Radio communication system, characteristic amount estimation device, radio communication method, and radio communication program
The wireless communication system enhances feature estimation accuracy by using learning models to combine individual feature estimations, addressing the limitations of existing methods in combining multiple feature quantities.
Patent Information
- Application Number
- JP2024023993
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-02-20
- Publication Date
- 2025-09-01
AI Technical Summary
Existing methods for estimating multiple feature quantities related to a structure from wireless signals fail to improve estimation accuracy by combining individual estimation results, limiting the precision of the overall assessment.
A wireless communication system and method utilizing a first and second learning model to derive differential information from propagation channel data, perform individual and combined estimation processes to enhance the accuracy of feature estimation.
The system effectively combines multiple feature quantities, improving the estimation accuracy of structural features through joint learning processes.
Smart Images
Figure 2025127322000001_ABST
Abstract
Description
[Technical Field]
[0001] The present disclosure relates to a technique for estimating characteristics of a structure from wireless signals exchanged between a transmitter and a receiver. [Background technology]
[0002] Wireless sensing technologies have been developed that estimate the external shape of structures based on the received signal strength indicator (RSSI) and signal-to-noise ratio (SNR) obtained from wireless signals exchanged between a transmitter and a receiver, or propagation channel information for each OFDM subcarrier. Furthermore, technologies for estimating the internal state of structures are also advancing (see, for example, non-patent documents 1 and 2). [Prior art documents] [Non-patent literature]
[0003] [Non-Patent Document 1] Y. Liu, L. Jiang, L. kong, Q. Xiang, X. Liu, and G. Chen, “WiFruit: See through fruits with smart devices”, ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol.5, no.169, pp.1- 29, 2021. [Non-patent document 2] Chen Wang, Jian Liu, Yingying Chen, Hongbo Liu and Yan Wang, “Towards In-baggage Suspicious Object Detection Using Commodity WiFi”, IEEE Conferene on Communications and Network Security, 30 May 2018 - 01 June 2018. Summary of the Invention [Problem to be solved by the invention]
[0004] Incidentally, when estimating multiple feature quantities from measured CSI, it is believed that the estimation accuracy will be improved by combining multiple feature quantities rather than estimating each feature quantity individually. This is something that we humans do intuitively, for example, by first grasping the size of a structure and then grasping the material, color, etc. of the details, we can obtain a more precise understanding of the structure's features.
[0005] However, the above-described method does not allow for further estimation after combining the estimation results of multiple feature quantities related to the structure, which poses a problem in that it is not possible to improve the estimation accuracy of multiple feature quantities related to the structure.
[0006] In order to solve the above-mentioned problems, a first object of the present disclosure is to provide a wireless communication system that can improve the estimation accuracy of multiple feature quantities related to a structure.
[0007] A second object of the present disclosure is to provide a feature estimation device that can improve the estimation accuracy of multiple feature quantities related to a structure.
[0008] A third object of the present disclosure is to provide a wireless communication method that can improve the accuracy of estimating multiple feature quantities related to a structure.
[0009] A fourth object of the present disclosure is to provide a wireless communication program that can improve the accuracy of estimating multiple feature quantities related to a structure. [Means for solving the problem]
[0010] A first aspect of the present disclosure is 1. A wireless communication system that estimates characteristics of a structure present on a propagation path of a wireless signal based on the wireless signal, comprising: a transmitter for transmitting the radio signal; a receiver for deriving propagation channel information from the radio signal; a feature estimation device that receives the propagation channel information from the receiver; Equipped with The feature estimation device includes: a first learning model and a second learning model; deriving differential information of the radio signal between the transmitter antenna and the receiver antenna from the propagation channel information; an individual estimation process of inputting the difference information into the first learning model and estimating a plurality of feature amounts; a combined estimation process of inputting the plurality of feature quantities into the second learning model and estimating a plurality of corrected feature quantities; configured to run In the joint estimation process, a process of estimating a first modified feature based on the feature; It is preferable that the method further includes a process of estimating a second modified feature by combining the feature and the first modified feature.
[0011] The second aspect is 1. A feature estimation device that estimates features of a structure present in a propagation path of a wireless signal, based on the wireless signal exchanged between a transmitter and a receiver, comprising: a first learning model and a second learning model; receiving propagation channel information from the receiver; deriving differential information of the radio signal between the transmitter antenna and the receiver antenna from the propagation channel information; an individual estimation process of inputting the difference information into the first learning model and estimating a plurality of feature amounts; a combined estimation process of inputting the plurality of feature quantities into the second learning model and estimating a plurality of corrected feature quantities; configured to run In the joint estimation process, a process of estimating a first modified feature based on the feature; It is preferable that the method further includes a process of estimating a second modified feature by combining the feature and the first modified feature.
[0012] The third aspect is 1. A wireless communication method for estimating characteristics of a structure present in a propagation path of a wireless signal, based on the wireless signal exchanged between a transmitter and a receiver, comprising: deriving propagation channel information from the wireless signal; deriving differential information of the radio signal between the transmitter antenna and the receiver antenna from the propagation channel information; executing an individual estimation process for inputting the difference information into a first learning model and estimating a plurality of feature quantities; inputting the plurality of feature quantities into a second learning model and executing a joint estimation process to estimate a plurality of corrected feature quantities; Including, In the joint estimation process, estimating a first modified feature based on the feature; It is preferable that the method further includes combining the feature amount and the first modified feature amount to estimate a second modified feature amount.
[0013] The fourth aspect is 1. A wireless communication program to be executed by a feature estimation device that estimates features of a structure present in a propagation path of a wireless signal, based on the wireless signal exchanged between a transmitter and a receiver, the program comprising: The feature estimation device includes: receiving propagation channel information from the receiver; deriving differential information of the radio signal between the transmitter antenna and the receiver antenna from the propagation channel information; an individual estimation process of inputting the difference information into a first learning model and estimating a plurality of feature quantities; a combined estimation process of inputting the plurality of feature quantities into a second learning model and estimating a plurality of corrected feature quantities; a program for executing In the joint estimation process, a process of estimating a first modified feature based on the feature; and estimating a second modified feature quantity by combining the feature quantity and the first modified feature quantity. [Effects of the Invention]
[0014] According to an aspect of the present disclosure, the second learning model combines multiple feature quantities estimated individually in the first learning model and performs further estimation, thereby improving the accuracy of estimating multiple feature quantities related to structures. [Brief explanation of the drawings]
[0015] [Figure 1] 1 is a configuration example of a wireless communication system according to a first embodiment of the present disclosure. [Figure 2] 2 illustrates a configuration example of a transmitter according to the first embodiment of the present disclosure. [Figure 3] 2 illustrates a configuration example of a receiver according to the first embodiment of the present disclosure. [Figure 4] 1 is a block diagram illustrating an example configuration of a feature estimation device according to a first embodiment of the present disclosure. [Figure 5] FIG. 2 is a block diagram showing a first learning model according to the first embodiment of the present disclosure. [Figure 6] FIG. 2 is a block diagram showing a second learning model according to the first embodiment of the present disclosure. [Figure 7] 4 is a flowchart showing processing performed by the feature estimation device according to the first embodiment of the present disclosure during learning. [Figure 8] 4 is a flowchart showing a process performed by the feature estimation device according to the first embodiment of the present disclosure when estimating a structure. [Figure 9] 1 is a diagram illustrating a hardware configuration of a feature estimation device according to a first embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION
[0016] Embodiments of the present disclosure will be described with reference to the drawings. The same or corresponding components will be designated by the same reference numerals, and repeated description may be omitted.
[0017] Embodiment 1 1 shows a configuration example of a wireless communication system 100 according to a first embodiment of the present disclosure. A wireless LAN transmitter (hereinafter referred to as a transmitter) 110 includes one or more antennas 111-1, 111-2, ..., 111-n. The transmitter 110 uses these antennas 111 simultaneously to communicate with a wireless LAN receiver (hereinafter referred to as a receiver) 120. Note that although the present example shows a case where there are two transmitters 110 and two receivers 120, there may be one or more transmitters 110 and one or more receivers 120.
[0018] In the following description, the antennas 111-1, 111-2, and so on are applicable to each other, and when there is no need to distinguish between them, they will simply be referred to as antenna 111. Only when it is necessary to distinguish between them, will the respective reference symbols be used. The same applies to antennas 121-1, 121-2, and so on of receiver 120.
[0019] The radio signal transmitted between the transmitter 110 and the receiver 120 is affected by structures 50 that exist in the propagation path.
[0020] Receiver 120 receives radio signals from transmitter 110 using one or more antennas 121-1, 121-2, ..., 121-m. Note that if transmitter 110 has one antenna 111, receiver 120 has two or more antennas 121. However, transmitter 110 usually has two or more antennas 111.
[0021] Based on the received radio signal, the receiver 120 derives CSI (Channel State Information), which is information about the radio propagation path between the antenna 111 of the transmitter 110 and the antenna of the receiver 120. The CSI includes amplitude information and phase information for each OFDM subcarrier. It may also include SNR (Signal-to-Noise Ratio) information.
[0022] The receiver 120 notifies the feature estimation device 130 of the CSI.
[0023] The feature estimation device 130 estimates multiple features of the structure 50 based on the CSI received from the receiver 120 using a machine learning model based on a neural network or principal component analysis, which will be described later. The features of the structure 50 include density, volume, angle, shape, material, etc. Furthermore, if there are multiple structures 50, the features may be something that indicates the state of the multiple structures 50, such as the spacing between the structures 50.
[0024] In this way, the structure 50 present in the propagation path of the wireless signal exchanged between the transmitter 110 and the receiver 120 is measured using the CSI. The feature estimation device 130 analyzes the measured CSI, thereby making it possible to estimate multiple features of the structure 50.
[0025] The feature estimation device 130 may be built into the transmitter 110 or the receiver 120.
[0026] In the above description, the feature estimation device 130 receives notification of CSI directly from the receiver 120. However, the feature estimation device 130 may capture a radio frame used when the receiver 120 feeds back CSI compressed by a method compliant with IEEE802.11ac / ax to the transmitter 110.
[0027] 2 shows an example of the configuration of the transmitter 110 according to the first embodiment of the present disclosure. The signal generating circuit 112 generates a radio signal including a known frame. The transmitting circuit 113 transmits the radio signal to the receiver 120 via the antenna 111.
[0028] 3 shows a configuration example of a receiver 120 according to the first embodiment of the present disclosure. A receiving circuit 122 guides radio waves arriving at each antenna 121 to generate a power signal. A measuring circuit 123 derives CSI from the radio signal. A reporting circuit 124 reports the CSI to a feature estimation device 130.
[0029] FIG. 4 is a block diagram showing an example configuration of a feature estimation device 130 according to the first embodiment of the present disclosure. A preprocessing circuit 131 receives CSI from the receiver 120. The received CSI indicates the state of the propagation path in absolute values, but the preprocessing circuit 131 converts this into relative value information between the antenna 111 of the transmitter 110 and the antenna of the receiver 120. The preprocessing circuit 131 then extracts the amplitude and phase of the radio signal from the relative value information. At this time, the preprocessing circuit 131 performs unwrapping processing on the phase to interpolate the phase continuity. The preprocessing circuit 131 also reduces phase errors by correcting the difference in clock frequencies or the difference in reception timing between the transmitter 110 and the receiver 120.
[0030] In this way, the preprocessing circuit 131 extracts the amplitude and phase of the radio signal between the antenna 111 of the transmitter 110 and the antenna of the receiver 120 and inputs them as difference information 10 to the first learning model 132 .
[0031] The first learning model 132 executes an individual estimation process to individually estimate multiple feature quantities 20 of the structure 50 from the input difference information 10. The feature quantities 20 here are numerical values of the features of the structure 50, such as density, volume, angle size, or shape, material, etc. Hereinafter, the multiple feature quantities 20 estimated by the first learning model 132 will be referred to as a first feature quantity 20-1, a second feature quantity 20-2, ..., an Nth feature quantity 20-N. For example, the first feature quantity 20-1 is density, the second feature quantity 20-2 is volume, ..., and so on, and each feature quantity 20 numerically indicates a different feature of the structure 50.
[0032] The first learning model 132 inputs the estimated plurality of feature quantities 20 into the second learning model 133 .
[0033] On the other hand, during learning, the first learning model 132 executes a first learning process to learn so that a plurality of feature quantities 20 obtained by performing an individual estimation process on the input learning difference information 10 approaches the correct answer based on a loss function or an RMSE (Root Mean Squared Error) value, etc. The learning difference information 10 is difference information 10 obtained from CSI measured for a structure 50 whose feature quantities 20 are known.
[0034] The second learning model 133 executes a combined estimation process to estimate multiple modified features 30 for multiple features 20 input by the first learning model 132. Specifically, the second learning model 133 performs further estimation for the input first feature 20-1 and outputs the result to the output circuit 134 as a first modified feature 30-1. The second learning model 133 also performs further estimation for the input second feature 20-2 by combining it with the first modified feature 30-1 and outputs the result to the output circuit 134 as a second modified feature 30-2. Similarly, the second learning model 133 performs further estimation for the Nth feature 20-N by combining it with the first modified feature 30-1 to the (N-1)th modified feature 30-(N-1), and outputs the result to the output circuit 134 as the Nth modified feature 30-N.
[0035] In this way, the second learning model 133 combines multiple features 20 estimated individually in the first learning model 132 and then performs further estimation, thereby making it possible to improve estimation accuracy.
[0036] On the other hand, during learning, the second learning model 133 receives the plurality of features 20 estimated by the first learning model 132 for the learning difference information 10, and performs a second learning process to learn so that the plurality of corrected features 30 obtained by performing a joint estimation process approach the correct answer. Note that the learning here is also based on a loss function or RMSE, etc.
[0037] 5 is a block diagram showing the first learning model 132 according to the first embodiment of the present disclosure. Here, a case where the first learning model 132 is a neural network will be described. A batch 11-1 to be analyzed out of the difference information 10 is shown as an input to the first learning model 132. Note that the numerical values in each batch 11 indicate an example of the batch size.
[0038] In the first learning model 132, batch 11-1 is first input to input layer 1 and calculated. Next, the calculation result is subjected to a linear transformation using the function Linear (hereinafter simply referred to as linear transformation) to obtain a reduced-size batch 11-2. Next, batch 11-2 is input to middle layer 2 and calculated. In middle layer 2, the calculation result is first normalized using the function ReLu (hereinafter simply referred to as normalization) and linearly transformed to obtain a reduced-size batch 11-3. Further calculations are performed on batch 11-3. In this way, multiple features 20 are extracted by repeating neuron calculations on batch 11 in middle layer 2. Then, normalization and linear transformation are performed on the calculation result of batch 11-3 to obtain a further reduced-size batch 11-4. By shaping batch 11-4 using the function Data reshape, batch 12-1 containing multiple features 20 is obtained as the output of middle layer 2.
[0039] 6 is a block diagram showing the second learning model 133 according to the first embodiment of the present disclosure. Here, a case where the second learning model 133 is a neural network will be described. The second learning model 133 receives a batch 12-1 including a plurality of features 20 from the intermediate layer 2 of the first learning model 132. In general, a batch handled in the intermediate layer 2 has more data than a batch handled in the output layer (not shown), and can be said to be more suitable for analysis.
[0040] The second learning model 133 first performs convolution using the function Conv on batch 12-1, which includes multiple features 20, to obtain a smaller batch 12-2. Then, batch 12-2 is normalized using the functions BatchNorm2d and ReLu. The normalized result is then convolved again using the function Conv to obtain a smaller batch 12-3. This process of repeated convolution and normalization reduces the size of batch 12, ultimately resulting in batch 12-5. This process also involves processing such as sorting the data within batch 12 based on combinations of features that mutually improve the estimation results.
[0041] Furthermore, batch 12-5 is pooled using the function AdaptiveAvePpol2d, resulting in a further reduced batch 12-6.
[0042] If the estimation for batch 12-6 is the first time, this means that the second learning model 133 does not have any estimation results for features that have already been estimated. Therefore, the second learning model 133 imports batch 12-6 itself into the coupling layer 3. In the coupling layer 3, a linear transformation is performed on batch 12-6 to obtain a reduced-size batch 12-8. The same linear transformation is then performed on batch 12-8 to obtain a reduced-size batch 12-9. Batch 12-9 is then normalized and linearly transformed to reduce its size. In this process of reducing the size of batch 12 in the coupling layer 3, the second learning model 133 further estimates the first feature 20-1 to obtain a first corrected feature 30-1.
[0043] The second learning model 133 holds the first modified feature 30-1 as the combined information 6.
[0044] When the estimation for batch 12-6 is the second time, the second learning model 133 holds the first modified feature 30-1 as the combined information (Sequential Info) 6. In this case, the second learning model 133 creates batch 12-7 by combining the first modified feature 30-1 with batch 12-6, and then imports it into the combined layer 3.
[0045] The second learning model 133 further adds the second modified feature 30-2 to the combined information 6.
[0046] Similarly, the second learning model 133 combines the first modified feature 30-1 to the Nth modified feature 30-(N-1) with batch 12-6 to create batch 12-7. By inputting batch 12-7 to the combining layer 3, the Nth modified feature 30-N is obtained.
[0047] The position where the batch 12 and the combined information 6 are combined is an example, and may be determined by using a general-purpose automatic parameter optimization framework, for example.
[0048] 7 is a flowchart showing the processing performed by the feature estimation device 130 according to the first embodiment of the present disclosure during learning. First, the preprocessing circuit 131 derives difference information 10 for learning (step S01). Next, the first learning model 132 executes a first learning process based on the difference information 10 for learning (step S02). Furthermore, the second learning model 133 executes a second learning process based on the plurality of feature amounts 20 estimated by the first learning model (step S03).
[0049] 8 is a flowchart showing the processing performed by the feature estimation device 130 according to the first embodiment of the present disclosure when estimating a structure 50. First, the preprocessing circuit 131 derives the difference information 10 (step S11). Next, the first learning model 132 performs an individual estimation process on the difference information 10 to estimate multiple feature quantities 20 (step S12). Next, the second learning model 133 performs a joint estimation process on the multiple feature quantities 20 to obtain multiple corrected feature quantities 30 as estimation results (step S13).
[0050] 9 is a diagram illustrating a hardware configuration of the feature estimation device 130 according to the first embodiment of the present disclosure. The processing performed by the feature estimation device 130 may be executed by a program using a computer including a CPU and memory and storing a program in the memory. Alternatively, the processing may be executed by a program using an integrated circuit such as an FPGA (Field Programmable Gate Array). The program may be provided by being recorded on a storage medium or via a network.
[0051] The feature estimation device 130 has an input unit 40, an output unit 41, a communication unit 42, a CPU 43, a memory 44, and an HDD 45 connected via a bus 46, and functions as a computer. The feature estimation device 130 is also configured to be able to input and output data to and from a computer-readable storage medium 47.
[0052] The input unit 40 is, for example, a keyboard and a mouse, etc. The output unit 41 is, for example, a display device such as a display.
[0053] The communication unit 42 is, for example, a communication interface that communicates with a wireless device to be controlled.
[0054] The CPU 43 controls each component of the feature estimation device 130 and performs predetermined processing, etc. The memory 44 and HDD 45 store data, etc.
[0055] The storage medium 47 is capable of storing programs and the like that cause the feature estimation device 130 to execute the functions of the feature estimation device 130. Note that the architecture that configures the feature estimation device 130 is not limited to the example shown in the figure.
[0056] As described above, according to the present disclosure, the second learning model 133 combines multiple feature quantities 20 estimated individually in the first learning model 132 and performs further estimation. This makes it possible to provide a wireless communication system, a wireless communication method, a feature quantity estimation device 130, and a wireless communication program that can improve the estimation accuracy of multiple feature quantities related to a structure.
[0057] <Variation 1> The antenna 111 of the transmitter 110 and the antenna 121 of the receiver 120 may be polarized antennas. This increases the number of branches in the propagation path of the wireless signal and the amount of CSI that can be acquired, which is expected to improve the estimation accuracy of the feature estimation device 130.
[0058] <Variation 2> Note that when there are multiple transmitters 110 or receivers 120, the CSI measured for a transmitter-receiver combination with high estimation accuracy of the corrected feature 30 may be preferentially acquired and used to estimate the feature. Alternatively, the CSI obtained for that combination may be weighted more heavily than the CSI obtained for other combinations, and then the feature may be estimated. This enables efficient estimation of the feature of the structure 50. The feature estimation device 130 here performs a process of estimating the estimation accuracy of the corrected feature 30 for each combination based on a loss function or RMSE value, etc., and a process of determining a transmitter-receiver combination with high estimation accuracy.
[0059] <Variation 3> When a wireless signal is transmitted across multiple frequencies, the CSI at a frequency or frequency band where the estimation accuracy of the corrected feature 30 is high may be preferentially acquired to estimate the feature. Alternatively, the feature may be estimated by weighting the CSI at that frequency or frequency band more heavily than the CSI at other frequencies. This enables the feature of the structure 50 to be estimated efficiently. The frequency or frequency band where the estimation accuracy is high may be determined based on the loss function, RMSE value, or the like, as described above.
[0060] <Variation 4> In order to reduce noise, the feature estimation device 130 may perform a moving average process or the like on the CSI.
[0061] <Variation 5> When the modulation method of the wireless signal is OFDM, the feature estimation device 130 may determine subcarriers with high received power or subcarriers with high estimation accuracy of the corrected feature 30, and acquire the CSI of those subcarriers preferentially. This makes it possible to efficiently estimate the feature of the structure 50. Note that the subcarriers with high estimation accuracy may be determined based on the loss function or RMSE value, as described above.
[0062] Variation 6 If a change that may affect the quality of the wireless signal, such as interference between wireless signals on different channels or a change in weather, is observed, the CSI measured while the change is observed may be excluded from the analysis target in the feature estimation device 130. Furthermore, if a fluctuation exceeding a threshold is observed in the measured CSI, the CSI having the fluctuation exceeding the threshold may be excluded from the analysis target. In this way, by limiting the CSI to be analyzed according to the quality of the wireless signal, improvement in estimation accuracy is expected.
[0063] <Variation 7> Of the multiple feature quantities 20 estimated by the first learning model 132, feature quantities expected to have high estimation accuracy may be weighted heavily and then input to the second learning model 133. The estimation results of feature quantities expected to have high estimation accuracy at the time of estimation by the first learning model 132 can be passed on to the second learning model 133, which is expected to improve the estimation accuracy. The feature quantity estimation device 130 here performs a process of estimating the estimation accuracy of each of the multiple feature quantities 20 estimated by the first learning model 132 based on a loss function, RMSE value, or estimation results, etc., and a process of determining feature quantities 20 with high estimation accuracy.
[0064] <Variation 8> Note that the modified features 30 may be weighted before being combined with the features 20 input from the first learning model 132. By assigning a large weight to a modified feature 30 that is expected to have high estimation accuracy, it is expected that the estimation accuracy of the feature 20 combined with that modified feature will also be improved. The feature estimation device 130 here performs a process of estimating the estimation accuracy of each of the multiple modified features 30 based on a loss function, an RMSE value, or estimation results, and a process of determining the modified feature 30 with high estimation accuracy.
[0065] <Variation 9> In the second learning model 133, multiple corrected features 30 may be combined together to be converted into new features.
[0066] Variation 10 The first learning model 132 and the second learning model 133 may be made to learn the time series information of the difference information 10. This makes it possible to predict the time change of the feature amount.
[0067] The present disclosure is not limited to the above-described embodiment, and various modifications can be made in the implementation stage without departing from the spirit of the present disclosure. Furthermore, the modifications may be implemented in appropriate combination, and in such a case, the combined effects can be obtained. [Explanation of symbols]
[0068] 1 input layer, 2 hidden layer, 3 connection layer, 6 connection information, 10 difference information, 11 batch, 12 batch, 20 feature, 20-1 first feature, 20-2 second feature, 20-N Nth feature, 30 corrected feature, 30-1 first corrected feature, 30-2 second corrected feature, 30-N Nth corrected feature, 40 input unit, 41 output unit, 42 communication unit, 43 CPU, 44 memory, 45 HDD, 46 bus, 47 storage medium, 50 structure, 100 wireless communication system, 110 transmitter, 111 antenna, 112 signal generation circuit, 113 transmission circuit, 120 receiver, 121 antenna, 122 receiving circuit, 123 measurement circuit, 124 notification circuit, 130 feature estimation device, 131 preprocessing circuit, 132 First learning model, 133 Second learning model
Claims
1. 1. A wireless communication system that estimates characteristics of a structure present on a propagation path of a wireless signal based on the wireless signal, comprising: a transmitter for transmitting the radio signal; a receiver for deriving propagation channel information from the radio signal; a feature estimation device that receives the propagation channel information from the receiver; Equipped with The feature estimation device includes: a first learning model and a second learning model; deriving differential information of the radio signal between the transmitter antenna and the receiver antenna from the propagation channel information; an individual estimation process of inputting the difference information into the first learning model and estimating a plurality of feature amounts; a combined estimation process of inputting the plurality of feature quantities into the second learning model and estimating a plurality of corrected feature quantities; configured to run In the joint estimation process, a process of estimating a first modified feature based on the feature; and estimating a second modified feature by combining the feature and the first modified feature.
2. During learning, the feature estimation device a process of training the first learning model so that a plurality of feature quantities obtained by performing the individual estimation process on learning difference information whose feature quantities have known values approach the known values; a process of training the second learning model so that a plurality of corrected feature quantities obtained by performing the joint estimation process on a plurality of feature quantities estimated by the first learning model approach the known value; The wireless communication system of claim 1 , configured to perform the following:
3. In the joint estimation process, a process of estimating estimation accuracy of the plurality of feature amounts estimated by the first learning model; a process of giving a large weight to the feature amount with high estimation accuracy and inputting the feature amount to the second learning model; The wireless communication system according to claim 1 or 2, further comprising:
4. In the joint estimation process, a process of estimating the estimation accuracy of the corrected feature; a process of combining the feature and the modified feature after assigning a large weight to the modified feature with high estimation accuracy; The wireless communication system according to claim 1 or 2, further comprising:
5. The transmitter or the receiver is plural, The feature estimation device includes: a process of determining a combination of the transmitter and the receiver that has a high estimation accuracy of the corrected feature; a process of preferentially acquiring the propagation channel information measured in the combination; The wireless communication system according to claim 1 or 2, further comprising:
6. 1. A feature estimation device that estimates features of a structure present in a propagation path of a wireless signal, based on the wireless signal exchanged between a transmitter and a receiver, comprising: a first learning model and a second learning model; receiving propagation channel information from the receiver; deriving differential information of the radio signal between the transmitter antenna and the receiver antenna from the propagation channel information; an individual estimation process of inputting the difference information into the first learning model and estimating a plurality of feature amounts; a combined estimation process of inputting the plurality of feature quantities into the second learning model and estimating a plurality of corrected feature quantities; configured to run In the joint estimation process, a process of estimating a first modified feature based on the feature; and estimating a second modified feature by combining the feature and the first modified feature.
7. 1. A wireless communication method for estimating characteristics of a structure present in a propagation path of a wireless signal, based on the wireless signal exchanged between a transmitter and a receiver, comprising: deriving propagation channel information from the wireless signal; deriving differential information of the radio signal between the transmitter antenna and the receiver antenna from the propagation channel information; executing an individual estimation process for inputting the difference information into a first learning model and estimating a plurality of feature quantities; inputting the plurality of feature quantities into a second learning model and executing a joint estimation process to estimate a plurality of corrected feature quantities; Including, In the joint estimation process, estimating a first modified feature based on the feature; and estimating a second modified feature by combining the feature and the first modified feature.
8. 1. A wireless communication program to be executed by a feature estimation device that estimates features of a structure present in a propagation path of a wireless signal, based on the wireless signal exchanged between a transmitter and a receiver, the program comprising: The feature estimation device includes: receiving propagation channel information from the receiver; deriving differential information of the radio signal between the transmitter antenna and the receiver antenna from the propagation channel information; an individual estimation process of inputting the difference information into a first learning model and estimating a plurality of feature quantities; a combined estimation process of inputting the plurality of feature quantities into a second learning model and estimating a plurality of corrected feature quantities; a program for executing In the joint estimation process, a process of estimating a first modified feature based on the feature; and estimating a second modified feature by combining the feature and the first modified feature.