that use beamforming to separate transmission signals from interfering signals cannot separate multiple transmission signals transmitted from the same direction, but the signal processing device 100 of this embodiment can also extract multiple transmission signals transmitted from the same direction.

[0077] The number of multiple transmission signals included in the interference signal is (N+1) or more, and the recurrent neural networks 31 to 34 are configured to have been trained in advance to extract multiple transmission signals before interference from an interference signal of (N+1) or more transmission signals.This makes it possible to determine the (N+1) or more transmission signals included in the interference signal even if the interference signal received simultaneously by N antennas contains (N+1) or more transmission signals.

[0078] In the signal processing device 100 of this embodiment, the calculation unit 30 is configured as a mixed expert model including multiple recurrent neural networks that have been trained in advance to determine multiple pre-interference transmission signals from an interference signal that is a combination of multiple transmission signals. Therefore, compared to a case in which the calculation unit 30 is configured with only a single recurrent neural network, versatility is achieved while maintaining high accuracy, and multiple transmission signals included in the interference signal can be determined more accurately.

[0079] A communication system according to one embodiment of the present invention includes a signal processing device 100 and N satellites flying in formation, each equipped with a different one of N antennas. The distance between the two most distant satellites among the N satellites is shorter than the product of the symbol interval and the speed of light in a vacuum. Because the distance between the two most distant satellites among the N satellites is shorter than the product of the symbol interval and the speed of light in a vacuum, the transmission delay of the baseband signal can be reduced to a negligible level, and the strength of signals simultaneously received by the N antennas from the same signal source can be made approximately equal.

[0080] <Second embodiment> In the signal processing device 100 in the first embodiment, it has been described that the first antenna 2A is mounted on the first artificial satellite 1A and the second antenna 2B is mounted on the second artificial satellite 1B. In contrast, in the signal processing device 100 in the second embodiment, two antennas are mounted on one artificial satellite.

[0081] FIG. 8 is a diagram schematically showing the configuration of first receivers 6A1 and 6A2 provided on a first artificial satellite 1A and second receivers 6B1 and 6B2 provided on a second artificial satellite 1B in a signal processing device 100 in the second embodiment.

[0082] The first receiver 6A1 includes a first antenna 2A1, a first amplifier 3A1, a first downconverter 4A1, and a first sampling circuit 5A1, and the first receiver 6A2 includes a first antenna 2A2, a first amplifier 3A2, a first downconverter 4A2, and a first sampling circuit 5A2.

[0083] The second receiver 6B1 includes a second antenna 2B1, a second amplifier 3B1, a second downconverter 4B1, and a second sampling circuit 5B1, and the second receiver 6B2 includes a second antenna 2B2, a second amplifier 3B2, a second downconverter 4B2, and a second sampling circuit 5B2.

[0084] In this embodiment, the directivities of the first antenna 2A1 and the second antenna 2B1 are aligned, and the directivities of the first antenna 2A2 and the second antenna 2B2 are aligned. The directivities of the first antenna 2A1 and the first antenna 2A2 are different. The directivities of the second antenna 2B1 and the second antenna 2B2 are also different.

[0085] The first downconverters 4A1 and 4A2 share a single first carrier frequency oscillator 8A. The second downconverters 4B1 and 4B2 share a single second carrier frequency oscillator 8B. Even if there is no data processing purpose, the first carrier frequency oscillator 8A and the second carrier frequency oscillator 8B are shared to reduce power consumption by using a common receiver configuration. The first receiver 6A1 and the second receiver 6B1 have the same configuration, and the first receiver 6A2 and the second receiver 6B2 have the same configuration.

[0086] The first receiver 6A1 and the second receiver 6B1 function in the same manner as the first receiver 6A and the second receiver 6B of the signal processing device 100 in the first embodiment, respectively. The first receiver 6A2 and the second receiver 6B2 function in the same manner as the first receiver 6A and the second receiver 6B of the signal processing device 100 in the first embodiment, respectively. The first receiver 6A1 and the second receiver 6B1 receive signals with a different directivity from that of the first receiver 6A2 and the second receiver 6B2.

[0087] According to the signal processing device 100 of the second embodiment, two antennas are mounted on one satellite, and therefore, for example, one antenna can be oriented with directivity forward and backward in the direction of satellite flight, and another antenna with directivity to the left and right in the direction of satellite flight, and the antennas with different directivities can be arranged so that the directivities do not overlap as much as possible. In this case, the antenna with directivity forward and backward in the direction of satellite flight can allocate its reception area to signals that are significantly affected by Doppler shift, and the antenna with directivity to the left and right can allocate its reception area to signals that are less affected by Doppler shift, which is more advantageous than mounting a single antenna with a wide directivity.

[0088] It is also possible to mount three or more antennas on one satellite.

[0089] The present invention is not limited to the above-described embodiment, and various applications and modifications can be made within the scope of the present invention.

[0090] For example, it is possible to combine beamforming technology with the signal processing device of the present invention. Beamforming reception is a method of combining signals simultaneously received by multiple antennas on a multi-antenna receiver capable of beamforming reception into a signal received by a single virtual antenna, and theoretically, under the best conditions, it can reduce radio waves from unnecessary directions up to (total number of antennas - 1). Therefore, after eliminating radio waves from unnecessary directions by beamforming, the signal processing device of the present invention can determine multiple transmission signals from interfering signals.

[0091] Furthermore, if the sampling is 8 points per symbol, the amount of information does not deteriorate even if the interference signal is subjected to a discrete Fourier transform with the same 8-point period, and this can be applied to the present invention. Other methods that can be applied include a bidirectional recurrent neural network that inputs an interference signal that has been reverse-played in the time direction, and a method that delays the output of the extracted signal by one time step or one symbol.

[0092] Additionally, although the calculation unit 30 is generally implemented using a digital calculation processing device, each recurrent neural network that constitutes the calculation unit 30 can also be implemented using reservoir computing using an optical analog calculator, etc. In addition to the input from the division unit 20, the calculation unit 30 can also input auxiliary information for calculation. Although the signal processing device of the present invention does not fly in formation, by applying it to N antennas with uniform directivity on one of the satellites that make up a satellite constellation, it is possible to construct a receiving network using the satellite constellation. [Explanation of symbols]

[0093] 1A First Satellite 1B Second Satellite 2A, 2A1, 2A2 First antenna 2B, 2B1, 2B2 Second antenna 3A, 3A1, 3A2 First Amplifier 3B, 3B1, 3B2 Second Amplifier 4A, 4A1, 4A2 First Down Converter 4B, 4B1, 4B2 Second Down Converter 5A, 5A1, 5A2 First sampling circuit 5B, 5B1, 5B2 Second sampling circuit 6A, 6A1, 6A2 First receiver 6B, 6B1, 6B2 Secondary Receiver 7A First signal source 7B Second signal source 7C Third signal source 8A First Carrier Frequency Oscillator 8B Second Carrier Frequency Oscillator 10 Sampling Department 20 Division 30 Arithmetic section 30A First intermediate calculation unit 30B Second intermediate calculation unit 30C Third intermediate calculation unit 31, 32, 33, 34 Recurrent Neural Networks 31A, 32A, 33A, 34A Recurrent neural network for the first input signal 31B, 32B, 33B, 34B Recurrent neural network for the second input signal 31C, 32C, 33C, 34C Recurrent neural network for the third input signal 40 Final output section 100 Signal processing device

Claims

1. A signal processing device for determining multiple transmission signals from N interference signals simultaneously received by N (N is a natural number) antennas with aligned directivity, using a modulation scheme employing phase-shift modulation with overlapping bandwidths on the same frequency channel that is not spread across frequencies. A sampling unit that samples the aforementioned interference signal, A division unit divides the interference signal sampled by the sampling unit into time windows of symbol interval width, The recurrent neural network is configured such that the relationship between the number of input time steps and the number of output time steps is many to many, and the calculation unit takes as input a signal in time window units of the symbol interval width obtained by the division by the division unit and outputs an extracted signal in time window units of the symbol interval width. A final output unit that determines the multiple transmission signals by arranging the extracted signals in time window units of the symbol interval width output from the calculation unit in a time series, Equipped with, The number of the multiple transmission signals included in the interference signal is (N+1) or more. The signal in units of the time window of the symbol interval width is input to the calculation unit at each time step of the recurrent neural network. The signal processing device is characterized in that the recurrent neural network is pre-trained to extract the multiple transmission signals before interference from interference signals of (N+1) or more transmission signals.

2. The signal processing device according to claim 1, characterized in that the N interfering signals are obtained by beamforming reception.

3. The signal processing device according to claim 1, characterized in that the calculation unit is configured as a mixed expert model comprising multiple recurrent neural networks that have been pre-trained to determine the multiple transmission signals before interference from an interference signal which is a combination of multiple transmission signals.

4. A signal processing apparatus according to any one of claims 1 to 3, N spacecraft flying in formation, each equipped with one different antenna out of the N antennas with the aforementioned directional properties, Equipped with, A communication system characterized in that, among the N spacecraft, the distance between the two furthest spacecraft is shorter than the distance obtained by the product of the symbol interval time of the plurality of transmitted signals and the speed of light in a vacuum.

Images