Methods, programs, and apparatus for estimating the quality of wireless communications

Abstract

This technology enhances the accuracy, robustness, and generalizability of prediction models for received signal strength in outdoor wireless communication systems. [Solution] A method for estimating an index representing the quality of wireless communication includes the steps of: accessing a first model obtained by learning data obtained by wireless communication in a first area; performing wireless communication between a transmitter and a first receiver among a plurality of receivers in a second area different from the first area; creating a dataset using the results of the wireless communication; deriving a second model by adjusting the first model using the dataset; and estimating an index representing the quality of wireless communication between a transmitter and a second receiver among a plurality of receivers, using an elevation matrix, a building height matrix, a terrain category matrix, a frequency matrix, and the second model obtained for the second receiver.

Description

[Technical Field]

[0001] This disclosure relates to wireless communication, and more specifically, to a technique for estimating received signal strength or bit error rate. [Background technology]

[0002] Advances in outdoor wireless communication systems have revolutionized modern connectivity, enabling ubiquitous access to information. From urban areas to suburbs, outdoor wireless communication systems serve as the backbone of modern communication infrastructure, supporting a wide range of applications, including voice, data, and multimedia transmission.

[0003] Accurate prediction of received signal strength (hereinafter sometimes referred to as "RSS") is crucial for the seamless operation of outdoor wireless communication systems. Received signal strength is a key metric in wireless communication, quantifying the power level of the signal propagating through the environment from the transmitter to the receiver. Reliable prediction of received signal strength is essential for optimizing network performance, ensuring adequate coverage, and enabling efficient resource allocation.

[0004] Traditionally, empirical models based on path loss equations, such as Frith's transmission equations and the Okumura-Hata model, have been used to predict received signal strength. While these models are easy to implement, they often lack the granularity necessary to capture the complex propagation phenomena observed in outdoor environments. Factors such as topographic irregularities, building morphology, vegetation, and atmospheric conditions significantly influence signal propagation, greatly affecting the propagation of conventional models and posing challenges to their accuracy. To address these limitations, the application of machine learning techniques for predicting received signal strength is being explored. Machine learning algorithms, particularly supervised learning-based machine learning algorithms, utilize past measurements of received signal strength to learn the complex relationships between environment variables and signal strength. Techniques such as regression, decision trees, support vector machines, and neural networks have been applied. While machine learning-based approaches offer greater flexibility and adaptability compared to empirical models, they also have their own unique challenges.

[0005] Historically, researchers and engineers have employed a variety of techniques to address the challenge of receiving signal strength in outdoor wireless communication systems. These techniques can be broadly categorized into empirical models, deterministic models, and more recently, machine learning methods.

[0006] Empirical models such as Frith's transmission equation (Non-Patent Literature 1), the Okumura-Hata model (Non-Patent Literature 2), and the COST 231 Hata model (Non-Patent Literature 3) have long been used as the basis for predicting received signal strength in wireless communication. These models are based on simplified assumptions about signal propagation and typically rely on parameters such as distance, frequency, and antenna height to estimate losses. Empirical models provide a rapid estimate of received signal strength and are computationally inexpensive.

[0007] Deterministic models, also known as ray tracing or ray propagation models (Non-Patent Literature 4), provide more detailed, physics-based methods for predicting received signal strength. These models can simulate electromagnetic waves through the environment by tracing individual rays and considering reflection, diffraction, and scattering phenomena. Deterministic models can provide relatively accurate predictions in certain scenarios.

[0008] In recent years, machine learning techniques have gained popularity in the field of predicting received signal strength because they can learn complex patterns from data. Supervised learning algorithms such as regression, decision trees, support vector machines, and neural networks have been applied to model relationships between input features, for example, geographic coordinates, topographic characteristics, and received signal strength.

[0009] By learning from historical received signal strength measurements collected from various environments, machine learning models can capture nonlinear relationships and adapt to changes in environmental conditions.

[0010] Non-Patent Document 5 presents a machine learning scheme based on geographical features for predicting received signal strength. Non-Patent Document 5 discloses a method for precisely selecting four features closely related to received signal strength from readily available geographical datasets and computing them at low cost. The experiments described in Non-Patent Document 5 were conducted in large-scale outdoor scenarios on 3.5G or 5G networks where actual received signal strength was collected by field measurements in urban areas. The experimental results demonstrate the effectiveness of the extracted features, showing that the proposed method is superior to existing methods in terms of model accuracy and computational efficiency. Furthermore, Non-Patent Document 5 discusses the potential of the proposed method to build a universal, accurate, and efficient machine learning-based method for predicting received signal strength based on vast amounts of field data in 5G and B5G networks.

[0011] Patent Document 1 discloses "a received power estimation device that can accurately estimate received power taking into account differences in quality at each location and time variations in quality" (see [Problem] in [Abstract]).

[0012] Non-patent document 6 presents an alternative method for predicting propagation path loss in urban environments, which is based on artificial neural networks (NNs) used in machine learning (ML) techniques. The results demonstrate the effectiveness of obtaining results with sufficient accuracy in a short time.

[0013] Non-patent document 7 proposes an ML-based method for rapidly predicting propagation path loss in urban areas using data extracted from online sources such as OpenStreetMap and other geographic information systems, and discloses a technology that supports the estimation of mobile phone coverage in a given area. [Prior art documents] [Patent Documents]

[0014] [Patent Document 1] Japanese Patent Publication No. 2020-191555 [Non-patent literature]

[0015] [Non-Patent Document 1] A. Goldsmith, Wireless Communications. Stanford University, 2004. [Non-Patent Document 2] M. Hata, “Empirical formula for propagation loss in land mobile radio services,” IEEE Transactions on Vehicular Technology, vol. 29, pp. 317-325, August 1980. [Non-Patent Document 3] V. Abhayawardhana, I. Wassell, D. Crosby, M. Sellars, and M. Brown, “Comparison of empirical propagation path loss models for fixed wireless access systems,” in 2005 IEEE 61st Vehicular Technology Conference, (Stockholm, Sweden), pp. 1-5, IEEE, December 2005. [Non-Patent Document 4] K. Schaubach, N. Davis, and T. Rappaport, “A ray tracing method for predicting path loss and delay spread in microcellular environments,” in Vehicular Technology Society 42nd VTS Conference - Frontiers of Technology, (Denver, CO, USA), pp. 1-4, IEEE, August 1992. [Non-Patent Document 5] Y. Liu, J. Dong, W. Huangfu, J. Liu, and K. Long, “3.5 ghz outdoor radio signal strength prediction with machine learning based on low-cost geographic features,” IEEE Transactions on Antennas and Propagation, vol. 70, pp. 4155-4170, May 2022. [Non-Patent Document 6] S. P. Sotiroudis, S. K. Goudos, K. A. Gotsis, K. Siakavara, and J. N. Sahalos, “Application of a composite differential evolution algorithm in optimal neural network design for propagation path-loss prediction in mobile communication systems,” IEEE Antennas and Wireless Propagation Letters, vol. 12, pp. 364-367, March 2013. [Non-Patent Document 7] I. F. M. Rafie, S. Y. Lim, and M. J. H. Chung, “Path loss prediction in urban areas: A machine learning approach,” IEEE Antennas and Wireless Propagation Letters, vol. 22, pp. 809-813, April 2023. [Summary of the Invention] [Problems to be Solved by the Invention]

[0016] Empirical models such as the Friis transmission equation (Non-Patent Document 1), the Okumura-Hata model (Non-Patent Document 2), and the COST 231 Hata model (Non-Patent Document 3) have been the basis for RSS prediction in wireless communication. These models are based on simplified assumptions regarding signal propagation and typically estimate path loss depending on parameters such as distance, frequency, and antenna height. Although empirical models provide a quick estimate of RSS and have a low computational cost, they often cannot capture the complex effects of environmental factors such as terrain, buildings, and vegetation on signal propagation.

[0017] Deterministic models, also known as ray tracing or ray propagation models (Non-Patent Literature 4), provide a more detailed, physics-based approach to RSS prediction. These models simulate the propagation of electromagnetic waves through the environment by tracing individual rays and by considering reflection, diffraction, and scattering phenomena. While deterministic models can yield relatively accurate predictions in certain scenarios, they require detailed knowledge of the surrounding environment, including topographic elevation data, building layouts, and material properties. As a result, deterministic models are often computationally intensive and impractical for large-scale deployments.

[0018] Machine learning models for RSS prediction in outdoor wireless communication systems can capture nonlinear relationships and adapt to changing environmental conditions by learning from historical RSS measurements collected in various environments (Patent Document 1, Non-Patent Documents 5-7). While each machine learning-based approach offers the potential to improve RSS prediction accuracy, they each have their own unique challenges.

[0019] One major constraint is the need for large, representative training datasets to train an accurate model. Collecting and annotating RSS measurements across diverse geographical regions and environmental conditions is resource-intensive and time-consuming.

[0020] Furthermore, machine learning models trained on data from one region may not be generalizable to other regions due to differences in topography, vegetation, or urban morphology. As a result, the aforementioned data collection and annotation work must be repeated. This lack of generalizability not only hinders the scalability and applicability of machine learning-based RSS prediction methods but also impairs their effectiveness in real-world deployment scenarios. Moreover, this challenge also stems from the inherent heterogeneity of outdoor environments, where topography, vegetation density, and urban infrastructure vary significantly from place to place. Machine learning models trained on data from a specific geographical region may inadvertently learn region-specific biases and specificities, potentially leading to decreased performance when applied to new regions.

[0021] Furthermore, the complex interactions between environmental factors and signal propagation present further obstacles to generalization. For example, the presence of dense foliage in rural areas may result in different attenuation of radio signals compared to urban canyons characterized by high-rise buildings and crowded streets. Similarly, differences in topography can lead to multipath and shadowing phenomena, which have different effects on signal strength across different landscapes.

[0022] Furthermore, the dynamic nature of outdoor wireless channels often presents challenges for conventional machine learning algorithms that assume a stationary data distribution. Changes in environmental conditions, such as seasonal shifts and urban development, can alter signal propagation characteristics over time, requiring continuous model adaptation and retraining.

[0023] Despite these challenges, there is growing interest in new approaches that incorporate domain-specific geographical knowledge into the RSS forecasting process.

[0024] This disclosure is made in view of the above-described background, and in a certain respect, the purpose is to provide a technology that enhances the accuracy, robustness, and generalizability of RSS prediction models in outdoor wireless communication systems. [Means for solving the problem]

[0025] According to one embodiment, a method is provided that is performed on a computer using a model for estimating the quality of wireless communication between a transmitter and a plurality of receivers. The dataset used by the model includes an elevation matrix, a building height matrix, a terrain category matrix, a frequency matrix, and a numerical value representing quality, all obtained by the wireless communication. The method includes the steps of: accessing a first model obtained by learning data obtained by wireless communication in a first region; performing wireless communication between a transmitter and a first receiver among a plurality of receivers in a second region different from the first region; creating a dataset using the results of the wireless communication between the transmitter and the first receiver; adjusting the first model using the dataset to derive a second model; and, for a second receiver among the plurality of receivers, estimating an index representing the quality of wireless communication between the transmitter and the second receiver using the elevation matrix, building height matrix, terrain category matrix, frequency matrix, and the second model obtained for the second receiver.

[0026] In a given context, an indicator representing the quality of wireless communication may include either the received signal strength or a numerical value representing the influence of the environment in which the wireless communication takes place.

[0027] According to another embodiment, a method is provided for a computer to create a database used for estimating the received signal strength of wireless communication between a transmitter and a receiver. The database includes an elevation matrix, a building height matrix, a terrain category matrix, a frequency matrix, and numerical values ​​representing environmental impacts. This method includes the steps of: calculating an elevation matrix using the latitude and longitude of the transmitting device, the latitude and longitude of the receiving device, and a pre-prepared elevation model for deriving altitude; calculating a building height matrix using the latitude and longitude of the transmitting device, the latitude and longitude of the receiving device, and a pre-prepared building height model for deriving building height; calculating a terrain category matrix using the latitude and longitude of the transmitting device, the latitude and longitude of the receiving device, and a pre-prepared terrain category model for deriving terrain categories; calculating a frequency matrix using the frequency of the signal transmitted by the transmitting device; and calculating a numerical value representing the environmental impact based on the received signal strength measured by the receiving device, the transmission output of the transmitting device, the respective antenna gains of the transmitting and receiving devices, the frequency, and the spatial distance between the transmitting and receiving devices.

[0028] In other embodiments, a program is provided that causes a computer to perform the method described in any of the above-described methods.

[0029] In yet another embodiment, an apparatus is provided comprising a memory storing a program that causes a computer to perform the method described in any of the above-described methods, and a processor that executes the program.

[0030] The above and other objects, features, aspects and advantages of this invention will become apparent from the following detailed description relating to this invention, which will be understood in conjunction with the accompanying drawings. [Brief explanation of the drawing]

[0031] [Figure 1] This is a diagram showing the general configuration of a communication network. [Figure 2]This is a block diagram showing the hardware configuration of the computer device 200 that implements the estimation device 100. [Figure 3] This figure shows a comparison between existing methods for estimating received signal strength and the method for estimating received signal strength according to this embodiment. [Figure 4] This diagram illustrates an overview of the communication equipment used for conducting measurement experiments in Area 1. [Figure 5] This diagram illustrates an overview of the technical concept applied to the estimation device 100. [Figure 6] This diagram conceptually represents an example of a database 600 configured on the hard disk 5 of the estimation device 100. [Figure 7] This diagram illustrates the relationship between two coordinate systems. [Figure 8] This figure illustrates an example of calculating the elevation of each sample point when a second coordinate system is applied to map 800. [Figure 9] This diagram illustrates an example of model training. [Figure 10] This diagram shows a state in which multiple terminal devices 110 are deployed in a new area. [Figure 11] This figure shows an example of a dataset 1100 configured on the hard disk 5 of a computer device 200 that functions as an estimation device 100. [Figure 12] This is a diagram illustrating the fine-tuning of the model. [Figure 13] This diagram illustrates the procedure for estimating the environmental impact 670 or average BER 680 using the estimation device 100. [Modes for carrying out the invention]

[0032] Embodiments of the present invention will be described below with reference to the drawings. In the following description, identical parts are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed descriptions of them will not be repeated.

[0033] First, to overcome the limitation of machine learning-based RSS prediction requiring large and representative training datasets, and to enhance the generalization of machine learning models across diverse regions with different topography, vegetation, and urban forms, transfer learning (TL) techniques are employed in machine learning-based RSS prediction. Here, transfer learning is a machine learning technique in which a model trained on one task is used to improve performance on related tasks. In short, knowledge gained by solving one problem is transferred and reused to solve different but related problems. This approach is particularly effective when training the model is computationally expensive, or when the task in question has only limited labeled data.

[0034] According to one embodiment, the process of using transfer learning to solve the RSS prediction problem in machine learning-based methods includes five main steps: (1) selecting a pre-trained model, (2) selecting a transfer task, (3) determining a transfer strategy, (4) training a transfer model, and (5) evaluating its performance.

[0035] (1) Step of selecting a pre-trained model This step involves selecting a pre-trained model that has been trained on a large dataset for a specific task, such as RSS prediction for outdoor wireless communication as described above. This pre-trained model has been trained on a large amount of historically acquired datasets to learn general functions that can be broadly applied to various environmental effects and communication scenarios.

[0036] (2) Step of selecting a transfer task This step involves identifying the target task to which transfer learning should be applied. This target task may be related to the task in which the pre-trained model was initially trained, or it may be related to a completely different task. The point here is to identify a task in which the trained representation of the pre-trained model may be useful. For example, a pre-trained model trained on the Tokyo dataset could be used with a transfer learning technique to address the problem of RSS prediction in Kyoto. Here, the pre-trained model trained on the Tokyo dataset is selected as the pre-trained model, and the problem of RSS prediction in Kyoto is selected as the target task (transfer task).

[0037] (3) Steps to determine the transfer strategy Once a pre-trained model is selected and a transfer task is defined, this embodiment determines a transfer strategy to which fine-tuning has been applied. The fine-tuning involves retraining the entire pre-trained model on the new task (in the example above, the RSS prediction problem in Kyoto) using a smaller learning rate to prevent forgetting of the learned representations. This allows the pre-trained model to apply the learned features to the details of the target task while still benefiting from the knowledge gained during pre-training.

[0038] (4) Steps to train the transfer model Once the transfer strategy is determined, the transfer model is trained using the approach selected as that transfer strategy. This step typically involves fine-tuning the pre-trained model with the finer details of the target task's dataset.

[0039] (5) Steps to evaluate performance This step involves evaluating the performance of the transfer model on a separate validation or test dataset to assess its effectiveness for the target task. This allows for identifying how well the fine-tuned pre-trained model generalizes to new data and identifying room for improvement or further fine-tuning.

[0040] To address the challenges arising from the complex interaction between environmental factors and signal propagation dynamics, which present further obstacles to the generalization of trained models, and to tackle the challenges arising from the dynamic nature of outdoor radio channels that lead to shifts in signal propagation characteristics, this embodiment discloses a feature extraction approach that takes geographical knowledge into account for generating feature vectors in a machine learning process.

[0041] First, in this embodiment, the received signal strength (RSS) η in the transmit / receive link is interpreted as a value consisting of a constant C derived from the free-space loss between the transceiver and the receiver, and a variable E that changes continuously based on changes in the local environment. This received signal strength η (dBm) is shown in equation (1).

[0042]

number

[0043] The constant C is given by equation (2).

[0044]

number

[0045] Here, the constant P represents the transmitter's output. The constant G t The constant G represents the gain of the transmitter's antenna. r The constant λ represents the gain of the receiver's antenna. The constant d represents the operating wavelength. The constant d represents the distance between the transmitter and the receiver.

[0046] The variable E is given by equation (3).

[0047]

number

[0048] Here, f1,···,f N These are N environmental factors or layers that influence the change in received signal strength.

[0049] In this embodiment, the transmitter antenna and the receiver antenna may be either directional or omnidirectional. In the case of a directional antenna, the antenna gain generally changes based on the antenna's directivity, which means that the gain changes with the position of the transmitter and receiver. In this embodiment, for a given transmit / receive link, the antenna gain is considered constant, as is the transmitter output.

[0050] In this embodiment, the correlation between local environmental factors and the variable E of the received signal intensity caused by the environment is determined by machine learning techniques.

[0051] In light of the principles of deep learning in image recognition, the regional environment can be represented as multiple 2D layers, namely f1,...,f N It can be considered a composite image consisting of these layers. Each layer can be obtained from various publicly available geographic websites, such as the OpenStreetMap project or the PLATEAU project, and can individually record information that may affect the magnitude of the received signal strength.

[0052] This composite image, consisting of multiple layers, can be considered as a feature vector of the received signal intensity on the transmission / reception link, facilitating machine learning. As a result, in this embodiment, the correlation between the composite image and the received signal intensity can be determined, similar to the principle of individually learning each RGB color in image recognition.

[0053] In addition, to avoid complex explanations, the area where the transceiver is installed is conceptualized as a composite image composed of multiple layers such as altitude layer f1, building height layer f2, terrain distribution layer f3, frequency layer f4, bandwidth layer f5, and humidity layer f6.

[0054] The altitude layer f1 records the relative altitude at each sampling point within the area observed along the line from the transmitter to the receiver. The value of the relative altitude at each sampling point is calculated by Equation (4).

[0055]

Equation

[0056] α, β, and γ represent coordinates in a coordinate system with the connection between the transmitter and the receiver as the x-axis. γ is the value of the relative altitude required in this embodiment. x, y, and z represent coordinates in the original coordinate system. The coordinates of the transmitter in the original coordinate system are represented as (x t , y t , z t ). The coordinates of the receiver are represented as (x r , y r , z r ). Therefore, cosφ, sinφ, cosθ, and sinθ are calculated by Equations (5) to (8), respectively.

[0057]

Equation

[0058]

Equation

[0059]

Equation

[0060]

number

[0061] Similarly, the building height layer f2 records the relative building height at each sampling point within the region observed along the line from the transmitter to the receiver. Note that this relative building height value at each sampling point can also be calculated using equation (4), taking into account that (x,y,z) represents the building's coordinates in the original coordinate system.

[0062] The terrain distribution layer f3 records the terrain type at each sampling point in the region. Terrain types include, for example, "rural area," "open land," "forest," and "water surface," and each type has a different impact on the characteristics of wireless communication.

[0063] The frequency layer f4 records the center frequency of the communication equipment being used. The bandwidth layer f5 records the bandwidth of the communication equipment being used.

[0064] Weather-related layers, such as the humidity layer f6, record weather and seasonal values ​​at each sampling point within the region.

[0065] In this embodiment, the composite image, consisting of multiple layers, constitutes the generated feature vectors, which, along with the received signal intensity, form a complete dataset awaiting training. Finally, the correlation φ can be trained using this dataset, either by employing the feature extraction approach proposed above or by employing transfer learning in a machine learning-based prediction method. This training process is shown by equation (9).

[0066]

number

[0067] Here, M prerepresents a pre-trained model. The value of variable E is derived from experiments. The structure of model M depends on which machine learning method is used, including support vector machines, neural networks, etc. When the model is trained according to equation (9), the received signal strength of the given transceiver line is calculated as shown in equation (10).

[0068]

number

[0069] [Schematic configuration] Referring to Figure 1, an example of a communication network to which the technical concept according to this embodiment is applied will be described. Figure 1 is a diagram showing the schematic configuration of a communication network. Multiple terminal devices 110-1...110-4 are connected to the network 190 via a base station device 120. Terminal device 110-N is connected to the network 190 via a base station device 121. These terminal devices are collectively referred to as terminal device 110. Terminal device 110 is, for example, a mobile phone or other wireless communication device. In a certain scenario, terminal devices 110-1 to 110-4 communicate wirelessly with the base station device 120 in a first area. On the other hand, terminal device 110-N communicates with a base station device 121 located in a second area, which is different from the first area.

[0070] Network 190 is further connected to an estimation device 100 that estimates the received signal. The terminal device 110 and the estimation device 100 can communicate via network 190.

[0071] Furthermore, in other scenarios, a configuration may be adopted in which an estimation circuit consisting of multiple circuit elements that realize the functions realized by the estimation device 100 is included in the terminal device 110 and base station devices 120, 121. In yet another scenario, the functions of the estimation device 100 can be realized by storing a computer program containing multiple instructions that realize the functions realized by the estimation device 100 in the terminal device 110 and base station devices 120, 121 and executing the computer program.

[0072] [Configuration of the estimation device] The configuration of the estimation device 100 will be described with reference to Figure 2. Figure 2 is a block diagram showing the hardware configuration of the computer device 200 that implements the estimation device 100. In one scenario, the estimation device 100 is implemented by a server or other computer device equipped with communication functions and arithmetic functions. In other scenarios, the estimation device 100 can be implemented as a combination of circuit elements equipped with communication circuits and arithmetic circuits, or by a program that implements communication processing and arithmetic processing being executed by a terminal device 110 or base station devices 120, 121.

[0073] The computer device 200 comprises, as its main components, a CPU (Central Processing Unit) 1 for executing programs, a mouse 2 and keyboard 3 for receiving input instructions from the user of the computer device 200, RAM 4 for volatilely storing data generated by the execution of programs by the CPU 1 or data input via the mouse 2 or keyboard 3, a hard disk 5 for non-volatilely storing data, an optical disc drive 6, a monitor 8, and a communication interface 7. Each component is connected to the others by a bus. A CD-ROM 9 or other optical disc is mounted in the optical disc drive 6. The communication interface 7 includes, but is not limited to, a USB (Universal Serial Bus) interface, a wired LAN (Local Area Network), a wireless LAN, a Bluetooth® interface, etc.

[0074] Processing in the computer device 200 is realized by each piece of hardware and software executed by the CPU 1. Such software may be pre-stored on the hard disk 5. Alternatively, the software may be stored on a CD-ROM 9 or other computer-readable non-volatile data recording medium and distributed as a program product. Or, the software may be provided as a downloadable program product by an information provider connected to the Internet or other network. Such software is read from the data recording medium by an optical disc drive 6 or other data reading device, or downloaded via the communication IF 7, and then temporarily stored on the hard disk 5. The software is then read from the hard disk 5 by the CPU 1 and stored in the RAM 4 in the form of an executable program. The CPU 1 then executes the program.

[0075] The components of the computer device 200 shown in Figure 2 are common. Therefore, it can be said that the most essential part of this embodiment is the program stored in the computer device 200. The operation of each hardware component of the computer device 200 is well known, so a detailed explanation will not be repeated.

[0076] Furthermore, the data recording medium is not limited to CD-ROMs, FDs (Flexible Disks), and hard disks, but may also be non-volatile data recording media that permanently store programs, such as magnetic tapes, cassette tapes, optical discs (MO (Magnetic Optical Disc) / MD (Mini Disc) / DVD (Digital Versatile Disc)), IC (Integrated Circuit) cards (including memory cards), optical cards, mask ROMs, EPROMs (Electronically Programmable Read-Only Memory), EEPROMs (Electronically Erasable Programmable Read-Only Memory), and semiconductor memories such as flash ROMs.

[0077] The term "program" as used here may include not only programs that can be directly executed by the CPU, but also programs in source code format, compressed programs, encrypted programs, and so on.

[0078] If the terminal device 110 or base station devices 120, 121 have communication functions, memory functions, and calculation functions implemented by the computer device 200, the estimation device 100 may also be implemented by the terminal device 110 or base station devices 120, 121.

[0079] [Estimation of received signal strength] Referring to Figure 3, a method for estimating the received signal strength according to this embodiment will be described. Figure 3 is a diagram showing a comparison between an existing method for estimating the received signal strength and the method for estimating the received signal strength according to this embodiment.

[0080] As shown in Figure 3(A), numerous measurement experiments have been conducted in multiple (for example, N measurement areas from the 1st to the Nth) measurement areas, and the longitude and latitude of the transceiver, the received signal strength, and the bit error rate have been collected for each measurement area.

[0081] In contrast, as shown in Figure 3(B), the estimation device 100 according to this embodiment uses data obtained from wireless communication in a new measurement area different from the area where wireless communication has already been performed (i.e., the N+1 measurement area, also called the second area) (latitude, longitude, altitude, terrain category, and frequency of the receiving device that performed the wireless communication) to fine-tune the already created model (i.e., a model learned from the transceiver's longitude and latitude, received signal strength, and bit error rate acquired from wireless communication in the first area), and estimates indicators representing the quality of wireless communication (for example, received signal strength, average bit error rate (BER), and indicators representing environmental impact) for the new receiving device.

[0082] [Data collection in Area 1] Referring to Figure 4, the configuration for collecting data used to train the model will be described. Figure 4 is an example diagram illustrating the overview of the communication equipment used to conduct measurement experiments in Area 1. The measurement experiments are conducted between a base station device 120 as a transmitting device and multiple terminal devices 110-1, 110-2, 110-3, and 110-4 as receiving devices. Note that the number of terminal devices 110 is not limited to those shown in Figure 4.

[0083] For example, the base station device 120 is installed in a certain location. Its location information (latitude and longitude) is determined, for example, using the GPS (Global Positioning System) or other positioning functions of the base station device 120.

[0084] Terminal devices 110-1, 110-2, 110-3, and 110-4 each receive signals from base station device 120. Each terminal device 110-1, 110-2, 110-3, and 110-4 acquires its moving speed, location information (latitude and longitude), received signal strength, and bit error rate (average BER). The acquired data is stored in estimation device 100.

[0085] [Operation of the estimation device] Referring to Figure 5, an overview of the technical concept related to this embodiment will be described. Figure 5 is a diagram illustrating an overview of the technical concept applied to the estimation device 100. The estimation device 100 includes a model training process 510 and a data acquisition process 530.

[0086] The model training process 510 includes the steps of creating a dataset 512 from past experimental data 511, creating a pre-trained model 514 using a pre-prepared learning algorithm 513 on the dataset 512, predicting the received signal strength using the pre-trained model 514 in step 515, evaluating the pre-trained model 514 using the prediction results and the actually measured received signal strength in step 516, and modifying the pre-trained model 514 using the results of the evaluation in step 517.

[0087] On the other hand, the data acquisition process 530 includes the steps of collecting received signal strength (RSS) 531 as experimental data in the second area, and performing feature extraction 533 using open data 532 to obtain altitude 534, terrain height 535, terrain distribution 536, and frequency 537 used in the wireless communication experiment. The data for altitude 534, terrain height 535, terrain distribution 536, and frequency 537 are stored in a database as a dataset 540.

[0088] The CPU 1 of the estimation device 100 performs transfer learning 520 on the pre-trained model 514 to create a trained model 560. The pre-trained model 514 is, for example, a model trained using each data acquired in the first area to estimate the received signal strength.

[0089] More specifically, CPU1 reads out altitude 534, terrain height 535, terrain distribution 536, and frequency 537 data from dataset 540. Using the read data, a pre-prepared learning algorithm 550, and a transfer-learned pre-trained model 514, CPU1 creates a trained model 560. The trained model 560 is derived by fine-tuning the pre-trained model. Using the created trained model 560, CPU1 predicts the received signal (step 570), and evaluates the model using the predicted received signal strength and the actual measurement results (step 580).

[0090] [Data structure] Referring to Figure 6, an example of the data structure of the estimation device 100 will be described. Figure 6 is a conceptual diagram showing an example of a database 600 configured on the hard disk 5 of the estimation device 100. The database 600 contains data items acquired by wireless communication experiments in the first area. More specifically, the database 600 includes an identification number 610, an elevation matrix 620, a building height matrix 630, a terrain category matrix 640, a frequency matrix 650, an additional matrix 660, an environmental impact 670, and an average BER 680.

[0091] Identification number 610 identifies terminal device 110. Elevation matrix 620 is the altitude of the sample point acquired by terminal device 110. When the latitude and longitude of the sample point are input into base map information prepared in advance for public use, the elevation is derived.

[0092] The building height matrix 630 contains height information for buildings in the city. The building height matrix 630 is derived by inputting latitude and longitude into an application prepared by "PLATEAU," a project by the Ministry of Highways and Transport to provide open data for 3D city models.

[0093] The topographic category matrix 640 contains data classifying topography by morphology, formation, or properties. The topographic category matrix 640 is derived by inputting latitude and longitude into a publicly available database.

[0094] The frequency matrix 650 contains data on the frequencies used in wireless communication between the base station device 120 and the terminal device 110. The additional matrix 660 is a reserve area for other data that can be stored.

[0095] Environmental impact 670 quantifies the environmental impact of communication between the base station equipment 120 and the terminal equipment 110. In a certain phase, environmental impact E n (dB) is calculated as equation (11), which is derived from equations (1) and (2), as follows:

[0096]

number

[0097] The average BER680 is the average bit error rate measured in wireless communication between the base station device 120 and the terminal device 110. C0 is the speed of light and f is the frequency.

[0098] [Coordinate system] The calculation of the elevation matrix will be explained with reference to Figures 7 and 8. Figure 7 is a diagram illustrating the relationship between the two coordinate systems. The first coordinate system, shown by the solid line, is defined by standard longitude and longitude and elevation. The second coordinate system, shown by the dashed line, is defined by rotating the first coordinate system by an angle θ on the surface defining longitude and latitude (also referred to as the latitude-longitude plane), and then rotating it by an angle φ around the elevation axis.

[0099] In deriving the elevation matrix, the input (the longitude x of each of the n terminal devices 110) n , latitude y n , Longitude x of base station equipment 120 BS , latitude y BS When the field of view (angle) α) is given to the numerical elevation model M, the elevation matrix An The output is as follows. The digital elevation model can be obtained from the Geospatial Information Authority of Japan's base map information download service. Angle θ, φ and the longitude x of each terminal device 110. n And, latitude y n And the longitude x of base station equipment 120 BS And, latitude y BS The relationship between this and the numerical elevation model M is as follows:

[0100]

number

[0101]

number

[0102]

number

[0103] Figure 8 shows an example of calculating the elevation of each sample point when a second coordinate system is applied to map 800. The distance d between the location Tx where the transmitting base station device 120 is located and the location Rx of the receiving terminal device 110 is calculated by the following formula.

[0104]

number

[0105] The distance d is divided into J segments. In the second coordinate system, the coordinates of location Tx of base station device 120 are (0,0,0). The coordinates of location Rx of terminal device 110 are (d,0,0). The field of view α (0≦α≦180°) is a predetermined parameter that affects the number of spatial samples. That is, the larger the value of α, the more spatial sample points there are, and the larger the dataset.

[0106] Furthermore, the number of sample points in areas other than the region between the base station device 120 and the terminal device 110, namely the region 810 behind the base station device 120 as seen from the terminal device 110's location Rx, and the region 820 behind the terminal device 110 as seen from the base station device 120, are calculated using the ceil function as N × ceil(k × J), where k > 0 and is a pre-specified parameter. In the example shown in Figure 8, the number of each sample point is 4.

[0107] Therefore, the computer device 200, which functions as the estimation device 100, can calculate the elevation matrix as follows. (Step 1) CPU1 creates sample points in RAM4 in the second coordinate system illustrated in Figure 8. (Step 2) CPU1 calculates the longitude and latitude of each sample point in the second coordinate system based on the Tx and Rx coordinates in the second coordinate system. (Step 3) CPU1 checks whether each sample point in the second coordinate system exists in the first coordinate system. If the sample point exists in the first coordinate system, CPU1 calculates the longitude and latitude of the sample point. (Step 4) CPU1 uses the latitude and longitude calculated in Step 3 above and the numerical elevation model M to determine the elevation of each sample point in the first coordinate system. (Step 5) CPU1 converts the elevation in the first coordinate system obtained in Step 4 above to the elevation in the second coordinate system, and generates the elevation matrix A n Calculate.

[0108] [Model Training] Refer to Figure 9 to explain the model training. Figure 9 is a diagram showing an example of model training. The estimation device 100 trains the model using the database shown in Figure 6. That is, for each terminal device 110, the elevation matrix A n (Elevation matrix 620 in Figure 6), building height matrix H n (Building height matrix 630), terrain category matrix C n (Geographic category matrix 640), and frequency matrix F nThe frequency matrix (650) is input to the model. At this time, the model is trained for each wireless network consisting of a transmitter and a receiver. The estimation device 100 saves the trained model to the hard disk 5.

[0109] [Estimates in Area 2] Next, we will explain a method for estimating the received signal strength in a new area, referring to Figures 10 to 13. Figure 10 shows a state in which multiple terminal devices 110 are arranged in a new area.

[0110] In Figure 10, terminal devices 110-2 (UE2), 110-3 (UE3), 110-5 (UE5), and 110-6 (UE6) are terminal devices that are the targets for estimation of received signal strength and average BER. First, the estimation device 100 collects the frequency, transmission power, and longitude of the signal transmitted from the base station device 120.

[0111] For terminal devices 110-1 (UE1) and 110-4 (UE4), their movement speed, latitude and longitude, received signal strength, and average BER are actually measured, and the measurement results are collected. The collected results are input to the estimation device 100. Next, the latitude and longitude of terminal devices 110-2, 110-3, 110-5, and 110-6 are estimated. In this case, the latitude and longitude do not need to be actually measured, and approximate values ​​may be used. The estimation results are input to the estimation device 100. The estimation device 100 uses the latitude and longitude of these terminal devices to estimate the received signal strength and average BER for each of terminal devices 110-2, 110-3, 110-5, and 110-6.

[0112] [Dataset] The dataset created in the estimation device 100 will be described with reference to Figure 11. Figure 11 shows an example of a dataset 1100 configured on the hard disk 5 of the computer device 200, which functions as the estimation device 100. In addition to the configuration of the database 600 shown in Figure 6, the dataset 1100 includes data items from terminal devices 110-5 and 110-6.

[0113] In other words, for all terminal devices, the elevation matrix, building height matrix, terrain category matrix, and frequency matrix are calculated using the method described above.

[0114] For terminal devices 110-1 (UE1) and 110-4 (UE4), the environmental impact 670 and average BER are calculated using conventional methods. On the other hand, for terminal devices 110-2 (UE2), 110-3 (UE3), 110-5 (UE5), and 110-6 (UE6), the environmental impact 670 and average BER are estimated using a trained model that has been fine-tuned according to this embodiment.

[0115] [Fine-tuning the model] Refer to Figure 12 to explain the fine-tuning of the transferred model. Figure 12 is a diagram illustrating the fine-tuning of the model.

[0116] For terminal devices 110-1 (UE1) and 110-4 (UE4), the elevation matrix 620, building height matrix 630, terrain category matrix 640, and frequency matrix 650, calculated from data acquired through actual wireless communication in the second area, are input into the transferred model. The transferred model is fine-tuned as it learns from this data. Environmental impact 670 or average BER 680 are also input into the model.

[0117] [Estimation using a finely tuned model] Refer to Figure 13 to explain the estimation using the finely tuned model. Figure 13 is a diagram showing the procedure for estimating the environmental impact 670 or average BER 680 using the estimation device 100.

[0118] The estimation device 100 inputs the elevation matrix 620, building height matrix 630, terrain category matrix 640, and frequency matrix 650 of terminal devices 110-2 (UE2), 110-3 (UE3), 110-5 (UE5), and 110-6 (UE6) into a finely tuned model to calculate an index indicating the environmental impact 670. Furthermore, the estimation device 100 can calculate the received signal strength using equation (16) based on equation (11).

[0119]

number

[0120] Here, n = 2, 3, 5, 6. C0 is the speed of light, and f is the frequency. <Examples of application> The following describes an example of the application of this embodiment. To maintain generality, we will consider a case where a signal receiver is to be installed on the roof of a building (referred to as the receiving building).

[0121] In this application example, the receiver (terminal device 110) needs to receive signals transmitted from a fishing ground (base station device 120) located on the coast far from the receiving building. Due to the complex buildings and natural environment surrounding the receiving building, it is impossible to transmit signals directly between the transmitter and the receiver. Therefore, it is necessary to install a relay station in an appropriate location. The location of this relay station must ensure high received signal strength for receiving information from vessels in the fishing ground, while also guaranteeing high received signal strength to the receiver. The relay station installed in such a location receives signals from vessels in the fishing ground and transmits the received signals to the receiving building, thereby realizing relay communication between the vessels in the fishing ground and the receiving building.

[0122] This example is not limited to vessels in fishing grounds; conducting similar experiments in other regions would allow us to accumulate a considerable amount of experimental data and experience.

[0123] In this application example, in order to install a relay station, it is necessary to first find a suitable location within a predetermined distance (for example, a radius of 10 km) from the receiving building. Since this is a very time-consuming and costly task, it is desirable to predict the received signal strength at different locations as accurately as possible using various methods and identify the optimal installation location.

[0124] Until now, the simplest method has been to predict received signal strength using empirical models discussed in Non-Patent Documents 1, 2, and 3. While the calculations are straightforward, the accuracy of the results is questionable. To improve accuracy, computer simulations, such as those described in Non-Patent Document 4, can also be used to predict received signal strength. This method can improve accuracy, but it requires modeling cities within a specified range and is time-consuming. Overall, this method is not cost-effective.

[0125] Currently, machine learning-based methods, such as those discussed in Patent Document 1 or Non-Patent Documents 5, 6, and 7, are widely used to predict received signal strength at various locations. These methods may have many advantages, such as low computational cost and relatively high accuracy. On the other hand, they also have disadvantages, such as the need to collect a large amount of experimental data in advance.

[0126] [Effects of application examples] In contrast, this application example can effectively address some of the difficulties encountered when putting machine learning-based methods into practical use. First, data previously collected in other regions is used to generate a model with high generalization ability. This highly generalized model can then be applied to the cases considered through transfer learning. Therefore, the method following this application example requires only the collection of a small amount of data in advance for fine-tuning, and the generalized model can be adapted when a new case is considered. Such an approach can significantly reduce the need for large-scale data collection.

[0127] Secondly, the feature extraction approach following this application example can effectively address the influence of both static and dynamic information (including geographical information and seasonal variations) on communication characteristics such as received signal strength, thereby significantly improving prediction accuracy.

[0128] Finally, the collected data and extracted feature vector information, if considered, can further enrich the database, thus contributing to the growth and scalability of the technical concept according to this embodiment.

[0129] <Effects of the Embodiment> The embodiments described above provide a technology that enhances the accuracy, robustness, and generalizability of RSS prediction models in outdoor wireless communication systems. Specifically, the accuracy, robustness, and generalizability of RSS prediction models in outdoor wireless communication systems are enhanced by incorporating information on topographic features, land use patterns, and geographic topology.

[0130] More specifically, firstly, a feature extraction approach that considers transfer learning methods and geographical knowledge is employed in machine learning-based received signal strength prediction. This eliminates the need for large, representative training datasets to train an accurate model. It also eliminates the need to collect and annotate received signal strength measurements across various geographical regions and environmental conditions, which is usually computationally intensive and time-consuming.

[0131] Secondly, this embodiment addresses the problem that machine learning models cannot be generalized to various regions due to changes in topography, vegetation, and urban form. As a result, the need for repeated data collection and annotation is eliminated, improving the scalability, applicability, and real-world effectiveness of the machine learning-based method for predicting received signal strength.

[0132] Thirdly, this embodiment can address the challenges arising from the complex interactions between environmental factors and signal propagation dynamics. This ensures better generalization of the model, even when dealing with dense foliage in rural areas or canyons in urban areas with towering skyscrapers, and allows for effective management of the impact of topographic elevation differences and signal intensity.

[0133] Finally, this embodiment can address the challenges posed by the dynamic nature of outdoor wireless channels, which conventional machine learning algorithms fail to do because they assume stationarity of the data distribution. The method according to this embodiment adapts to changes in environmental conditions such as seasonal changes and urban development, reducing the need for continuous model retraining.

[0134] The embodiments disclosed herein should be considered in all respects to be illustrative and not restrictive. The scope of the present invention is indicated by the claims rather than by the foregoing description, and all modifications within the meaning and scope equivalent to the claims are intended to be included. [Explanation of Symbols]

[0135] 1 CPU, 2 Mouse, 3 Keyboard, 4 RAM, 5 Hard disk, 6 Optical disc drive, 7 Communication interface, 8 Monitor, 9 CD-ROM, 100 Estimation device, 110, 110-1, 110-2, 110-3, 110-4, 110-5, 110-6 Terminal device, 120, 121 Base station device, 190 Network, 200 Computer device.

Claims

1. A method performed on a computer using a model for estimating the quality of wireless communication between a transmitting device and a plurality of receiving devices, wherein the dataset used by the model includes an elevation matrix, a building height matrix, a terrain category matrix, a frequency matrix, and a numerical value representing the quality, and the method is: A step of accessing a first model obtained by learning data obtained by wireless communication in a first region, The steps include: performing wireless communication between the transmitting device and a first receiving device among the plurality of receiving devices in a second region different from the first region; The steps include creating the dataset using the results of wireless communication between the transmitting device and the first receiving device, The steps include: adjusting the first model using the aforementioned dataset to derive a second model; A method comprising the step of estimating an index representing the quality of wireless communication between a transmitter and a second receiver, using an elevation matrix, a building height matrix, a terrain category matrix, a frequency matrix, and the second model obtained for the second receiver among the plurality of receivers.

2. The method according to claim 1, wherein the indicator representing the quality of the wireless communication includes either the received signal strength or a numerical value representing the influence of the environment in which the wireless communication is performed.

3. A method performed on a computer to create a database used for estimating the received signal strength of wireless communication between a transmitting device and a receiving device, wherein the database includes an elevation matrix, a building height matrix, a terrain category matrix, a frequency matrix, and numerical values ​​representing environmental impacts, and the method is: A step of calculating an elevation matrix using the latitude and longitude of the transmitting device, the latitude and longitude of the receiving device, and an elevation model prepared in advance for deriving altitude. A step of calculating a building height matrix using the latitude and longitude of the transmitting device, the latitude and longitude of the receiving device, and a building height model prepared in advance for deriving the building height. The steps include: calculating a terrain category matrix using the latitude and longitude of the transmitting device, the latitude and longitude of the receiving device, and a terrain category model prepared in advance for deriving terrain categories; The steps include: calculating a frequency matrix using the frequency of the signal transmitted by the transmitting device; A method comprising the step of calculating a numerical value representing the environmental impact based on the received signal strength measured by the receiving device, the transmission output of the transmitting device, the antenna gains of the transmitting device and the receiving device, the frequency, and the spatial distance between the transmitting device and the receiving device.

4. A program that causes a computer to perform the method described in any one of claims 1 to 3.

5. A memory containing a program that causes a computer to execute the method described in any one of claims 1 to 3, A device comprising a processor that executes the aforementioned program.

Images