Estimation device, estimation method, and program

Abstract

To provide an estimation device, an estimation method, and a program that can improve the accuracy of estimating the direction of arrival.SOLUTION: An estimation device according to the present disclosure includes acquisition means for acquiring an evaluation function indicating the characteristics of a determination target signal, which are obtained by inputting the determination target signal into a trained model, and the number of waves of the determination target signal, and estimation means for estimating the direction of arrival of the determination target signal using peak information in the evaluation function and the number of waves.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an estimation device, an estimation method, and a program. [Background technology]

[0002] In wireless communication, it is important to estimate the direction and position of a radio wave transmission source such as a communication terminal, and various techniques related to this have been proposed.

[0003] For example, Patent Document 1 discloses a radar system for identifying a preceding vehicle. The radar system includes an independent multi-beam antenna that outputs a received signal in response to an incoming wave, and a signal processing circuit in which a trained neural network is constructed. The signal processing circuit uses the neural network to output a signal indicating the number of incoming waves based on the received signal. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2016-142526 Summary of the Invention [Problem to be solved by the invention]

[0005] The MUSIC (Multiple Signal Classification) method, which is based on the subspace method using eigenvalue decomposition of the correlation matrix of received data, is a widely known method for estimating the direction of arrival of a radio wave source when an array antenna is used as a receiver. The MUSIC method performs a peak search on the MUSIC spectrum, which is an evaluation function of the calculation results, to obtain an estimated result of the direction of arrival.

[0006] However, in recent years, a new direction-of-arrival estimation method using deep learning has been proposed. In this direction-of-arrival estimation method, a peak search is performed on an evaluation function (spectrum) that is the result of inference by a neural network, thereby obtaining an estimated direction-of-arrival result. However, since the shape of the evaluation function is distorted due to the effects of noise, multipath fading, etc., there is a possibility that the number of peaks that appear in the evaluation function will be greater than the actual number of arriving waves. As a result, a problem may arise in which the direction of arrival is erroneously determined based on an incorrect peak that does not actually exist. The technology described in Patent Document 1 does not consider using learning in the evaluation function, and therefore does not solve this problem.

[0007] One of the objectives that the embodiments of the present disclosure aim to achieve is to provide an estimation device, an estimation method, and a program that can improve the accuracy of estimating the direction of arrival. It should be noted that this objective is only one of multiple objectives that the embodiments disclosed herein aim to achieve. Other objectives or problems and novel features will become apparent from the description of this specification or the accompanying drawings. [Means for solving the problem]

[0008] An estimation device according to one aspect includes: An acquisition means for acquiring an evaluation function indicating characteristics of the determination target signal and a wave number of the determination target signal, the evaluation function being obtained by inputting the determination target signal to a trained model; an estimation means for estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number; Equipped with.

[0009] An estimation method according to one aspect includes the steps of: An evaluation function indicating the characteristics of the determination target signal and the wave number of the determination target signal, which are obtained by inputting the determination target signal to a trained model, are obtained; estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number; This is what a computer does.

[0010] In one aspect, the program An evaluation function indicating the characteristics of the determination target signal and the wave number of the determination target signal, which are obtained by inputting the determination target signal to a trained model, are obtained; estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number; This is what causes a computer to execute the above. [Effects of the Invention]

[0011] According to the present disclosure, it is possible to provide an estimation device, an estimation method, and a program that can improve the accuracy of estimating the direction of arrival. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a block diagram illustrating an example of an estimation device according to the present disclosure. [Figure 2] FIG. 2 is a flowchart showing an example of a typical process performed by the estimation device. [Figure 3] FIG. 3 is a block diagram illustrating an example of an estimation device according to the present disclosure. [Figure 4] FIG. 4 is a block diagram illustrating an example of the inference unit. [Figure 5] FIG. 5 is a flowchart showing an example of the overall operation of the estimation device. [Figure 6] FIG. 6 is a flowchart illustrating an example of a model learning step of the estimation device. [Figure 7] FIG. 7 is a flowchart illustrating an example of the estimation steps of the estimation device. [Figure 8] FIG. 8 is a block diagram illustrating an example of an estimation device according to the present disclosure. [Figure 9] FIG. 9 is a block diagram illustrating an example of the inference unit. [Figure 10] FIG. 10 is a flowchart showing an example of the overall operation of the estimation device. [Figure 11] FIG. 11 is a flowchart illustrating an example of a model learning step of the estimation device. [Figure 12] FIG. 12 is a flowchart illustrating an example of the estimation steps of the estimation device. [Figure 13] FIG. 13 is a block diagram illustrating an example of an estimation device according to the present disclosure. [Figure 14] FIG. 14 is a block diagram illustrating an example of the inference unit. [Figure 15] FIG. 15 is a flowchart showing an example of the overall operation of the estimation device. [Figure 16] FIG. 16 is a flowchart illustrating an example of a model learning step of the estimation device. [Figure 17] FIG. 17 is a flowchart illustrating an example of estimation steps of the estimation device. [Figure 18] FIG. 18 is a block diagram illustrating an example of an estimation device and a learning device according to the present disclosure. [Figure 19A] FIG. 19A is a sequence diagram illustrating an example of the operations of the estimation device, the learning device, and the operation terminal according to the present disclosure. [Figure 19B] FIG. 19B is a sequence diagram illustrating an example of the operations of the estimation device, the learning device, and the operation terminal according to the present disclosure. [Figure 20] FIG. 20 is a block diagram showing an example of the hardware configuration of an information processing device that executes the processing of the system or device described in the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. Note that the following descriptions and drawings in the embodiments have been omitted or simplified as appropriate for clarity of explanation. Furthermore, in this disclosure, unless otherwise specified, when multiple items are defined as "at least one of multiple items," the definition may mean any one item, or any multiple items including all items. Each embodiment can be combined with other embodiments as appropriate.

[0014] Each drawing referenced in the embodiments is merely an example for describing one or more embodiments. Each drawing may not relate to only one particular embodiment, but may also relate to one or more other embodiments. As will be understood by those skilled in the art, various features or steps described with reference to any one drawing can be combined with features or steps shown in one or more other drawings to create, for example, an embodiment not explicitly shown or described. Not all features or steps shown in any one drawing are necessarily required to describe an exemplary embodiment, and some features or steps may be omitted. The order of steps described in any drawing may be changed as appropriate. Furthermore, components or processes described with the same reference numerals throughout multiple drawings indicate the same or corresponding components or processes.

[0015] Embodiment 1 [Configuration Description] 1 is a block diagram showing an example of an estimation device according to the present disclosure. The estimation device is a device that estimates the direction of arrival of a signal, and includes an acquisition unit 102 and an estimation unit 104. Each unit of the estimation device 10 will be described below.

[0016] The acquisition unit 102 acquires an evaluation function indicating characteristics of the determination target signal and the wave number of the determination target signal, which are obtained by inputting the determination target signal to the trained model. Here, a first trained model may output the evaluation function based on the determination target signal, and a second trained model different from the first trained model may output the wave number based on the determination target signal. However, the first trained model and the second trained model may be configured with some common components or may be configured as a common model in their entirety.

[0017] The evaluation function represents the spectrum of the signal to be evaluated. This spectrum includes peak information indicating peaks specific to the signal. The peak information includes, for example, at least one of information indicating the presence or absence of a peak and information indicating the magnitude of the peak. The magnitude of the peak may be represented by a peak value, which is a quantitative numerical value, or by qualitative information indicating the degree (for example, information indicating whether the peak is large / small).

[0018] The target signal is a signal that is the target of direction-of-arrival estimation. The target signal may be a signal obtained by a receiver in a real environment, or may be a virtually set signal.

[0019] The trained model is realized by a model that has been trained in advance before the target signal is input using a machine learning (e.g., deep learning) technique such as a neural network or a monotonic neural network. This model is trained by inputting training data that includes multiple sets of signals or feature quantities related to the signals and evaluation functions or wave numbers corresponding to the signals.

[0020] The trained model may be installed inside the estimation device 10, or may be installed in a device external to the estimation device 10. When the trained model is installed inside the estimation device 10, the acquisition unit 102 can acquire the evaluation function and wave number by inputting the target signal to the trained model. Alternatively, when a device external to the estimation device 10 acquires the evaluation function and wave number using the trained model, the estimation device 10 may acquire the evaluation function and wave number from that device.

[0021] The estimation unit 104 estimates the arrival direction of the target signal using the peak information and the wave number in the evaluation function acquired by the acquisition unit 102. The estimation unit 104 may, for example, identify peaks present in the evaluation function, the number of which corresponds to the wave number. The number corresponding to the wave number may be the value of the wave number, or may be a value obtained by increasing or decreasing the value of the wave number by a predetermined value. Alternatively, the estimation unit 104 may identify the peak values ​​or magnitudes of the evaluation function indicated by the peak information in descending order, the number of which corresponds to the wave number. That is, the estimation unit 104 may identify the number of peaks corresponding to the wave number by sorting the peaks indicated by the peak information in descending order of their magnitude. As another example, the estimation unit 104 may identify peaks whose peak values ​​or magnitudes indicated by the peak information are equal to or greater than a predetermined standard, the number of which corresponds to the wave number. The estimation unit 104 estimates one or more arrival directions corresponding to the identified peaks as the arrival direction of the target signal.

[0022] The information on the arrival direction estimated by the estimation unit 104 may be, for example, information on the arrival angle, or information on the direction indicating a predetermined range of the arrival angle (for example, 0° to 30°, 90° to 105°, etc.).

[0023] [Flow description] 2 is a flowchart showing an example of a typical process of the estimation device 10. This flowchart explains the process of the estimation device 10. Note that the details of each process are as described above, and therefore will not be explained again.

[0024] First, the acquisition unit 102 acquires an evaluation function indicating the characteristics of the target signal, which is obtained by inputting the target signal to a trained model, and the wave number of the target signal (step S12: acquisition step).The estimation unit 104 then estimates the direction of arrival of the target signal using peak information in the evaluation function and the wave number (step S14: estimation step).

[0025] [Effect description] As described above, the estimation unit 104 estimates the direction of arrival of the target signal by using not only the evaluation function obtained by the trained model but also the wave number. Therefore, even if the evaluation function includes an erroneous peak that does not actually exist, the estimation unit 104 can prevent erroneous estimation of the direction of arrival as the direction of arrival corresponding to the erroneous peak by further using wave number information in the estimation. Therefore, the estimation device 10 can improve the accuracy of estimating the direction of arrival.

[0026] The estimation device 10 may be configured as a single computer device or as a distributed system having multiple computer devices. In a distributed system, the processing performed by the estimation device 10 can be shared and executed by multiple computer devices. In other words, the acquisition unit 102 and the estimation unit 104 may be distributed and installed on two or more computer devices.

[0027] In each of the following embodiments, a specific example of the estimation device 10 described in embodiment 1 will be disclosed. However, the specific example of the estimation device 10 described in embodiment 1 is not limited to the one shown below. Furthermore, the configurations and processes described below are merely examples, and are not limited to these.

[0028] Embodiment 2 [Configuration Description] 3 is a block diagram showing an example of an estimation device according to the present disclosure. The estimation device 20 includes a receiving unit 202, a matrix generating unit 204, a teacher data storage unit 206, a learning unit 208, an inference unit 210, and a determination unit 212.

[0029] The learning unit 208 has a first learning unit 222 and a second learning unit 224, the inference unit 210 has a spectrum generation unit 226 and a wave number estimation unit 228 (described later), and the determination unit 212 has a peak search unit 262 and an angle-of-arrival determination unit 264. Each unit of the estimation device 20 is controlled by a hardware controller (not shown).

[0030] The receiving unit 202 is connected to an array antenna including a plurality of antenna elements as its components. The receiving unit 202 receives signals (received signals) wirelessly transmitted from a transmitting device (not shown) via the array antenna. There are at least two types of received signals: training data signals used for model learning, and target signals for determining the angle of arrival (hereinafter simply referred to as target signals). The receiving unit 202 outputs the received signals to the matrix generating unit 204.

[0031] The matrix generation unit 204 calculates the correlation matrix of the received signal output from the receiving unit 202 as a feature corresponding to the received signal. When the matrix generation unit 204 calculates the correlation matrix of the training data signal, it outputs the calculated correlation matrix to the training data storage unit 206. When the matrix generation unit 204 calculates the correlation matrix of the target signal, it outputs the calculated correlation matrix to the inference unit 210.

[0032] The training data storage unit 206 is a database that stores, as training data, a lower triangular matrix extracted from the correlation matrix output from the matrix generation unit 204, together with arrival angle information corresponding to the training data signal related to the correlation matrix. The lower triangular matrix corresponding to the training data signal is associated with information on the arrival angle of one or more waves in the training data signal and stored in the training data storage unit 206. If the training data signal is a signal from a plurality of different arrival angles, the lower triangular matrix corresponding to the received signal is associated with information on the plurality of different arrival angles and stored in the training data storage unit 206. The arrival angle information may be input by, for example, a user of the estimation device 20 or may be input from a device external to the estimation device 20. The stored information can be referenced by the learning unit 208.

[0033] The learning unit 208 generates a neural network, which is a trained model trained using a deep learning method, in the model training stage. The learning unit 208 trains the neural network so that, when a correlation matrix related to a signal is input, an evaluation spectrum indicating the characteristics of the signal is output as a node output from the output layer of the neural network.

[0034] In detail, the first learning unit 222 generates a first trained model for outputting an evaluation function (evaluation spectrum) based on a lower triangular matrix of the training data stored in the training data storage unit 206. The first learning unit 222 trains the first trained model by inputting training data including multiple sets of a lower triangular matrix and an evaluation function corresponding to the lower triangular matrix into the first learning model.

[0035] Furthermore, the second learning unit 224 generates a second trained model for outputting wavenumber values ​​based on a lower triangular matrix of the training data stored in the training data storage unit 206. The second learning unit 224 trains the second trained model by inputting training data including multiple sets of a lower triangular matrix and the wavenumber corresponding to that lower triangular matrix into the second training model. The trained models generated by the first learning unit 222 and the second learning unit 224 are two different and independent neural network models. The generated first trained model and second trained model are stored in the inference unit 210.

[0036] The inference unit 210 inputs the correlation matrix calculated by the matrix generation unit 204 to the neural network, which is a trained model. By inputting the correlation matrix calculated by the matrix generation unit 204 to the neural network, the inference unit 210 acquires information on the evaluation function and wave number corresponding to the received signal as an inference result.

[0037] FIG. 4 is a block diagram showing an example of the inference unit 210. The inference unit 210 infers information necessary to estimate the arrival direction of a target signal by using a trained model installed in the inference unit 210. In detail, the spectrum generation unit 226 executes inference by inputting the correlation matrix calculated by the matrix generation unit 204 to the first trained model generated by the first learning unit 222. The spectrum generation unit 226 acquires the evaluation function output by the first trained model. The wave number estimation unit 228 executes inference by inputting the correlation matrix calculated by the matrix generation unit 204 to the second trained model generated by the second learning unit 224. The wave number estimation unit 228 acquires the wave numbers output by the second trained model. The spectrum generation unit 226 and the wave number estimation unit 228 output the acquired evaluation function and wave numbers to the determination unit 212, respectively.

[0038] The first trained model may be configured as, for example, a classification model, and the second trained model may be configured as, for example, a regression model. The wave number estimation unit 228 may use machine learning (e.g., deep learning) as described above to estimate the number of incoming waves, or may use a non-learning method. Examples of non-learning methods include AIC (Akaike Information Criteria) and MDL (Minimum Description Length) based on maximum likelihood methods, but the methods that can be used are not limited to these.

[0039] As shown in FIG. 4, the first trained model has a feature extraction unit 242 and a fully connected layer 244. The feature extraction unit 242 is a part that extracts features of the input correlation matrix, and may include layers such as a convolutional layer that extracts features and a pooling layer that handles discrepancies in input information, as appropriate. The feature extraction unit 242 also uses functions such as an activation function and a normalization function as appropriate. The fully connected layer 244 is a layer that combines the features extracted by the feature extraction unit 242 and outputs an evaluation function.

[0040] 4, the second trained model has a feature extraction unit 246 and a fully connected layer 248. The feature extraction unit 246 is a part that extracts features of the input correlation matrix, and the detailed components thereof are the same as those of the feature extraction unit 242. The feature extraction unit 246 may be composed of components different from those of the feature extraction unit 242, or may be composed of the same components as those of the feature extraction unit 242. The fully connected layer 248 is a layer that combines the features extracted by the feature extraction unit 242 and outputs the number of incoming waves.

[0041] However, the configurations of the first trained model and the second trained model are not limited to those described above. For example, the first trained model may be composed of only the fully connected layer 244. The same applies to the second trained model.

[0042] The determining unit 212 determines the estimated arrival angle of the target signal. Specifically, the peak search unit 262 determines whether the number of arriving waves output by the wave number estimating unit 228 is a value equal to or greater than 1, or is zero.

[0043] If the number of arriving waves is one or more, the peak search unit 262 performs a peak search on the evaluation function output by the spectrum generation unit 226 and detects peaks present in the evaluation function. The arrival angle determination unit 264 sorts the peak values ​​of the detected peaks in descending order. Then, the arrival angle determination unit 264 selects peaks in descending order of peak value, the number of which is the same as the number of arriving waves output by the wave number estimation unit 228. The arrival angle determination unit 264 outputs the angle indicated by the selected peak as the estimated arrival angle.

[0044] If the number of arriving waves is 0, peak search unit 262 does not perform a peak search on the evaluation function output by spectrum generation unit 226. In other words, arrival angle determination unit 264 does not determine the arrival angle of the target signal. In this case, peak search unit 262 performs a peak search on the evaluation function when the received signal changes and the number of arriving waves output by wave number estimation unit 228 becomes a value of 1 or more. Arrival angle determination unit 264 determines the arrival angle using the result of the peak search.

[0045] The arrival angle determination unit 264 may output the estimated arrival angle to a display unit connected to the estimation device 20, thereby visualizing information about the arrival angle for the user. Even when the arrival angle of the determination target signal is not determined, the arrival angle determination unit 264 can output the information to the display unit. Alternatively, the arrival angle determination unit 264 may output the estimated arrival angle or information indicating that the arrival angle of the determination target signal is not determined to a device external to the estimation device 20.

[0046] [Flow description] Fig. 5 is a flowchart showing an example of the overall operation of the estimation device 20. Fig. 6 is a flowchart showing an example of the model learning steps of the estimation device 20. Fig. 7 is a flowchart showing an example of the estimation steps of the estimation device 20. An example of the operation of the above-mentioned estimation device 20 will be described below with reference to the flows of Figs. 5 to 7.

[0047] 5, when the received signal is a training data signal, the estimation device 20 performs model learning to generate a first trained model and a second trained model (step S22). When the received signal is a target signal, the estimation device 20 performs inference processing using the generated trained models to obtain an evaluation function and a wave number. Then, the estimation device 20 estimates the angle of arrival using the evaluation function and the wave number (step S24).

[0048] FIG. 6 is a flowchart illustrating an example of the operation of step S22 in detail. First, the receiving unit 202 acquires a received signal as a training data signal (step S222). The matrix generating unit 204 calculates a correlation matrix using data from the received signal (step S224). The training data storage unit 206 stores, as training data, a lower triangular matrix extracted from the correlation matrix and angle-of-arrival information corresponding to the training data signal related to the correlation matrix (step S226). The first learning unit 222 generates a first trained model for inferring an evaluation function using the training data (step S228). Furthermore, the second learning unit 224 generates a second trained model for inferring a wave number using the training data (step S230). Note that the processing of step S228 and the processing of step S230 may be performed in any order, or both processes may be performed in parallel. The above-described processing is performed as a preliminary preparation before receiving a target signal.

[0049] FIG. 7 is a flowchart illustrating an example of the operation of step S24 in detail. First, the receiving unit 202 acquires a received signal as a target signal (step S242). The matrix generating unit 204 calculates a correlation matrix using data of the received signal (step S244). The spectrum generating unit 226 generates an evaluation function by performing inference using the first trained model (step S246). The wave number estimating unit 228 generates a value of the number of incoming waves by performing inference using the second trained model (step S248). Note that the processing of step S246 and the processing of step S248 may be performed in any order, or both processes may be performed in parallel.

[0050] The peak search unit 262 determines whether the number of incoming waves output by the wave number estimation unit 228 is 1 or greater, or 0 (step S250). If the number of incoming waves is 1 or greater (Yes in step S250), the peak search unit 262 executes a peak search on the evaluation function (step S252). The arrival angle determination unit 264 determines the estimated arrival angle using the result of the peak search (step S254). The arrival angle determination unit 264 outputs the determined arrival angle.

[0051] If the number of incoming waves is 0 (No in step S250), the peak search unit 262 does not perform a peak search. In this case, the estimation device 20 returns to step S242 and executes the process.

[0052] As long as the receiver 202 continues to receive the comparison signal, each unit of the estimation device 20 executes steps S242 to S254 in a loop. This allows the estimation device 20 to change the estimated arrival angle in accordance with the change in the comparison signal, even if the comparison signal received by the receiver 202 changes.

[0053] [Effect description] When using deep learning to estimate the angle of arrival of a received signal in a real environment, accurate data may not be obtained as the evaluation function, which is the inference result of the neural network. For example, if the number of peaks in the evaluation function is greater than the actual number of arriving waves, the device may erroneously determine the angle of arrival based on an incorrect peak that does not actually exist. Also, if a false peak appears in the evaluation function even when the number of arriving waves is zero, the device may erroneously determine the angle of arrival based on that peak.

[0054] In contrast, the estimation device 20 infers the evaluation function and the number of incoming waves of the received signal in the inference unit 210. The estimation device 20 then estimates the angle of arrival of the target signal using not only the evaluation function but also information on the number of incoming waves. Therefore, even in a situation where using only the evaluation function would result in an erroneous peak determination and therefore an erroneous angle of arrival, the estimation device 20 uses the number of incoming waves to suppress erroneous peak determination, and as a result, can suppress erroneous determination of the angle of arrival. Therefore, the estimation device 20 can improve the accuracy of estimating the direction of arrival even when there is little training data. Therefore, the estimation device 20 can be said to be a device suitable for applying deep learning methods to real environments.

[0055] Alternatively, arrival angle determination unit 264 may select peaks in the evaluation function the number of which corresponds to the wave number, and estimate the arrival angle (arrival direction) indicated by the selected peak as the arrival angle of the target signal. This allows arrival angle determination unit 264 to accurately detect the peaks of the evaluation function, and therefore accurately detect the arrival angle.

[0056] Furthermore, the inference unit 210 does not need to estimate the angle of arrival of the target signal when the wave number value is 0. This can reduce the possibility that the inference unit 210 will erroneously estimate the angle of arrival.

[0057] Embodiment 3 In the following third to fifth embodiments, variations of the estimation device 20 described in the second embodiment will be disclosed. However, the variations of the estimation device 20 are not limited to those described below. Furthermore, the configurations and processes described below are merely examples, and are not intended to be limiting.

[0058] [Explanation of the premise] In the second embodiment, the learning unit 208 generates two different independent models as inference models for the evaluation function and wave number, as shown in Fig. 4. In contrast, in the third embodiment, a method for generating a common inference model for the evaluation function and wave number by performing multitask learning will be described.

[0059] Multi-task learning is a technique for improving the accuracy of feature estimation for multiple tasks by having one model simultaneously learn multiple tasks when the tasks are related to each other. Applying multi-task learning reduces the probability that the model misses information that is useful as a feature, and also has the effect of suppressing overfitting. Therefore, in a third embodiment, the configuration and operation required when applying multi-task learning to an estimation device according to the present disclosure will be described.

[0060] [Configuration Description] Fig. 8 is a block diagram showing an example of an estimation device according to the present disclosure. Compared to the estimation device 20 shown in Fig. 3, the estimation device 30 has the learning unit 208 and the inference unit 210 changed to a learning unit 302 and an inference unit 304, respectively. Each unit of the estimation device 30 is controlled by a hardware controller (not shown). Hereinafter, the points already explained in the actual embodiment 2 will be omitted as appropriate, and the components and processing unique to the embodiment 3 will be particularly explained.

[0061] The learning unit 302 generates a trained model using a deep learning technique in the model training stage. Specifically, the learning unit 302 generates a trained model for outputting an evaluation function (evaluation spectrum) and wavenumbers based on a lower triangular matrix of the training data stored in the training data storage unit 206. The learning unit 302 trains the trained model by inputting training data including multiple sets of a lower triangular matrix and an evaluation function and wavenumber corresponding to the lower triangular matrix into the model.

[0062] 9 is a block diagram showing an example of the inference unit 304. The inference unit 304 infers information necessary to estimate the arrival direction of a target signal by using a trained model installed in the inference unit 304. In detail, the inference unit 304 executes inference by inputting the correlation matrix calculated by the matrix generation unit 204 into the trained model generated by the learning unit 302. The trained model outputs an evaluation function and a wave number as a result of executing the inference. The inference unit 304 outputs this evaluation function and wave number to the determination unit 212.

[0063] Here, the trained model shown in FIG. 9 includes a feature extraction unit 282 and fully connected layers 284 and 286. In FIG. 4, the first trained model and the second trained model each include a different feature extraction unit and fully connected layer. In contrast, in the trained model shown in FIG. 9, the feature extraction unit 282 that extracts features of an input correlation matrix is ​​provided in a common manner for deriving an evaluation function and wavenumber. The feature extraction unit 282 may include layers such as a convolutional layer that extracts features and a pooling layer that accommodates deviations in input information, as appropriate. Furthermore, the feature extraction unit 282 uses functions such as an activation function and a normalization function as appropriate.

[0064] 9, the trained model includes a fully connected layer for each task. That is, the fully connected layer 284 outputs an evaluation function (evaluation spectrum) by combining the features extracted by the feature extraction unit 282, and the fully connected layer 286 outputs the number of incoming waves by combining the features extracted by the feature extraction unit 282.

[0065] In this way, by applying multi-task learning, a single trained model with one input and two outputs is generated, and the learning unit 302 and the inference unit 304 are each configured as a single component.

[0066] [Flow description] Fig. 10 is a flowchart showing an example of the overall operation of the estimation device 30. Fig. 11 is a flowchart showing an example of the model learning steps of the estimation device 30. Fig. 12 is a flowchart showing an example of the estimation steps of the estimation device 30. An example of the operation of the above-mentioned estimation device 30 will be described below with reference to the flows of Figs. 10 to 12. Note that a description of the same processes as those of the estimation device 20 will be omitted.

[0067] 10, when the received signal is a training data signal, the estimation device 30 executes model learning to generate a one-input, two-output trained model (step S32). When the received signal is a target signal, the estimation device 20 executes inference processing using the generated trained model to obtain an evaluation function and a wave number. Then, the estimation device 20 estimates the angle of arrival using the evaluation function and the wave number (step S34).

[0068] Fig. 11 is a flowchart for explaining in detail an example of the operation of step S32. The processes of steps S222 to S226 in Fig. 11 are the same as the processes of steps S222 to S226 in Fig. 6, respectively, and therefore will not be described again. The learning unit 302 uses training data to generate a trained model for inferring an evaluation function and wave number (step S322). The above-described processes are performed as advance preparation before receiving a target signal.

[0069] Fig. 12 is a flowchart illustrating an example of the operation of step S34 in detail. The processes of steps S242 to S244 in Fig. 12 are the same as the processes of steps S242 to S244 in Fig. 7, respectively, and therefore will not be described again. The inference unit 304 generates an evaluation function and a wave number by performing inference using the trained model (step S342). The processes of steps S250 to S254 in Fig. 12, which are processes executed using the generated evaluation function and wave number, are the same as the processes of steps S250 to S254 in Fig. 7, respectively, and therefore will not be described again.

[0070] [Effect description] As described above, the estimation device 30 can generate a one-input, two-output trained model for the evaluation function and wave number by applying multi-task learning to the neural network. As a result, in addition to the effects described in the second embodiment, the estimation device 30 can further improve the estimation accuracy of the direction of arrival and improve the generalizability of the device. As a secondary effect, the feature extraction unit of the trained model can be used in common for the evaluation function and wave number, which also has the effect of reducing the size of the trained model.

[0071] Embodiment 4 [Explanation of the premise] In the second embodiment, inference is performed using one model as the first trained model for all the expected numbers of incoming waves in the determination target signal (i.e., the number of antenna elements minus 1). In this case, as the number of antenna elements increases, the expected number of incoming waves also increases, and it was necessary to increase the number of parameters of the first trained model to accommodate the number of incoming waves. As a result, there is a potential problem that the inference time of the first trained model increases.

[0072] In order to solve this problem, in the fourth embodiment, a plurality of first trained models used to generate an evaluation function are provided as models specialized for each number of incoming waves, and a process of switching one first trained model to be used is executed based on the estimated number of incoming waves.

[0073] [Configuration Description] Fig. 13 is a block diagram showing an example of an estimation device according to the present disclosure. Compared to the estimation device 20 shown in Fig. 3, the estimation device 40 has the first learning unit 222 changed to a first learning unit 402, and the inference unit 210 changed to an inference unit 410. Each unit of the estimation device 40 is controlled by a hardware controller (not shown). Hereinafter, the points already explained in the actual embodiment 2 will be omitted as appropriate, and the components and processing unique to the embodiment 4 will be particularly explained.

[0074] In the model learning stage, the first learning unit 402 generates a first trained model for outputting an evaluation function (evaluation spectrum) based on the lower triangular matrix of the training data stored in the training data storage unit 206. At this time, the first learning unit 402 acquires a wave number related to the training data and determines which wave number, among all the expected numbers of incoming waves, the wave number corresponds to.

[0075] In this example, as learning models, models corresponding to the number of expected incoming waves (for example, the total number of expected incoming waves) are provided. The first learning unit 402 selects a learning model corresponding to the determined number of waves and supplies training data to that model, thereby causing that model to learn. Details of the learning method are as described in the second embodiment, and therefore will not be described here.

[0076] 14 is a block diagram showing an example of the inference unit 410. The inference unit 410 has a wave number estimation unit 228, a model selection unit 412, and a spectrum generation unit 414. Here, the wave number estimation unit 228 infers a wave number using the second trained model. Details of this inference process are as explained in the second embodiment, and will not be explained again. The wave number estimation unit 228 outputs the inferred wave number to the model selection unit 412.

[0077] The model selection unit 412 acquires the inferred wave number and determines which wave number, out of all the expected numbers of incoming waves, the wave number corresponds to. As described above, the number of first trained models is equal to the number of all the expected numbers of incoming waves. The model selection unit 412 selects the first trained model that corresponds to the determined wave number from among the multiple first trained models. The model selection unit 412 outputs information on the selected model to the spectrum generation unit 414.

[0078] The spectrum generation unit 414 executes inference by inputting the correlation matrix calculated by the matrix generation unit 204 to the first trained model indicated by the information output by the model selection unit 412. Details of this inference process are as described above. As a result, the spectrum generation unit 414 acquires the evaluation function output by the selected first trained model. The spectrum generation unit 414 and the wave number estimation unit 228 output the acquired evaluation function and wave number to the determination unit 212, respectively.

[0079] The number of training models provided does not have to be the same as the number of all expected numbers of incoming waves. For example, multiple wave number types may be set based on all expected numbers of incoming waves or a range of the number of incoming waves, and training models may be provided in the number corresponding to the types. In such a case, the first training unit 402 selects a training model corresponding to the type corresponding to the wave number that is the determination result. Furthermore, the model selection unit 412 selects a first trained model corresponding to the type corresponding to the wave number that is the determination result. Other processing by the first training unit 402 and the model selection unit 412 is the same as described above.

[0080] [Flow description] Fig. 15 is a flowchart showing an example of the overall operation of the estimation device 40. Fig. 16 is a flowchart showing an example of the model learning steps of the estimation device 40. Fig. 17 is a flowchart showing an example of the estimation steps of the estimation device 40. An example of the operation of the above-mentioned estimation device 40 will be described below with reference to the flows of Figs. 15 to 17.

[0081] First, in FIG. 15, when the received signal is a training data signal, the estimation device 40 generates a first trained model and a second trained model by performing model learning (step S42). Here, the first trained models are generated for all expected numbers of incoming waves. When the received signal is a target signal, the estimation device 20 performs inference processing using the generated trained models to obtain an evaluation function and a wave number. Then, the estimation device 20 estimates the angle of arrival using the evaluation function and the wave number (step S44).

[0082] FIG. 16 is a flowchart illustrating an example of the operation of step S42 in detail. The processes of steps S222 to S226 in FIG. 16 are the same as the processes of steps S222 to S226 in FIG. 6, respectively, and therefore will not be described again. The first learning unit 222 uses the training data to generate a plurality of first trained models for inferring an evaluation function for each number of incoming waves (step S422). The process of step S230 in FIG. 16 is the same as the process of step S230 in FIG. 6, and therefore will not be described again. Note that either the process of step S422 or the process of step S230 may be executed first, or both processes may be executed in parallel. The processes described above are performed as advance preparation before receiving a target signal.

[0083] Fig. 17 is a flowchart for explaining in detail an example of the operation of step S24. The processes of steps S242 to S244 in Fig. 17 are the same as the processes of steps S242 to S244 in Fig. 7, respectively, and therefore will not be described again.

[0084] The wave number estimation unit 228 generates a wave number value by performing inference using the second trained model (step S442). The model selection unit 412 switches the first trained model to be used by selecting the first trained model corresponding to the wave number estimated in step S442 (step S444). The model selection unit 412 outputs information on the selected model to the spectrum generation unit 414. The spectrum generation unit 414 generates an evaluation function by performing inference using the selected first trained model (step S446). The processes of steps S250 to S254 in FIG. 17 are the same as the processes of steps S250 to S254 in FIG. 7, and therefore their explanations are omitted. As long as the receiving unit 202 continues to receive the comparison signal, each unit of the estimation device 40 executes steps S242 to S254 in a loop.

[0085] [Effect description] As described above, the estimation device 40 generates multiple first trained models according to the number of waves in the training data and switches the first trained model to be used according to the number of waves in the received signal. Because the estimation device 40 does not need to always use a single first trained model, an increase in the number of parameters of the first trained model can be suppressed. Therefore, in addition to the effect described in the second embodiment, the estimation device 40 has the effect of suppressing an increase in the inference time of the first trained model per inference process.

[0086] Fifth embodiment [Explanation of the premise] In the fifth embodiment, a configuration example will be described in which the function for generating a learning model in the estimation device 20 shown in the second embodiment is separated from the estimation device 20 and provided as a separate device (learning device). In this configuration example, the learning device is installed in one location (for example, on the cloud side), and the estimation device is installed in a location remote from the learning device. For convenience, the estimation device will be described as a single device in the following description, but multiple estimation devices may be provided corresponding to one learning device. The multiple estimation devices are installed at multiple locations where direction-of-arrival estimation is performed, respectively. In addition, a console operated by an operator is connected to the learning device and the estimation device via a network. The operator can remotely control the learning device and the estimation device using the console.

[0087] [Configuration Description] Fig. 18 is a block diagram showing an example of an estimation device and a learning device according to the present disclosure. The estimation device 50 has the configuration of the estimation device 20 shown in Fig. 3, but the teacher data storage unit 206 and the learning unit 208 are separated from each other. The learning device 60 includes the teacher data storage unit 206 and the learning unit 208.

[0088] Furthermore, at least between the estimation device 50 and the learning device 60, the matrix generation unit 204 and the teacher data storage unit 206, and the learning unit 208 and the inference unit 210 are connected, thereby enabling transmission and reception of data necessary for processing. For example, the correlation matrix calculated by the matrix generation unit 204 is transmitted to the teacher data storage unit 206 of the learning device 60 via a network. After generating the first trained model and the second trained model, the learning unit 208 transmits the generated first trained model and the second trained model to the inference unit 210 of the estimation device 50 via a network. This allows the inference unit 210 to use the trained models.

[0089] However, the first trained model and the second trained model may be stored in the learning device 60 even after they are generated. In this case, the inference unit 210 transmits the correlation matrix calculated by the matrix generation unit 204 to the learning device 60. The first trained model and the second trained model perform inference using the transmitted correlation matrix and output an evaluation function and a wave number, respectively. The learning device 60 transmits the evaluation function and wave number to the inference unit 210. In this way, the estimation device 50 can obtain information on the evaluation function and wave number required for estimating the angle of arrival by using a trained model stored in another learning device 60.

[0090] Each unit of the estimation device 50 and the learning device 60 is controlled by a hardware controller (not shown) provided in each device. Details of the processes executed by each unit of the estimation device 50 and the learning device 60 are as described in the second embodiment, and therefore will not be described again.

[0091] [Flow description] 19A and 19B are sequence diagrams showing an example of the operation of the estimation device, learning device, and operation terminal according to the present disclosure. The operation terminal 70 is provided with an operation console that is operated by an operator to control the estimation device 50 and the learning device 60. The operation shown in FIG. 19A shows a series of operations executed in step S22 (model learning processing) of FIG. 5. The operation shown in FIG. 19B shows a series of operations executed in step S24 (arrival angle estimation processing) of FIG. 5. Below, an example of the operation of each device will be described with reference to FIGS. 19A and 19B.

[0092] (1) First, in FIG. 19A , an operator operates the operation terminal 70 to output an instruction to the estimation device 50 and the learning device 60 to start collecting teacher data. (2) After receiving the instruction, the receiving unit 202 of the estimation device 50 acquires the received signal as a teacher data signal in accordance with the instruction. (3) The matrix generating unit 204 calculates a correlation matrix using the data of the received signal. (4) The matrix generating unit 204 transfers the lower triangular matrix of the calculated correlation matrix to the learning device 60 via the network. (5) The learning device 60 receives the transferred lower triangular matrix and stores it as teacher data in the teacher data storage unit 206 together with angle-of-arrival information corresponding to the teacher data signal related to the correlation matrix.

[0093] 19A correspond to the processes of step S222, step S224, and step S226 in FIG. 6, respectively. When the learning device 60 receives an instruction related to (1) from the operation terminal 70, the learning device 60 appropriately performs preparations for receiving matrix data from the estimation device 50.

[0094] (6) Next, after storing the teacher data in the teacher data storage unit 206, the learning device 60 transmits a report indicating that collection of the teacher data has been completed to the operation terminal 70. The operation terminal 70 displays the report on the operation console. (7) The operator outputs an instruction to the learning device 60 to generate a trained model through learning according to the displayed content. (8) After receiving the instruction, the learning unit 208 of the learning device 60 (more specifically, the first learning unit 222 and the second learning unit 224) generates a first trained model and a second trained model using the teacher data. (9) After generating the first trained model and the second trained model, the learning device 60 transmits a report indicating that learning has been completed and that the trained models have been generated to the operation terminal 70. The display of this report on the operation console allows the operator to recognize that learning has been completed. (10) Furthermore, the learning device 60 transfers the generated first trained model and second trained model to the estimation device 50 via the network.

[0095] Note that the process (8) in Figure 19A corresponds to the processes of steps S228 and S230 in Figure 6. Also, either the process (9) or the process (10) in Figure 19A may be executed first, or both processes may be executed in parallel. Furthermore, the learning device 60 may automatically execute the process (8) even when it does not receive an instruction to execute generation of a trained model.

[0096] (11) In FIG. 19B, an operator operates the operation terminal 70 to output an instruction to the estimation device 50 to execute inference processing regarding the arrival angle of the target signal. (12) After receiving the instruction, the receiving unit 202 of the estimation device 50 acquires the received signal as the target signal in accordance with the instruction. (13) The matrix generation unit 204 calculates a correlation matrix using data on the received signal. (14) The inference unit 210 generates an evaluation function and a wave number by performing inference using the first trained model and the second trained model. (15) The determination unit 212 performs a peak search on the evaluation function based on the generated value of the number of arriving waves, and determines the estimated arrival angle using the peak search result. (16) The determination unit 212 outputs information on the determined arrival angle to the operation terminal 70 via the network. This arrival angle information is displayed on the console, allowing the operator to grasp the arrival angle information.

[0097] The processes of (12) and (13) in Fig. 19B correspond to the processes of steps S242 to S244 in Fig. 7. The process of (14) in Fig. 19B corresponds to the processes of steps S246 to S248 in Fig. 7. Furthermore, the process of (15) in Fig. 19B corresponds to the processes of steps S250 to S254 in Fig. 7.

[0098] [Effect description] As described above, the configuration for executing the learning process is separated from the estimation device 50 and provided as the learning device 60, and even when the estimation device 50 and the learning device 60 are installed separately, the processing described in embodiment 2 can be executed.

[0099] In a situation where estimation devices 50 are provided at multiple locations where direction-of-arrival estimation is performed for one learning device 60, multiple first trained models may be provided and a process of switching between the single first trained model to be used may be implemented, as shown in embodiment 4. In this case, the total time required for inference in the estimation device 50 can be shortened as the number of estimation devices 50 increases, compared to when a single first trained model is generated and used. Therefore, the effects of embodiment 4 can be further obtained.

[0100] The estimation devices 20 to 50 shown in the second to fifth embodiments and the learning device 60 shown in the fifth embodiment may be configured as a single computer device, similar to the estimation device 10 shown in the first embodiment. Alternatively, these devices may also be configured as a distributed system having a plurality of computer devices.

[0101] The estimation device and learning device according to the present disclosure can be used for applications such as radio wave monitoring, cognitive radio, dynamic frequency sharing, and police and security services, but the uses of the devices are not limited to these.

[0102] In the above-described embodiments, the present disclosure has been described as a hardware configuration, but the present disclosure is not limited to this. The present disclosure can also be realized by causing a processor in a computer to execute a computer program to perform the processing of each device constituting the estimation devices 10 to 50 and the learning device 60 described in the above-described embodiments.

[0103] 20 is a block diagram showing an example of the hardware configuration of an information processing device (in other words, a computer) on which the processing of the system or device described in the present disclosure is executed. Referring to FIG. 20, the information processing device 90 includes a signal processing circuit 91, a processor 92, and a memory 93.

[0104] The signal processing circuit 91 is a circuit for processing signals in accordance with the control of the processor 92. The signal processing circuit 91 may include a communication circuit for receiving signals from a transmitting device.

[0105] The processor 92 is connected to the memory 93, and performs the processing of the system described in the above embodiment by reading and executing a computer program from the memory 93. As an example of the processor 92, one of a CPU (Central Processing Unit), an MPU (Micro Processing Unit), an FPGA (Field-Programmable Gate Array), a DSP (Demand-Side Platform), and an ASIC (Application Specific Integrated Circuit) may be used, or a plurality of these may be used in parallel.

[0106] The memory 93 is configured with a volatile memory, a nonvolatile memory, or a combination thereof. The memory 93 is not limited to one, and multiple memories may be provided. The volatile memory may be, for example, a RAM (Random Access Memory) such as a DRAM (Dynamic Random Access Memory) or an SRAM (Static Random Access Memory). The nonvolatile memory may be, for example, a ROM (Read Only Memory) such as a PROM (Programmable Random Only Memory) or an EPROM (Erasable Programmable Read Only Memory), a flash memory, or an SSD (Solid State Drive).

[0107] The memory 93 is used to store one or more instructions. Here, the one or more instructions are stored as programs in the memory 93. The processor 92 can perform the processes described in the above embodiments by reading and executing these programs from the memory 93.

[0108] The memory 93 may include memory built into the processor 92 in addition to memory provided outside the processor 92. The memory 93 may also include storage located away from the processors constituting the processor 92. In this case, the processor 92 can access the memory 93 via an I / O (Input / Output) interface.

[0109] As described above, one or more processors included in each system in the above-described embodiments execute one or more programs including instructions for causing a computer to execute the algorithms described using the drawings. Execution of the programs enables the information processing described in each embodiment to be realized.

[0110] The program includes instructions or software code that, when loaded into a computer, causes the computer to perform one or more functions described in the embodiments. The program may be stored in a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, computer-readable media or tangible storage media include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disk (DVD), Blu-ray disc or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices. The program may also be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, transitory computer-readable media or communication media include electrical, optical, acoustic, or other forms of propagated signals. The transitory computer-readable medium or communication medium may provide the program to the computer via a wired communication path, such as electrical wires and optical fibers, or via a wireless communication path.

[0111] A part or all of the above-described embodiments can be described as, but not limited to, the following supplementary notes. (Appendix 1) An acquisition means for acquiring an evaluation function indicating characteristics of the determination target signal and a wave number of the determination target signal, the evaluation function being obtained by inputting the determination target signal to a trained model; and an estimation means for estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number. Estimation device. (Appendix 2) the estimation means selects peaks in the evaluation function in a number corresponding to the wave number, and estimates the arrival direction indicated by the selected peaks as the arrival direction of the target signal. 10. The estimation apparatus of claim 1. (Appendix 3) The estimation means does not estimate the direction of arrival of the target signal when the value of the wave number is 0. 3. The estimation device according to claim 1 or 2. (Appendix 4) The estimation device includes: Further provided is a generation unit that generates the trained model having a feature extraction unit that is common to the evaluation function and the wave number, a first fully connected layer that outputs the evaluation function, and a second fully connected layer that outputs the wave number. 4. The estimation device according to any one of appendixes 1 to 3. (Appendix 5) The estimation device includes: A generation unit that generates the trained model that outputs the evaluation function for each of a plurality of wave numbers expected in the determination target signal, The acquisition means selects one of the plurality of trained models according to the wave number of the determination target signal, and acquires the evaluation function using the selected trained model. 4. The estimation device according to any one of appendixes 1 to 3. (Appendix 6) An evaluation function indicating the characteristics of the determination target signal and the wave number of the determination target signal, which are obtained by inputting the determination target signal to a trained model, are obtained; estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number; This is an estimation method performed by a computer. (Appendix 7) The computer selecting peaks in the evaluation function by the number corresponding to the wavenumber, and estimating the arrival direction indicated by the selected peaks as the arrival direction of the target signal; Estimation method described in Appendix 6. (Appendix 8) The computer When the value of the wave number is 0, the arrival direction of the target signal is not estimated. The estimation method described in Appendix 6 or 7. (Appendix 9) The computer generating the trained model having a feature extraction unit common to the evaluation function and the wave number, a first fully connected layer that outputs the evaluation function, and a second fully connected layer that outputs the wave number; 9. The estimation method according to any one of appendixes 6 to 8. (Appendix 10) An evaluation function indicating the characteristics of the determination target signal and the wave number of the determination target signal, which are obtained by inputting the determination target signal to a trained model, are obtained; estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number; A program that makes a computer do something.

[0112] Some or all of the elements (e.g., configurations and functions) described in Appendix 5 that are dependent on Appendix 1 may also be dependent on Appendix 6 in the same dependency relationship as Appendix 5. Furthermore, some or all of the elements (e.g., configurations and functions) described in Appendix 2 to Appendix 5 that are dependent on Appendix 1 may also be dependent on Appendix 10 in the same dependency relationship as Appendix 2 to Appendix 5. Some or all of the elements described in any appendix may be applied to various hardware, software, recording means for recording software, systems, and methods.

[0113] Although the present disclosure has been described above with reference to the embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate. [Explanation of symbols]

[0114] 10, 20, 30, 40, 50 estimator 60 Learning Device 102 Acquisition part 104 Estimation part 202 receiving unit 204 matrix generating unit 206 Teacher Data Storage Unit 208 Learning Unit 210 inference unit 212 decision unit 222 First Learning Section 224 Second Learning Section 226 Spectrum generation unit 228 Wave number estimation unit 242 Feature extraction part 244 Fully connected layer 246 Feature extraction part 248 Fully connected layer 262 Peak search unit 264 Arrival angle determination unit 282 Feature extraction section 284, 286 Fully connected layer 302 Learning section 304 Inference section 410 Inference unit 412 Model selection unit 414 Spectrum Generation Unit

Claims

1. An acquisition means for acquiring an evaluation function indicating characteristics of the determination target signal and a wave number of the determination target signal, the evaluation function being obtained by inputting the determination target signal to a trained model; and an estimation means for estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number. Estimation device.

2. the estimation means selects peaks in the evaluation function in a number corresponding to the wave number, and estimates the arrival direction indicated by the selected peaks as the arrival direction of the target signal. The estimation device according to claim 1 .

3. the estimation means does not estimate the direction of arrival of the target signal when the value of the wave number is 0; The estimation device according to claim 1 or 2.

4. The estimation device includes: Further provided is a generation unit that generates the trained model having a feature extraction unit that is common to the evaluation function and the wave number, a first fully connected layer that outputs the evaluation function, and a second fully connected layer that outputs the wave number. The estimation device according to claim 1 or 2.

5. The estimation device includes: A generation unit that generates the trained model that outputs the evaluation function for each of a plurality of wave numbers expected in the determination target signal, The acquisition means selects one of the plurality of trained models according to the wave number of the determination target signal, and acquires the evaluation function using the selected trained model. The estimation device according to claim 1 or 2.

6. An evaluation function indicating the characteristics of the determination target signal and the wave number of the determination target signal, which are obtained by inputting the determination target signal to a trained model, are obtained; estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number; This is an estimation method performed by a computer.

7. The computer selecting peaks in the evaluation function by the number corresponding to the wavenumber, and estimating the arrival direction indicated by the selected peaks as the arrival direction of the target signal; The estimation method according to claim 6.

8. The computer When the value of the wave number is 0, the arrival direction of the target signal is not estimated. The estimation method according to claim 6 or 7.

9. The computer generating the trained model having a feature extraction unit common to the evaluation function and the wave number, a first fully connected layer that outputs the evaluation function, and a second fully connected layer that outputs the wave number; The estimation method according to claim 6 or 7.

10. An evaluation function indicating the characteristics of the determination target signal and the wave number of the determination target signal, which are obtained by inputting the determination target signal to a trained model, are obtained; estimating the direction of arrival of the target signal using peak information in the evaluation function and the wave number; A program that makes a computer do something.

Images