Classification device, training method for classification device, and computer program

Abstract

A discrimination device (1) is provided with: a discriminating means (110) for discriminating signals; and a learning means (120) for employing training data (X) having periodicity to execute machine learning so that discrimination results from the discriminating means indicate periodicity. According to the foregoing discrimination device, suitably executing machine learning in order to discriminate signals is possible. This makes it possible to suitably discriminate signals having periodicity, such as radar waves and sound waves, emanating from a plurality of transmission sources, for example.

Description

[Technical field]

[0001] The present invention relates to the technical fields of a discrimination device for discriminating a signal, a training method for the discrimination device, and a computer program. [Background technology]

[0002] Known devices of this type are those that identify (in other words, classify) multiple signals. For example, Patent Document 1 discloses a technique for analyzing radio waves transmitted from multiple transmitting devices and identifying and classifying the transmitting devices individually. Patent Document 2 discloses a technique for separating pulses into groups based on the pulse width, pulse repetition period, pulse amplitude, etc. of the radar pulse. Patent Document 3 discloses a technique for separating radar pulses into signals based on the pulse arrival period. Patent Document 4 discloses a technique for classifying local periodic components of radar waves into a predetermined number of categories based on their characteristics.

[0003] As another related technique, Patent Document 5 discloses a technique for performing perturbation processing on training samples to generate additional training samples. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2016-158209 A [Patent Document 2] JP 2008-139056 A [Patent Document 3] JP 2005-283276 A [Patent Document 4] JP 2002-303667 A [Patent Document 5] JP 2012-014270 A Summary of the Invention [Problem to be solved by the invention]

[0005] When identifying multiple radar waves, a method of identifying them using, for example, the signal strength for each frequency of the radar pulse is conceivable. However, an identification device employing such a method is relatively costly. In particular, when optimizing an identification device (e.g., a neural network) by machine learning, labeling training data is costly. The above-mentioned patent documents do not anticipate such a problem, and there is room for improvement.

[0006] The present invention has been made in consideration of the above problems, and an object of the present invention is to provide a classification device, a training method for a classification device, and a computer program capable of appropriately executing machine learning for classifying signals. [Means for solving the problem]

[0007] One aspect of the identification device of the present invention includes an identification means for identifying a signal, and a learning means for performing machine learning using training data having periodicity so that the identification result of the identification means exhibits periodicity.

[0008] One aspect of the training method for a classification device of the present invention is a training method for a classification device that classifies a signal, which performs machine learning using training data having periodicity so that the classification results of the classification device exhibit periodicity.

[0009] One aspect of the computer program of the present invention is a training method for a classification device that classifies signals, which operates a computer to perform machine learning using training data having periodicity so that the classification results of the classification device exhibit periodicity. Effect of the Invention

[0010] According to one aspect of each of the above-described identification device, the learning method for the identification device, and the computer program, it is possible to appropriately execute machine learning for identifying signals, thereby making it possible to appropriately identify signals emitted from, for example, multiple sources. [Brief description of the drawings]

[0011] [Figure 1] 1 is a block diagram showing a hardware configuration of a recognition device according to a first embodiment. [Diagram 2] 1 is a block diagram showing a functional configuration of a discrimination device according to a first embodiment. [Diagram 3] 5 is a flowchart showing a flow of a learning operation of the identification device according to the first embodiment. [Figure 4] FIG. 2 is a conceptual diagram illustrating an example of an input and an output in a classifier according to the first embodiment. [Diagram 5] 1 is a graph showing an example in which an autocorrelation function exhibits periodicity. [Figure 6] FIG. 11 is a block diagram showing a functional configuration of a discrimination device according to a second embodiment. [Figure 7] 10 is a flowchart showing a flow of a learning operation of the discrimination device according to the second embodiment. [Figure 8] FIG. 13 is a block diagram showing a functional configuration of a discrimination device according to a third embodiment. [Figure 9] 13 is a flowchart showing the flow of a learning operation of the identification device according to the third embodiment. [Figure 10] FIG. 13 is a block diagram showing a functional configuration of a discrimination device according to a fourth embodiment. [Figure 11] 13 is a flowchart showing the flow of a learning operation of the identification device according to the fourth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, embodiments of a classification device, a training method for a classification device, and a computer program will be described with reference to the drawings.

[0013] First Embodiment An identification device according to a first embodiment will be described with reference to Fig. 1 to Fig. 5. In the following, an example will be described in which the identification device is applied as a device for identifying radar waves.

[0014] (Hardware configuration) First, the hardware configuration of the identification device according to the first embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the hardware configuration of the identification device according to the first embodiment.

[0015] 1, the identification device 1 according to the first embodiment includes a CPU (Central Processing Unit) 11, a RAM (Random Access Memory) 12, a ROM (Read Only Memory) 13, and a storage device 14. The identification device 1 may further include an input device 15 and an output device 16. The CPU 11, the RAM 12, the ROM 13, the storage device 14, the input device 15, and the output device 16 are connected via a data bus 17.

[0016] The CPU 11 reads a computer program. For example, the CPU 11 is configured to read a computer program stored in at least one of the RAM 12, the ROM 13, and the storage device 14. Alternatively, the CPU 11 may read a computer program stored in a computer-readable storage medium using a storage medium reading device (not shown). The CPU 11 may obtain (i.e., read) a computer program from a device (not shown) disposed outside the identification device 1 via a network interface. The CPU 11 controls the RAM 12, the storage device 14, the input device 15, and the output device 16 by executing the computer program that the CPU 11 reads. In particular, in this embodiment, when the CPU 11 executes the computer program that it reads, a functional block for learning the identification device 1 is realized in the CPU 11.

[0017] The RAM 12 temporarily stores a computer program executed by the CPU 11. The RAM 12 temporarily stores data that is temporarily used by the CPU 11 when the CPU 11 is executing the computer program. The RAM 12 may be, for example, a D-RAM (Dynamic RAM).

[0018] The ROM 13 stores computer programs executed by the CPU 11. The ROM 13 may also store other fixed data. The ROM 13 may be, for example, a programmable ROM (P-ROM).

[0019] The storage device 14 stores data to be stored long-term by the identification device 1. The storage device 14 may operate as a temporary storage device for the CPU 11. The storage device 14 may include, for example, at least one of a hard disk device, a magneto-optical disk device, an SSD (Solid State Drive), and a disk array device.

[0020] The input device 15 is a device that receives an input instruction from a user of the identification device 1. The input device 15 may include, for example, at least one of a keyboard, a mouse, and a touch panel.

[0021] The output device 16 is a device that outputs information related to the identification device 1 to the outside. For example, the output device 16 may be a display device (for example, a display) that can display information related to the identification device 1.

[0022] (Functional structure for learning) Next, a functional configuration for learning of the classification device 1 according to the first embodiment will be described with reference to Fig. 2. Fig. 2 is a block diagram showing the functional configuration for learning of the classification device according to the first embodiment.

[0023] As shown in FIG. 2, the classification device 1 according to the first embodiment includes a classifier 110 and a learning unit 120 as functional blocks or processing circuits for realizing the functions thereof.

[0024] The classifier 110 is configured to be able to distinguish between a plurality of signals. For example, the classifier 110 is configured to include a neural network or the like, and is configured to be able to distinguish between a plurality of signals emitted from different emission sources, one for each emission source. An example of a signal that the classifier 110 can distinguish is a signal having periodicity, such as a radar wave or a sound wave. However, the classifier 110 may also be able to distinguish between signals that do not have periodicity.

[0025] The learning unit 120 is configured to be able to read the training data X and execute a learning operation (specifically, machine learning) of the classifier 110. The learning unit 120 includes a periodicity loss function calculation unit 201 and a classifier training unit 202.

[0026] The periodicity loss function calculation unit 201 calculates a periodicity loss function (hereinafter, appropriately referred to as a “first loss function”) from the training data X read into the learning unit 120. The first loss function is a function that becomes small when the classification result of the classifier 110 indicates periodicity. The first loss function will be described in detail later. The value of the first loss function calculated by the periodicity loss function calculation unit 201 is configured to be output to the classifier training unit 202.

[0027] The classifier training unit 202 trains the classifier 110 based on the first loss function calculated by the periodicity loss function calculation unit 201. Specifically, the classifier training unit 202 executes a learning operation so as to reduce the first loss function. For example, the classifier training unit 202 adjusts the parameters of the neural network by a known method such as backpropagation so as to reduce the first loss function.

[0028] In the above example, the learning unit 120 executes the learning operation using the first loss function (i.e., periodicity loss function), but the learning unit 120 may execute the learning operation without using a loss function. That is, the method of the learning operation by the learning unit 120 is not particularly limited as long as it uses the periodicity of the signal.

[0029] (Operation flow) Next, the flow of operations of the classification device 1 according to the first embodiment (particularly, the flow of operations for training the classifier 110) will be described with reference to Fig. 3. Fig. 3 is a flowchart showing the flow of learning operations of the classification device according to the first embodiment.

[0030] 3, during operation of the classification device 1 according to the first embodiment, the learning unit 120 first initializes parameters (i.e., values ​​changed by the learning operation) of the classifier 110 (step S101). Note that, when the learning operation is repeatedly executed, the process of step S101 may be omitted as appropriate.

[0031] Next, the learning unit 120 reads the training data X (step S102). Note that the training data X is data having periodicity, but the specific period may be unknown. In other words, the training data X may be data to which no correct answer label has been assigned.

[0032] Next, the periodicity loss function calculation unit 201 calculates a first loss function using the loaded training data X (step S103). Then, the classifier training unit 202 updates (i.e., changes) the parameters of the classifier 110 so that the first loss function becomes smaller (step S104).

[0033] Thereafter, the learning unit 120 determines whether the classification result of the classifier 110 has converged (step S105). The learning unit 120 determines the convergence, for example, by whether the absolute value or the decrease amount of the loss function is below a preset threshold value. If the classification result has converged (step S105: YES), the series of processes ends. On the other hand, if the classification result has not converged (step S105: NO), the process starts again from step S103. Therefore, the processes of steps S103 and S104 are repeatedly executed until the classification result converges.

[0034] (Specific operation example) Next, a specific operation example (i.e., a specific example of a learning operation) of the classification device 1 according to the first embodiment will be described with reference to Fig. 4 and Fig. 5. Fig. 4 is a conceptual diagram showing an example of an input and an output in the classifier according to the first embodiment. Fig. 5 is a graph showing an example in which the autocorrelation function shows periodicity.

[0035] As shown in Fig. 4, the classifier 110 is configured as a neural network having layers from the first layer to the Nth layer. In this case, training data X (x1, x2, ..., x t . . ) is input to the discriminator 110, the discrimination result y (y1, y2, . . . , y t ...) is obtained. The classification result y indicates a class posterior probability series, and is output as a value indicating the likelihood of each classified class, as shown in the example in the figure. In the example in the figure, the probability of being the first class is "0.8", the probability of being the second class is "0.01", and the probability of being the third class is "0.05". The classification results y here are adjusted by a softmax function so that the sum becomes 1.

[0036] For example, the entropy of an autocorrelation function can be used to train the above-mentioned classifier 110. The autocorrelation function R(τ) and its entropy L1 can be expressed as the following formulas (1) and (2), respectively.

[0037]

number

[0038]

number

[0039] In addition, the expected value E[·] in the above formulas (1) and (2) may be an average, or simply a sum. In this case, the autocorrelation function R(τ) and its entropy L1 can be expressed as the following formulas (3) and (4), respectively.

[0040]

number

[0041]

number

[0042] If the training data X includes multiple observed signals i=1, 2, ..., the sums of these signals can also be calculated. In this case, the above formulas (3) and (4) become the following formulas (5) and (6), respectively.

[0043]

number

[0044]

number

[0045] As shown in FIG. 5, when the entropy L1 of the autocorrelation function R(τ) becomes small, clear peaks (peak A and peak B in the figure) are observed in the autocorrelation function R(τ). These peaks A and B are caused by signals having different periodicities. Therefore, if the learning operation is carried out so as to reduce the entropy L1 of the autocorrelation function R(τ), it becomes possible to identify the signal according to the periodicity. Note that the above formulas (3) and (5) are not limited to the autocorrelation function as long as they have the same properties as the autocorrelation function. In other words, instead of the autocorrelation function, it is also possible to use a function other than the autocorrelation function. For example, it is possible to use a function obtained as a result of applying a discrete Fourier transform to the identification result y.

[0046] (Technical effect) Next, technical effects obtained by the identification device 1 according to the first embodiment will be described.

[0047] As described in FIG. 1 to FIG. 5, in the classification device 1 according to the first embodiment, machine learning is executed so that the classification result of the classifier 110 shows periodicity. The classification device 1 thus trained can appropriately classify a signal having periodicity. Furthermore, particularly in this embodiment, the period of the training data X used in the learning operation may be unknown. In other words, appropriate machine learning can be performed without assigning a correct answer label to the training data X. Therefore, the cost of preparing the training data X can be kept low.

[0048] <Second embodiment> Next, the identification device 1 according to the second embodiment will be described with reference to Figures 6 and 7. The second embodiment differs from the first embodiment already described only in part of the configuration and operation, and other parts are generally similar. Therefore, the following will describe in detail the parts that differ from the first embodiment, and will omit explanations of overlapping parts as appropriate.

[0049] (Functional structure for learning) First, the functional configuration for learning of the classification device 1 according to the second embodiment will be described with reference to Fig. 6. Fig. 6 is a block diagram showing the functional configuration for learning of the classification device according to the second embodiment. Among the components shown in Fig. 6, the same components as those shown in Fig. 2 are denoted by the same reference numerals.

[0050] 6, the classification device 1 according to the second embodiment includes a data extension unit 130 in addition to the components of the first embodiment (see FIG. 2). The learning unit 120 also includes a class consistency loss function calculation unit 203.

[0051] The data extension unit 130 is configured to be able to extend the training data by adding perturbation (noise) to the training data X. Note that, since the addition of perturbation can be achieved by appropriately adopting existing technology, a detailed description thereof will be omitted here.

[0052] The class consistency loss function calculation unit 203 calculates a class consistency loss function (hereinafter, appropriately referred to as a "second loss function") from the training data X extended by the data extension unit 130. The second loss function is a function that becomes small when the class that is the classification result of the extended training data X is the same as the class of the training data X before extension. It is not necessary to know which class the data specifically belongs to. The second loss function will be described in detail later. The value of the second loss function calculated by the class consistency loss function calculation unit 203 is configured to be output to the classifier training unit 202.

[0053] (Operation flow) Next, the flow of operations of the classification device 1 according to the second embodiment (particularly, the flow of operations for training the classifier 110) will be described with reference to Fig. 7. Fig. 7 is a flowchart showing the flow of learning operations of the classification device according to the second embodiment. Note that among the processes shown in Fig. 7, the same reference numerals are used for the processes similar to those shown in Fig. 3.

[0054] As shown in FIG. 7, when the classification device 1 according to the second embodiment operates, first, the learning unit 120 initializes parameters of the classifier 110 (step S101) and reads the training data X (step S102).

[0055] Next, in the second embodiment, the data extension unit 130 adds noise to the training data X (step S201). As a result, the training data X is extended.

[0056] Next, the periodicity loss function calculation unit 201 calculates a first loss function using the expanded training data X (step S103). In addition, the class consistency loss function calculation unit 203 calculates a second loss function using the expanded training data X (step S202). The second loss function can be expressed, for example, as the following formula (7) or (8).

[0057]

number

[0058]

number

[0059] Next, the classifier training unit 202 updates (i.e., changes) the parameters of the classifier 110 based on the first loss function and the second loss function (step S104). At this time, the classifier training unit 202 updates the parameters of the classifier 110 so that, for example, L1+λL2 (λ is a predetermined coefficient) becomes small.

[0060] After that, the learning unit 120 judges whether or not the classification result of the classifier 110 has converged (step S105). If the classification result has converged (step S105: YES), the series of processes ends. On the other hand, if the classification result has not converged (step S105: NO), the process starts again from step S103.

[0061] (Technical effect) Next, the technical effects obtained by the identification device 1 according to the second embodiment will be described.

[0062] 6 and 7, in the classification device 1 according to the second embodiment, machine learning is performed by the classifier 110 so that the training data X before extension and the training data X after extension belong to the same class. By learning in this way, even if the training data X is extended, learning using periodicity can be appropriately performed. Therefore, it is possible to appropriately classify a signal having periodicity while suppressing an increase in cost.

[0063] <Third embodiment> Next, the identification device 1 according to the third embodiment will be described with reference to Figures 8 and 9. The third embodiment differs from the first and second embodiments already described only in part of the configuration and operation, and other parts are generally similar. Therefore, the following will describe in detail the parts that differ from the first and second embodiments, and will omit explanations of overlapping parts as appropriate.

[0064] (Functional structure for learning) First, a functional configuration for learning of the classification device 1 according to the third embodiment will be described with reference to Fig. 8. Fig. 8 is a block diagram showing a functional configuration for learning of the classification device according to the third embodiment. Among the components shown in Fig. 8, the same components as those shown in Fig. 2 are denoted by the same reference numerals.

[0065] As shown in Fig. 8, in the classification device 1 according to the third embodiment, in addition to the components of the first embodiment (see Fig. 2), the learning unit 120 includes an error loss function calculation unit 204. Note that, while the first and second embodiments have been described by way of examples in which the classifier 110 classifies signals using a neural network, the classifier 110 according to the third embodiment classifies signals using non-negative matrix factorization (NMF). Since NMF is an existing technology, a detailed description of the classification method using NMF will be omitted.

[0066] The error loss function calculation unit 204 calculates a loss function of the reproduction error used in NMF (hereinafter, appropriately referred to as the "basic loss function") from the training data X. For example, when the training data X is decomposed into a base spectrum H and an activation U, the basic loss function in NMF can be expressed as the following formula (9).

[0067]

number

[0068] The value of the basic loss function calculated by the error loss function calculation unit 204 is output to the classifier training unit 202 .

[0069] (Operation flow) Next, the flow of operations of the classification device 1 according to the third embodiment (particularly, the flow of operations for training the classifier 110) will be described with reference to Fig. 9. Fig. 9 is a flowchart showing the flow of learning operations of the classification device according to the third embodiment. Note that among the processes shown in Fig. 9, the same reference numerals are used for the processes similar to those shown in Fig. 3.

[0070] As shown in FIG. 9, when the classification device 1 according to the third embodiment operates, first, the learning unit 120 initializes parameters of the classifier 110 (step S101) and reads the training data X (step S102).

[0071] Next, in the third embodiment, the error loss function calculation unit 204 calculates a basic loss function from the training data X (step S301). Then, the periodicity loss function calculation unit 201 calculates a first loss function using the training data X (step S103). Here, the first loss function (autocorrelation function R(τ) and entropy L1) can be expressed as the following formulas (10) and (11).

[0072]

number

[0073]

number

[0074] Furthermore, the expected value E[·] in the above formulas (10) and (11) may be an average, or simply a sum, as described in the first embodiment (see formulas (3) and (4)).

[0075] Next, the classifier training unit 202 updates (i.e., changes) the parameters of the classifier 110 based on the basic loss function and the first loss function (step S104). At this time, the classifier training unit 202 updates the parameters of the classifier 110 so that, for example, the following formula (12) becomes smaller.

[0076]

number

[0077] After that, the learning unit 120 judges whether or not the classification result of the classifier 110 has converged (step S105). If the classification result has converged (step S105: YES), the series of processes ends. On the other hand, if the classification result has not converged (step S105: NO), the process starts again from step S103.

[0078] (Technical effect) Next, the technical effects obtained by the identification device 1 according to the third embodiment will be described.

[0079] As described in Fig. 8 and Fig. 9, in the classification device 1 according to the third embodiment, a learning operation is performed by adding the first loss function to the basic loss function of NMF. In this way, the classifier 110 using NMF can perform learning under a constraint on the periodicity of a signal. Therefore, the classifier 110 using NMF can also appropriately classify a signal having periodicity.

[0080] <Fourth embodiment> Next, the identification device 1 according to the fourth embodiment will be described with reference to Fig. 10 and Fig. 11. The fourth embodiment differs from the first to third embodiments already described only in part of the configuration and operation, and other parts are generally similar. Therefore, the following will describe in detail the parts that differ from the first to third embodiments, and will omit explanations of overlapping parts as appropriate.

[0081] (Functional structure for learning) First, a functional configuration for learning of the classification device 1 according to the fourth embodiment will be described with reference to Fig. 10. Fig. 10 is a block diagram showing a functional configuration for learning of the classification device according to the fourth embodiment. Among the components shown in Fig. 10, the same components as those shown in Fig. 6 and Fig. 8 are denoted by the same reference numerals.

[0082] As shown in Fig. 10, the classification device 1 according to the fourth embodiment includes a data augmentation unit 130 in addition to the components of the third embodiment (see Fig. 2). Furthermore, the learning unit 120 includes a class consistency loss function calculation unit 203. That is, the classification device 1 according to the fourth embodiment is substantially a combination of the second and third embodiments. Note that the classifier 110 according to the fourth embodiment classifies a signal using NMF, similar to the third embodiment.

[0083] (Operation flow) Next, the flow of operations of the classification device 1 according to the fourth embodiment (particularly, the flow of operations for training the classifier 110) will be described with reference to Fig. 11. Fig. 11 is a flowchart showing the flow of learning operations of the classification device according to the fourth embodiment. Note that among the processes shown in Fig. 11, the same reference numerals are used for the processes similar to those shown in Figs. 7 and 9.

[0084] As shown in FIG. 11, when the classification device 1 according to the fourth embodiment operates, first, the learning unit 120 initializes parameters of the classifier 110 (step S101) and reads the training data X (step S102).

[0085] Next, in the fourth embodiment, the data extension unit 130 adds noise to the training data X (step S201). As a result, the training data X is extended.

[0086] Next, the error loss function calculation unit 204 calculates a basic loss function from the training data X (step S301). Then, the periodicity loss function calculation unit 201 calculates a first loss function using the training data X (step S103). Also, the class consistency loss function calculation unit 203 calculates a second loss function using the expanded training data X (step S202).

[0087] Next, the classifier training unit 202 updates (i.e., changes) the parameters of the classifier 110 based on the basic loss function, the first loss function, and the second loss function (step S104). That is, the classifier training unit 202 executes a learning operation that combines the learning operation of the second embodiment and the learning operation of the third embodiment already described.

[0088] After that, the learning unit 120 judges whether or not the classification result of the classifier 110 has converged (step S105). If the classification result has converged (step S105: YES), the series of processes ends. On the other hand, if the classification result has not converged (step S105: NO), the process starts again from step S103.

[0089] (Technical effect) Next, the technical effects obtained by the identification device 1 according to the fourth embodiment will be described.

[0090] As described in Fig. 10 and Fig. 11, in the classification device 1 according to the fourth embodiment, a learning operation is performed by adding the first loss function and the second loss function to the basic loss function of NMF. In this way, the classifier 110 using NMF can perform learning under a constraint on the periodicity of the signal. Furthermore, even if the training data X is expanded, learning utilizing the periodicity can be appropriately performed. Therefore, it is possible to appropriately classify a signal having periodicity while suppressing an increase in cost.

[0091] <Additional Notes> The following supplementary notes are further disclosed regarding the above-described embodiment.

[0092] (Appendix 1) The identification device described in Supplementary Note 1 includes an identification means for identifying a signal, and a learning means for performing machine learning using training data having periodicity so that the identification result of the identification means shows periodicity. The present invention relates to an identification device comprising:

[0093] (Appendix 2) The classification device described in Appendix 2 is the classification device described in Appendix 1, characterized in that the learning means performs the machine learning using a first loss function that becomes small when the classification result of the classification means shows periodicity.

[0094] (Appendix 3) The identification device according to claim 1 or 2, wherein the period of the training data is an unknown value.

[0095] (Appendix 4) The classification device described in Appendix 4 is characterized in that it further comprises an extension means for extending the training data by applying a perturbation to the training data, and the learning means performs the machine learning so that the classification result of the training data after the extension is in the same class as the classification result of the training data before the extension.

[0096] (Appendix 5) The classification device described in Appendix 5 is the classification device described in Appendix 4, characterized in that the learning means performs the machine learning using a second loss function that becomes small when the classification result of the training data after expansion is in the same class as the classification result of the training data before expansion.

[0097] (Appendix 6) The identification device described in Supplementary Note 6 is the identification device described in any one of Supplementary Notes 1 to 5, characterized in that the identification means identifies the signal using a neural network.

[0098] (Appendix 7) The identification device according to Supplementary Note 7 is the identification device according to any one of Supplementary Notes 1 to 5, characterized in that the identification means identifies the signal using non-negative matrix factorization.

[0099] (Appendix 8) The identification device described in Supplementary Note 8 is the identification device described in any one of Supplementary Notes 1 to 7, characterized in that the signal is a radar wave or a sound wave.

[0100] (Appendix 9) The training method for a classification device described in Supplementary Note 9 is a training method for a classification device that classifies signals, characterized in that machine learning is performed using training data having periodicity so that the classification results of the classification device exhibit periodicity. It is.

[0101] (Appendix 10) The computer program described in Supplementary Note 10 is a computer program used for training a classification device that classifies signals, characterized in that the computer program operates to perform machine learning using training data having periodicity so that the classification results of the classification device exhibit periodicity.

[0102] (Appendix 11) The identification device according to claim 11 is the identification device according to claim 2, wherein the learning means uses an autocorrelation function as the first loss function.

[0103] (Appendix 12) The classification device according to Supplementary Note 12 is the classification device according to Supplementary Note 2, characterized in that the learning means performs the machine learning using entropy of the first loss function.

[0104] (Appendix 13) The identification device according to claim 13 is the identification device according to claim 5, wherein the learning means uses an autocorrelation function as the second loss function.

[0105] (Appendix 14) The classification device according to Supplementary Note 14 is the classification device according to Supplementary Note 5, characterized in that the learning means performs the machine learning using the entropy of the second loss function.

[0106] The present invention can be modified as appropriate within the scope that does not contradict the gist or idea of ​​the invention that can be read from the claims and the entire specification, and the identification device, the learning method for the identification device, and the computer program involving such modifications are also included in the technical idea of ​​the present invention. [Explanation of symbols]

[0107] 1 Identification device 110 Classifier 120 Learning Department 130 Data Extension 201 Periodicity Loss Function Calculation Part 202 Classifier Training Department 203 Class consistency loss function calculation part 204 Error loss function calculation part X training data

Claims

1. a discrimination means for discriminating a signal from each of the one or more transmission sources from a mixed signal in which signals transmitted from the one or more transmission sources are mixed, according to periodicity of the signals; a learning means for performing machine learning using training data having periodicity so that the discrimination means outputs information indicating the periodicity of the signal; Equipped with The learning means performs machine learning using a first loss function that becomes smaller when the signal identified by the identification means exhibits periodicity, so that the information output by the identification means exhibits periodicity.

1. An identification device comprising:

2. 2. The apparatus according to claim 1, wherein the learning means executes the machine learning and changes parameters of the classification means.

3. 3. The identification device according to claim 1, wherein the period of the training data is an unknown value.

4. Further comprising an extension means for extending the training data by applying a perturbation thereto; The learning means executes the machine learning so that a classification result of the training data after the expansion is in the same class as a classification result of the training data before the expansion.

4. The identification device according to claim 1, wherein the first and second electrodes are arranged in a first direction.

5. The identification device according to claim 4, characterized in that the learning means performs the machine learning using a second loss function that becomes small when the classification result of the training data after the expansion is in the same class as the classification result of the training data before the expansion.

6. 6. The identification device according to claim 1, wherein the identification means identifies the signal using a neural network.

7. 6. The apparatus according to claim 1, wherein the discrimination means discriminates the signal using non-negative matrix factorization.

8. 8. The identification device according to claim 1, wherein the signal is a radar wave or a sound wave.

9. 1. A learning method for an identification device for identifying a signal for each of one or more transmission sources from a mixed signal in which signals emitted from one or more transmission sources are mixed, according to periodicity of the signals, comprising: performing machine learning using training data having periodicity so that the discrimination device outputs information indicating the periodicity of the signal; Machine learning is performed using a first loss function that becomes smaller when the signal identified by the identification device exhibits periodicity, so that the information output by the identification device exhibits periodicity. A method for training an identification device comprising the steps of:

10. 1. A computer program used for training an identification device that identifies signals from one or more transmission sources according to periodicity of the signals from a mixed signal in which signals from the one or more transmission sources are mixed, the computer program comprising: performing machine learning using training data having periodicity so that the discrimination device outputs information indicating the periodicity of the signal; Machine learning is performed using a first loss function that becomes smaller when the signal identified by the identification device exhibits periodicity, so that the information output by the identification device exhibits periodicity. A computer program for causing a computer to operate in such a manner.

11. 2. The identification device according to claim 1, wherein the learning means uses an autocorrelation function as the first loss function.

12. The identification device according to claim 1 , wherein the learning means executes the machine learning using an entropy of the first loss function.

13. 6. The identification device according to claim 5, wherein the learning means uses an autocorrelation function as the second loss function.

14. The identification device according to claim 5 , wherein the learning means executes the machine learning using an entropy of the second loss function.

Images