Signal processing device and signal processing method

Abstract

To more easily reduce a signal error rate.SOLUTION: A signal processing device 1 comprises: a setting unit 10 configured to set a data string in each pattern of a transmission signal whose data string has a set length as a normal signal; a generator 121 configured to generate a pseudo signal which is a data string serving as a different pattern with respect to the pattern of a data string of the normal signal; a learning unit 12 configured to train a machine learning model including a discriminator 122 for distinguishing between the pseudo signal generated by the generator 121 and the normal signal; a determination unit 13 configured to determine a threshold value for suppressing the error rate between the normal signal and the pseudo signal in a transmission signal to which noise has been added on the basis of the parameters of the machine learning model optimized through training by the learning unit 12; and a notification unit 15 configured to notify a device which receives the transmission signal of the determined threshold value.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a signal processing device and a signal processing method, and more particularly to a technology for processing a transmission signal to which noise has been added. [Background technology]

[0002] Conventionally, there have been known techniques for correcting signal errors caused by noise that occurs during signal transmission. Patent Document 1 discloses a technique for reducing data errors in transmission characteristics by performing error correction coding on data on the transmitting side and performing error correction according to the coding on the receiving side. Also, known error correction methods include, for example, Hamming code and BCH code.

[0003] However, conventional error correction techniques that add redundant bits to data during transmission have the problem of increasing communication traffic due to the addition of redundant bits, and also have the problem of the heavy processing load required for error correction. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7241851 [Patent Document 2] Japanese Patent Application Laid-Open No. 2013-038676 Summary of the Invention [Problem to be solved by the invention]

[0005] With conventional techniques, it has been difficult to reduce the signal error rate more easily.

[0006] The present invention has been made to solve the above-mentioned problems, and has as its object to more easily reduce the signal error rate. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems, the signal processing device of the present invention comprises a setting unit configured to set each pattern of a data sequence of a transmission signal having a data sequence of a set length as a normal signal; a learning unit configured to learn a machine learning model having a generator that generates a pseudo signal, which is a data sequence having a pattern different from the pattern of the data sequence of the normal signal; and a classifier that distinguishes between the pseudo signal generated by the generator and the normal signal; a determination unit configured to determine a threshold value for suppressing the error rate between the normal signal and the pseudo signal in the transmission signal to which noise has been added, based on parameters of the machine learning model optimized by learning by the learning unit; and a notification unit configured to notify a device receiving the transmission signal of the determined threshold value.

[0008] In addition, in the signal processing device according to the present invention, the decision unit may set the threshold value based on a first expected value at which the classifier will identify the normal signal as a normal signal, and a second expected value at which the classifier will identify the pseudo signal generated by the generator as a pseudo signal, both indicated by an objective function of the machine learning model.

[0009] In addition, in the signal processing device according to the present invention, the learning unit may adjust the value of the first expected value until the generator generates the pseudo signals having data sequence patterns different from each other for each pattern of the data sequence of the normal signal set by the setting unit, thereby performing adversarial training between the generator and the classifier.

[0010] Furthermore, the signal processing device of the present invention may further include a collection unit configured to collect the variance of the noise from the device, and the determination unit may determine the threshold value based on the variance collected by the collection unit, the first expected value, and the second expected value.

[0011] In order to solve the above-mentioned problems, the signal processing method of the present invention includes a setting step of setting each pattern of a data sequence of a transmission signal having a data sequence of a set length as a normal signal; a learning step of training a machine learning model having a generator that generates a pseudo signal, which is a data sequence having a pattern different from the pattern of the data sequence of the normal signal, and a classifier that distinguishes between the pseudo signal generated by the generator and the normal signal; a determination step of determining a threshold value for suppressing the error rate between the normal signal and the pseudo signal in the transmission signal to which noise has been added, based on parameters of the machine learning model optimized by learning in the learning step; and a notification step of notifying a device receiving the transmission signal of the determined threshold value.

[0012] In addition, in the signal processing method according to the present invention, the determining step may set the threshold value based on a first expected value at which the classifier will identify the normal signal as a normal signal, and a second expected value at which the classifier will identify the pseudo signal generated by the generator as a pseudo signal, both indicated by an objective function of the machine learning model.

[0013] In addition, in the signal processing method according to the present invention, the learning step may adjust the value of the first expected value until the generator generates the pseudo signal having a data sequence pattern different from each other for each pattern of the data sequence of the normal signal set in the setting step, and perform adversarial learning between the generator and the classifier.

[0014] Furthermore, the signal processing method according to the present invention may further include a collection step of collecting the variance of the noise from the device, and the determination step may determine the threshold value based on the variance collected in the collection step, the first expected value, and the second expected value. [Effects of the Invention]

[0015] According to the present invention, a threshold value for suppressing the error rate between a normal signal and a pseudo signal in a transmission signal to which noise has been added is determined based on parameters of a machine learning model optimized by learning by a learning unit, thereby making it possible to more easily reduce the error rate of the signal. [Brief explanation of the drawings]

[0016] [Figure 1] FIG. 1 is a block diagram showing the configuration of a signal processing system including a signal processing device according to an embodiment of the present invention. [Figure 2] FIG. 2 is a block diagram showing the configuration of a learning unit included in the signal processing device according to this embodiment. [Figure 3] FIG. 3 is a diagram illustrating a learning unit included in the signal processing device according to the present embodiment. [Figure 4] FIG. 4 is a diagram illustrating a learning unit included in the signal processing device according to the present embodiment. [Figure 5] FIG. 5 is a block diagram showing the hardware configuration of a signal processing device according to this embodiment. [Figure 6] FIG. 6 is a sequence diagram showing an outline of the operation of the signal processing system according to this embodiment. [Figure 7] FIG. 7 is a flowchart showing the operation of the signal processing device according to this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0017] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will now be described in detail with reference to FIGS.

[0018] [Configuration of signal processing system] First, an outline of a signal processing system including a signal processing device 1 according to an embodiment of the present invention will be described. Fig. 1 is a block diagram showing the configuration of a signal processing system including a signal processing device 1 according to an embodiment of the present invention.

[0019] A signal processing system according to this embodiment is provided in a communication network having, for example, a communication path L in a wireless section and a communication path C in a wired section. As shown in Fig. 1, the signal processing system includes a signal processing device 1, a communication terminal 2, a base station 3, and a core device 4. The signal processing device 1, the base station 3, and the core device 4 are each connected to a network NW such as a LAN, a WAN, or the Internet. A communication path L in a wireless section is provided between the communication terminal 2 and the base station 3, and a communication path C in a wired section is provided between the base station 3 and the core device 4.

[0020] The communication terminal 2 is a user terminal such as a smartphone or tablet terminal, and performs wireless communication with a base station 3 in the communication area where the communication terminal 2 is located via a wireless communication path L. The communication terminal 2 communicates with a core device 4 via the base station 3.

[0021] The base station 3 is a base station conforming to a predetermined communication standard such as LTE, 5G, or 6G, and establishes communication with the communication terminal 2 and transmits data to the communication terminal 2 via wireless communication. The base station 3 also communicates with the core device 4 via a communication path C in a wired section such as a backhaul link, exchanging data and control information. The base station 3 has a communication interface (not shown) for communicating with the signal processing device 1.

[0022] The core device 4 is a device provided in a core network that complies with a predetermined communication standard such as LTE, 5G, or 6G. For example, in the case of a 5G communication network, it includes a UPF (User Plane Function) in the U-plane and a UDR (Unified Data Repository) in the C-plane. The core device 4 has a communication interface (not shown) for communicating with the signal processing device 1.

[0023] Noise occurring on communication path L in the wireless section between communication terminal 2 and base station 3 may cause a code error, for example, when a digital transmission signal transmitted by communication terminal 2 is decoded by receiving base station 3. Furthermore, noise occurring on communication path C in the wired section between base station 3 and core device 4 may cause a code error, for example, when a digital transmission signal transmitted by base station 3 is decoded by receiving core device 4. The noise occurring on communication paths L and C is, in particular, Gaussian noise that follows a normal distribution.

[0024] To improve the error rate of transmission signals, a method is known in which maximum likelihood detection (MLD) is used to determine the most likely signal from received transmission signals and perform signal separation based on the maximum likelihood determination (see Patent Document 2). This method, known as maximum likelihood detection, creates replicas of the received signal for all possible combinations of transmitted symbols, compares them with the received signal, and determines the most likely one as having been transmitted.

[0025] In maximum likelihood detection, when Gaussian noise is added to a binary digital signal, the appropriate threshold ρ0 that minimizes the bit error rate is expressed by the following equation (1).

number

[0026] In the above formula (1), S0 is 0, S1 is 1, and σ 2 denotes the variance of Gaussian noise, P[S0] denotes the probability that S0 = 0, and P[S1] denotes the probability that S1 = 1. The threshold ρ0 expressed in the above equation (1) is known to be applied to, for example, a matched filter that minimizes the error rate of a binary signal.

[0027] In the signal processing system according to this embodiment, the threshold value ρ0 in the above formula (1) is used as a threshold value ρ0' for suppressing the error rate between a normal signal and a pseudo signal in a transmission signal, that is, for minimizing the error rate. A normal signal refers to, for example, a data string of each pattern of a transmission signal of a data string of a set length transmitted from a communication terminal 2 to a base station 3. For example, in the case of a transmission signal of a binary digital signal of 3 bits length, 2 3 There exists a normal signal with a data sequence of two types. On the other hand, a pseudo signal is a data sequence with a different pattern from the pattern of the data sequence of a normal signal. A normal signal is a data sequence that does not contain any bit errors, while a pseudo signal corresponds to a data sequence in which a bit error occurs in the data sequence of a normal signal due to the addition of Gaussian noise.

[0028] For example, 2 3 In the case of a normal signal with two data sequence patterns, eight data sequence patterns are defined as normal signals: 000, 001, 010, 011, 100, 101, 110, and 111. In contrast, for example, a pseudo signal for data sequence 111 of the eight normal signal patterns is a data sequence other than 111, and for example, any of 000, 001, 010, 011, 100, 101, and 110 could be a pseudo signal for data sequence 111 related to a normal signal.

[0029] The threshold value ρ0′ for minimizing the error rate between a normal signal and a pseudo signal in a transmission signal is expressed by the following equation (2) based on the above equation (1).

number

[0030] In the above equation (2), S0 is 0, i.e., indicates a false signal, S1 is 1, i.e., indicates a normal signal, and σ 2 denotes the variance of Gaussian noise. E D(x)=1 is the expected value (first expected value) at which the classifier 122 included in the learning unit 12 that performs learning of a GAN (generative adversarial network) described later identifies a normal signal as a normal signal (so-called real data), and E D(x)=0indicates an expected value (second expected value) at which the classifier 122 included in the learning unit 12, which will be described later, identifies a pseudo signal as a pseudo signal (so-called fake data).

[0031] In this way, the signal processing system according to this embodiment determines the threshold value ρ0' for suppressing the error rate between normal signals and pseudo signals using the above equation (2). That is, the base station 3 can use the determined threshold value ρ0' as a decision threshold for a transmission signal to which specific Gaussian noise generated in the communication environment including the communication path L in the wireless section and the base station 3 has been added. Therefore, when the receiving base station 3 decodes the transmission signal, it is possible to minimize the error rate between normal signals and pseudo signals. The same applies to the case of a transmission signal to which specific Gaussian noise generated in the communication environment including the communication path C in the wired section has been added.

[0032] [Signal processing unit functional blocks] As shown in FIG. 1, the signal processing device 1 includes a setting unit 10, a collecting unit 11, a learning unit 12, a determining unit 13, a storage unit 14, and a notifying unit 15.

[0033] The setting unit 10 sets each pattern of the data string of the transmission signal of the data string of the set length as a normal signal. Specifically, in the case of n (n is an integer of 2 or more) data strings, the setting unit 10 sets 2 n As described above, in the case of a 3-bit data string, each of the eight data strings is set as a normal signal.

[0034] The collection unit 11 collects the noise variance from the device. Specifically, the collection unit 11 collects the Gaussian noise variance σ from the target base station 3 or core device 4 for which the threshold ρ′ is to be calculated. 2 The variance of Gaussian noise is 2 The values ​​used are calculated from the distortion of the voltage value of the pilot signal in each of the base station 3 and the core device 4.

[0035] The learning unit 12 performs adversarial learning on a GAN having a generator 121 and a classifier 122. More specifically, the learning unit 12 learns a GAN (machine learning model) having a generator 121 that generates a pseudo signal, which is a data sequence having a pattern different from the pattern of a data sequence of a normal signal, and a classifier 122 that distinguishes between the pseudo signal generated by the generator 121 and a normal signal.

[0036] 3 and 4 are diagrams schematically illustrating the neural network configuration of the generator 121 and the classifier 122 of the GAN used by the learning unit 12. As shown in FIG. 3, the generator 121 is configured as a neural network having an input layer, a hidden layer, and an output layer. The generator 121 is a model that generates a pseudo signal from random noise. M randomly sampled vectors of Gaussian noise (z1 to z m The generator 121 performs a multiplication and accumulation operation on the input and weight parameters and threshold processing using an activation function to generate outputs G(z1) to G(z n ) is output.

[0037] 4 is configured as a neural network having an input layer, a hidden layer, and an output layer. In the example of FIG. 4, m sampled data strings x1 to x2 of training data are used as input. m is given. The classifier 122 performs a product-sum operation on the input and weight parameters and threshold processing using an activation function, and outputs, for example, 1 or 0. When the classifier 122 correctly identifies the training data related to the input normal signal as a normal signal, it outputs an output y=1. On the other hand, when the classifier 122 correctly identifies the training data related to the input pseudo signal as a pseudo signal, it outputs an output y=0. In this way, the classifier 122 is a model that correctly distinguishes between a normal signal, which is real data, and a pseudo signal, which is fake data generated by the generator 121.

[0038] The learning unit 12 calculates a first expected value E of the objective function E of the GAN (described later) until the generator 121 generates pseudo signals having different data sequence patterns for each pattern of the data sequence of the normal signal set by the setting unit 10.D(x)=1 The value of the first expected value E is adjusted, and the generator 121 and the classifier 122 are trained in an adversarial manner. D(x)=1 is the expected value at which the discriminator 122 discriminates a normal signal as a normal signal.

[0039] FIG. 2 is a block diagram for explaining the adversarial learning of GAN by the learning unit 12. The generator 121 of the GAN adopted by the learning unit 12 is represented as a function G, and the discriminator 122 is represented as a function D. Furthermore, genuine data that is a normal signal is represented as x, the predicted value that is output by the discriminator 122 is represented as y, and the correct label is represented as t. The correct label t is set to 1 for genuine data that is a normal signal, and 0 for fake data that is a pseudo signal generated by the generator 121. In this case, the discriminator 122 calculates the cross entropy E CE It can be expressed as:

[0040]

number

[0041] The first term in the brace of the above equation (3) represents t n lny n In this case, the predicted value y n is the correct label of the normal signal, t n = 1. On the other hand, the second term in the braces represents (1-t n )ln(1-y n ), the predicted value y n is the correct label value (1-t n ) = 0. In this way, the cross entropy E CE is the maximum value when the predicted value matches the correct label value.

[0042] Here, the generator 121 that constitutes the GAN has parameters w G ,θ G and the function G(w G ,θ G ) The classifier 122 also uses the parameter w D ,θD and function D(w D ,θ D ) The cross entropy E in the above equation (3) CE The objective function E of the GAN including the generator 121 and the discriminator 122 based on the above can be expressed by the following equation (4).

number

[0043] The first term in the above equation (4) represents E D(x)=1 lnD(w D ,θ D ) is the first expected value at which the discriminator 122 discriminates a normal signal as a normal signal. D(x)=0 ln(1-D(G(w G ,θ G ),(w D ,θ D )) is a second expected value at which the discriminator 122 discriminates a pseudo signal generated by the generator 121 as a pseudo signal. In GAN training, the generator 121 and the discriminator 122 are trained adversarially through min-max optimization of the objective function E. Therefore, the generator 121 is trained to be able to generate pseudo signals that deceive the discriminator 122, and the discriminator 122 is trained to discriminate the pseudo signal generated by the generator 121 as a pseudo signal.

[0044] In learning of the classifier 122, when a normal signal is given, the classifier 122 outputs an output close to y=1, thereby maximizing the first term of the objective function E in the above equation (4). On the other hand, when a pseudo signal is given, the classifier 122 learns to output an output close to y=0, thereby maximizing the second term of the objective function E.

[0045] In the learning of the generator 121, D(G(w G ,θ G ),(w D ,θ D )) (D(Gz))) in Figure 2 is close to 1. G ,θ G) (G(z) in FIG. 2), that is, by deceiving the classifier 122, the objective function E is minimized. The learning unit 12 uses a learning procedure in which the parameters of the generator 121 and the parameters of the classifier 122 are updated alternately. The learning procedure of the generator 121 and the classifier 122 by the learning unit 12 will be described in detail later.

[0046] The learning unit 12 learns two n For each normal signal in the data sequence of the pattern, the generator 121 and the classifier 122 are alternately trained. n The first expected value E is calculated until the generator 121 generates pseudo signals having different patterns of data strings for each normal signal of the pattern. D(x)=1 The value of is adjusted to train the generator 121 and the classifier 122 in an adversarial manner.

[0047] For example, consider a case where, among normal signals of eight patterns of 3-bit data sequences, a pseudo signal of the data sequence 101 is generated by the generator 121 for a normal signal of a certain data sequence 111, and a pseudo signal of the data sequence 101 is also generated by the generator 121 for a normal signal of another data sequence 110. In this case, the learning unit 12 continues to adjust the first expected value E until the pseudo signals generated for the data sequences 111 and 110 of the normal signals become data sequences with different patterns. D(x)=1 The value of is adjusted, and learning is repeated between the generator 121 and the discriminator 122. Therefore, the data strings of the pseudo signals are data strings with patterns that do not overlap with each other.

[0048] When the objective function E of the GAN is optimized, the learning unit 12 calculates a first expected value E D(x)=1 and the second expected value E D(x)=0 The value of is passed to the determination unit 13.

[0049] The determination unit 13 determines a threshold value for suppressing the error rate between a normal signal and a pseudo signal in a transmission signal to which noise has been added, based on the parameters of the machine learning model optimized by the learning unit 12. More specifically, the determination unit 13 determines a first expected value E obtained by optimizing the objective function E of the GAN by the learning unit 12. D(x)=1 and the second expected value E D(x)=0 The determination unit 13 calculates the threshold value ρ0' that minimizes the error rate between the normal signal and the pseudo signal by substituting the value of ρ0' into the above equation (2). The determination unit 13 also calculates the variance σ 2 Substitute into the above equation (2).

[0050] The storage unit 14 stores information about the models of the generator 121 and the classifier 122 of the GAN, and the above formulas (1) to (4) used by the learning unit 12 and the determination unit 13.

[0051] The notification unit 15 notifies the determined threshold value ρ0' to the base station 3 or core device 4, which is a device that receives the transmission signal.

[0052] [Hardware configuration of signal processing device] Next, an example of a hardware configuration for realizing the signal processing device 1 having the above-described functions will be described with reference to FIG.

[0053] 5, the signal processing device 1 can be realized by, for example, a computer including a processor 102, a main memory device 103, a communication interface 104, an auxiliary memory device 105, and an input / output (I / O) 106, which are connected via a bus 101, and a program that controls these hardware resources. The signal processing device 1 can also include a display device 107 connected via the bus 101.

[0054] The main memory device 103 pre-stores programs for the processor 102 to perform various controls and calculations. The processor 102 and the main memory device 103 implement the functions of the signal processing device 1, such as the setting unit 10, the collecting unit 11, the learning unit 12, the determining unit 13, and the notifying unit 15 shown in FIG.

[0055] The communication interface 104 is an interface circuit for connecting the signal processing device 1 to various external electronic devices via a network.

[0056] The auxiliary storage device 105 is composed of a readable / writable storage medium and a drive for reading and writing various information such as programs and data from and to the storage medium. The auxiliary storage device 105 can use a semiconductor memory such as a hard disk or flash memory as the storage medium.

[0057] The auxiliary storage device 105 has a program storage area for storing a GAN learning program executed by the signal processing device 1. The auxiliary storage device 105 realizes the storage unit 14 described in FIG. 1. The auxiliary storage device 105 also has an area for storing information about the base station 3, the core device 4, and the communication channels L and C for which the threshold value ρ0′ is to be set. Furthermore, the auxiliary storage device 105 may have, for example, a backup area for backing up the above-mentioned data, programs, etc.

[0058] The input / output I / O 106 is an input / output device that inputs signals from external devices and outputs signals to external devices.

[0059] The display device 107 is configured by an organic EL display, a liquid crystal display, etc. The notification unit 15 can also be realized by the display device 107.

[0060] [Operation of signal processing device] Next, the operation of the signal processing device 1 having the above-described configuration will be described with reference to the sequence diagram of FIG. 6 and the flowchart of FIG.

[0061] 6 is an operation sequence showing an outline of the operation of a signal processing system including the signal processing device 1. In the example of FIG. 6, the device for which the threshold value ρ0' is set is the core device 4. First, in the core device 4, the variance σ of Gaussian noise occurring in the communication path C of the wired section between the base station 3 and the core device 4 is calculated. 2 and transmits it to the signal processing device 1. The signal processing device 1 measures the variance σ of the intrinsic Gaussian noise measured by the core device 4. 2 are collected (step S1).

[0062] Next, the signal processing device 1 performs adversarial learning of the GAN to optimize parameters of the generator 121 and the classifier 122 (step S100). The learning process of step S100 will be described in detail later with reference to the flowchart in FIG.

[0063] Thereafter, the signal processing device 1 calculates the first expected value E of the objective function E of the above equation (4) obtained as a result of the learning process. D(x)=1 and the second expected value E D(x)=0 The value of σ, as well as the variance of the Gaussian noise specific to the core device 4 collected in step S1 2 Based on this, the signal processing device 1 determines a threshold value ρ0' that minimizes the error rate between a normal signal and a pseudo signal in the transmission signal (step S13). Thereafter, the signal processing device 1 notifies the core device 4 of the threshold value ρ0' determined in step S13 (step S14).

[0064] The core device 4 that has received the notification sets the voltage value of the determination threshold based on the threshold ρ0' (step S101). Note that the process shown in Fig. 6 is performed for each device, and for example, if there are multiple base stations 3, the threshold ρ0' is determined for each base station 3. Similarly, for the core device 4, the threshold ρ0' is determined for each UPF and UDR.

[0065] Next, the operation of the signal processing device 1 will be described with reference to the flowchart in Fig. 7 and the block diagram of the learning unit 12 in Fig. 2. Note that steps S1, S13, and S14 in Fig. 7 are the same as steps S1, S13, and S14 described in Fig. 6.

[0066] First, the collection unit 11 receives the Gaussian noise variance σ from the base station 3 or core device 4, which is the device for which the threshold ρ0' is to be determined. 2 (Step S1). Next, the setting unit 10 sets the data string of each pattern contained in the transmission signal of the data string of the set length as a normal signal (Step S2). Specifically, the setting unit 10 sets the data string of each pattern contained in the transmission signal of the n-bit length as a normal signal (Step S3). n Each of the eight signal patterns is set as a normal signal. For example, when n=3, eight patterns of data strings are set as normal signals. The normal signals set by the setting unit 10 are training data used by the learning unit 12.

[0067] 2, the normal signal set in step S2 is used as training data 124 to be input when training the classifier 122. In the example of FIG. 2, the training data 124 of the data sequence 111 is provided to the classifier 122.

[0068] Next, the learning unit 12 n Normal signals of the signal patterns are input to the classifier 122 as training data 124, and the parameter w of the classifier 122 is set to classify the normal signals as normal signals (y=1). D ,θ D (Step S3). In Step S3, the learning unit 12 can cause the classifier 122 to learn the normal signal using, for example, backpropagation. In Step S3, the classifier 122 that can distinguish a normal signal from another normal signal is constructed in advance.

[0069] Next, the learning unit 12 generates Gaussian noise 120 and provides a random vector of the generated Gaussian noise 120 as an input to the generator 121 (step S4). Subsequently, the generator 121 generates a random vector of the generated Gaussian noise 120 based on the input z and the weight parameter w G ,θ GA product-sum operation and threshold processing using an activation function are performed to generate a pseudo signal G(z) (step S5). Note that, for example, if the pattern of the data sequence of a normal signal is 111, a pseudo signal may be a data sequence with a pattern other than 111, such as 101.

[0070] Next, the classifier 122 is trained. The training of the classifier 122 is performed by using the parameter w D ,θ D First, the learning unit 12 provides the training data 124 of the normal signal set in step S2 as an input to the discriminator 122, and calculates the gradient dE / dw so that the objective function E of the above formula (4) is maximized. D ,dE / dθ D Calculate the parameter w by backpropagation etc. D ,θ D (Step S6). The label of the training data 124 is set to 1 (normal signal).

[0071] Next, the learning unit 12 inputs the pseudo signal generated by the generator 121 to the discriminator 122 in step S5, and calculates the gradient dE / dw so that the objective function E in the above equation (4) is maximized. D ,dE / dθ D Calculate the parameter w by backpropagation etc. D ,θ D (Step S7). That is, in steps S6 and S7, in order to maximize the objective function E in the above equation (4), the first term is updated as D(w D ,θ D )=1 is output, and the second term is D(G(w G ,θ G ),(w D ,θ D ))=0. In the training data 124, the label is set to 0 (pseudo signal).

[0072] The learning of the classifier 122 in steps S6 and S7 corresponds to the dashed arrows shown in FIG. 2, in which a classifier error is calculated from block 125 of objective function E based on output 123 from the classifier 122, and the error is then backpropagated to the classifier 122.

[0073] Next, the generator 121 is trained. In the training of the generator 121, the parameters of the discriminator 122 are fixed. The training unit 12 trains the generator 121 so that a pseudo signal is generated when random Gaussian noise 120 is given to the generator 121. Specifically, the training unit 12 calculates the gradient −dE / dw G ,-dE / dθ G Calculate the parameter w by backpropagation etc. G ,θ G is updated (step S8).

[0074] The learning in step S8 corresponds to the arrow indicated by the dashed line in which the error is backpropagated to the generator 121 shown in Fig. 2. That is, based on the output 123 when the pseudo signal generated by the generator 121 shown in Fig. 2 is input to the discriminator 122, a generator error is calculated in the block 125 of the objective function E, and this corresponds to the arrow indicated by the dashed line in which the error is backpropagated to the generator 121.

[0075] Thereafter, the learning of the classifier 122 and the generator 121 from step S4 to step S8 is repeated until the value of the objective function E reaches the Nash equilibrium and converges (step S9: NO). On the other hand, if the value of the objective function E has converged (step S9: YES), n The remaining 2 normal signal data strings of the pattern n For the data string of the -1 pattern, the processes from step S2 to step S9 are repeated until the generator 121 and the discriminator 122 are trained (step S10: NO).

[0076] Then the remaining 2 nWhen the generator 121 and the discriminator 122 have been trained on a data sequence of −1 pattern (step S10: YES), the training unit 12 determines whether or not the generator 121 has generated pseudo signals with mutually different data sequence patterns for each signal pattern of a normal signal (step S11). For example, in the case of a data sequence with a length of 3 bits and eight patterns, consider a case in which the generator 121 generates a pseudo signal of the data sequence 101 for a data sequence 111 of a normal signal, and also generates a pseudo signal of the data sequence 101 for a data sequence 110 of another normal signal. In this case, in step S11, it is determined whether or not the data sequences of the pseudo signals generated for the data sequences 111 and 110 of the normal signals are eight different data sequences. That is, the patterns of the data sequences of the pseudo signals generated for each of the eight normal signal patterns are such that they do not overlap with each other.

[0077] If no pseudo signals related to data sequences with different patterns are generated for normal signals of all patterns of data sequences (step S11: NO), the learning unit 12 calculates the first term of the objective function E, i.e., the first expected value E D(x)=1 The value of is changed and adjusted, and the learning process from step S2 to step S10 is repeated. Therefore, in the example of the eight patterns of data sequences above, the patterns of the eight pseudo signal data sequences generated for each of the eight patterns of normal signal data sequences are different from each other and there are no overlapping signals.

[0078] After that, if pseudo signals of data sequences with different patterns are generated for normal signals of all patterns of data sequences (step S11: YES), the learning unit 12 calculates the first expected value E of the optimized objective function E of the GAN. D(x)=1 and the second expected value E D(x)=0 (Step S12).

[0079] Next, the determination unit 13 applies the first expected value E determined in step S12 to the above equation (2). D(x)=1 and the second expected value E D(x)=0and the variance σ of the Gaussian noise collected from the base station 3 or core device 4, which is the device to be processed and collected in step S1, is then calculated. 2 and determines the threshold value ρ0′ (step S13). After that, the notification unit 15 notifies the determined threshold value ρ0′ to the processing target device, that is, the base station 3 or the core device 4 (step S14).

[0080] As described above, the signal processing device 1 according to this embodiment performs adversarial learning of a GAN having a generator 121 that sets each pattern of a transmission signal of a data sequence of a set length as a normal signal, generates pseudo signals that are data sequences with patterns different from the patterns of the data sequence of the normal signal, and a classifier 122 that distinguishes between the pseudo signals generated by the generator 121 and normal signals. Furthermore, a first expected value E of the objective function E optimized by learning is calculated. D(x)=1 and the second expected value E D(x)=0 The threshold ρ0' that minimizes the error rate between the normal signal and the pseudo signal of the transmission signal is determined using the above equation. Therefore, the error rate of the signal can be reduced more easily without adding redundant bits.

[0081] Furthermore, according to the signal processing device 1 of this embodiment, a GAN machine learning model including a generator 121 and a discriminator 122 is constructed for each base station 3 and each core device 4, so that measures to reduce the bit error rate for each device can be realized with a simpler configuration.

[0082] The above describes embodiments of the signal processing device and signal processing method of the present invention, but the present invention is not limited to the described embodiments, and various modifications that a person skilled in the art can make within the scope of the invention described in the claims are possible. [Explanation of symbols]

[0083] 1...signal processing device, 10...setting unit, 11...collection unit, 12...learning unit, 13...decision unit, 14...memory unit, 15...notification unit, 2...communication terminal, 3...base station, 4...core device, 101...bus, 102...processor, 103...main memory device, 104...communication interface, 105...auxiliary memory device, 106...input / output I / O, 107...display device, 120...noise, 121...generator, 122...discriminator, 123...output, 124...training data, 125...block of objective function E, C, L...communication channel, NW...network.

Claims

1. a setting unit configured to set a data sequence of each pattern of a transmission signal, which is a data sequence of a set length, as a normal signal; a learning unit configured to learn a machine learning model having a generator that generates a pseudo signal, which is a data sequence having a pattern different from a pattern of a data sequence of the normal signal, and a classifier that distinguishes between the pseudo signal generated by the generator and the normal signal; a determination unit configured to determine a threshold value for suppressing an error rate between the normal signal and the pseudo signal in the transmission signal to which noise has been added, based on parameters of the machine learning model optimized by learning by the learning unit; and a notification unit configured to notify a device receiving the transmission signal of the determined threshold value; A signal processing device comprising:

2. 2. The signal processing device according to claim 1, The determination unit sets the threshold value based on a first expected value at which the classifier will classify the normal signal as a normal signal and a second expected value at which the classifier will classify the pseudo signal generated by the generator as a pseudo signal, both of which are indicated by an objective function of the machine learning model. A signal processing device comprising:

3. 3. The signal processing device according to claim 2, The learning unit adjusts the value of the first expected value until the generator generates the pseudo signals having data sequence patterns different from each other for each pattern of the data sequence of the normal signal set by the setting unit, and performs adversarial learning between the generator and the classifier. A signal processing device comprising:

4. 3. The signal processing device according to claim 2, further comprising a collection unit configured to collect the variance of the noise from the device; The determination unit determines the threshold value based on the variance collected by the collection unit, the first expected value, and the second expected value. A signal processing device comprising:

5. a setting step of setting each pattern of a data string of a transmission signal having a set length as a normal signal; a learning step of learning a machine learning model having a generator that generates a pseudo signal, which is a data sequence having a pattern different from the pattern of the data sequence of the normal signal, and a classifier that distinguishes between the pseudo signal generated by the generator and the normal signal; a determining step of determining a threshold value for suppressing an error rate between the normal signal and the pseudo signal in the transmission signal to which noise has been added, based on parameters of the machine learning model optimized by learning in the learning step; a notification step of notifying a device receiving the transmission signal of the determined threshold value; A signal processing method comprising:

6. 6. The signal processing method according to claim 5, The determining step sets the threshold value based on a first expected value at which the classifier will classify the normal signal as a normal signal and a second expected value at which the classifier will classify the pseudo signal generated by the generator as a pseudo signal, both of which are indicated by an objective function of the machine learning model. A signal processing method comprising:

7. 7. The signal processing method according to claim 6, The learning step adjusts the value of the first expected value until the generator generates pseudo signals having data sequence patterns different from each other for each pattern of the data sequence of the normal signal set in the setting step, and performs adversarial learning between the generator and the classifier. A signal processing method comprising:

8. 7. The signal processing method according to claim 6, further comprising a collecting step of collecting the variance of the noise from the device; The determining step determines the threshold value based on the variance collected in the collecting step, the first expected value, and the second expected value. A signal processing method comprising:

Images