Communication quality prediction device, communication quality prediction method, and program

Abstract

To provide a communication quality prediction device capable of highly accurately predicting a communication quality with a small number of samples, a communication quality prediction method, and a program.SOLUTION: A communication quality prediction device comprises: a background point group generation section 11 for generating a point group under an environment that a pedestrian is not present; a pedestrian point group generation section 15 for generating a point group of a pedestrian being present within a region R1; a pedestrian point group processing section 31 for generating a deformed pedestrian point group which is obtained by deforming the pedestrian point group; and a communication quality calculation section 35 for calculating a first communication quality at a time when the deformed pedestrian point group is operated. Further, the communication quality prediction device comprises: a simulation point group generation section 34 for generating a simulation point group which is obtained by combining a background point group and a deformed pedestrian point group within the region R1; a first dataset generation section 36 for generating a first dataset which is a pair of the simulation point group and the first communication quality; a training section 37 for training a prediction model using the first dataset; and a communication quality prediction section 70 for predicting a communication quality within the region R1 using the trained prediction model.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a communication quality prediction device, a communication quality prediction method, and a program. [Background technology]

[0002] The recent increase in the Internet, IoT, and M2M traffic has led to a shortage of wireless communication bandwidth, and the use of higher frequencies is being considered. Next-generation mobile communications (5G, 6G) are expected to realize high-speed, large-capacity communications using frequencies above 30 GHz (millimeter waves).

[0003] Non-Patent Document 1 mentions that communications using high frequencies above the sub-6 GHz band are strongly affected by the surrounding environment. Specifically, it shows that in millimeter wave communications and terahertz wave communications, radio wave shielding by human bodies and other objects can cause a sudden drop in communication quality. For this reason, it is necessary to detect large changes in communication quality in advance based on the presence and movement of pedestrians and take measures such as transmission power control and handover.

[0004] Non-Patent Document 2 discloses that the communication path state is predicted in advance using machine learning based on measurement data obtained by measurement devices such as an RGB camera, a depth camera, and LiDAR, and transmission power control and handover are performed.

[0005] Non-Patent Document 3 discloses a pre-learning method that reduces the number of samples used in communication quality prediction and shortens the time required to acquire the samples by using a handmade model to perform radio wave propagation simulation and transfer learning. Specifically, a data set that pairs spatial information and communication quality to be used in pre-learning is generated using computer simulation, and a communication quality prediction model is trained using this data set. The pre-trained prediction model is subjected to transfer learning using a data set obtained at a location where the system is actually used so that the model is adapted to that location.

[0006] Non-Patent Document 4 discloses that segmentation is performed by applying a deep learning model such as "PointNet" to extract data on parts that affect communication quality, such as humans and vehicles. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] S. Collonge, G. Zaharia, and GE Zein, “Influence of the human activityon wide-band characteristics of the 60 GHz in-door radiochannel,” IEEE Trans. WirelessCommun., vol.3, no.6,pp.2396-2406, Nov. 2004. [Non-patent document 2] S. Ohta, T. Nishio, R. Kudo, K. Takahashi, and H. Nagata, “Point cloud-based proactive link quality prediction for millimeter-wave communications,” arXiv preprint, vol. 2301.00752, pp. 1-19, Jan. 2023. [Non-patent document 3] T. Mikuma,T. Nishio, M. Morikura, K. Yamamoto, Y. Asai and R. Miyatake, "Transfer Learning-Based ReceivedPower Prediction UsingRGB-D Camera in mmWave Networks," Proceedings of IEEE VTC-Spring, KualaLumpur, Malaysia, 2019,pp. 1-5. [Non-patent document 4] Qi, CharlesR., et al. "Pointnet: Deep learning on point sets for 3d classification and segmentation." Proceedings of the IEEE conference on computer visionand pattern recognition. 2017. Summary of the Invention [Problem to be solved by the invention]

[0008] However, in the method of predicting communication quality using radio wave propagation simulation and transfer learning on a handmade model described in Non-Patent Document 3, it is difficult to reproduce spatial information (for example, point cloud data) acquired from an actual sensor. As a result, it is not possible to generate a dataset with high accuracy, and high effectiveness of pre-training cannot be expected.

[0009] Furthermore, the techniques disclosed in Non-Patent Documents 2 and 3 are simulations for generating depth images, and do not mention generating dynamic spatial information for a communication quality prediction device. Furthermore, Non-Patent Document 4 does not disclose generating a dataset to be used for pre-training a prediction model.

[0010] The present disclosure has been made in consideration of the above circumstances, and its purpose is to provide a communication quality prediction device, a communication quality prediction method, and a program that are capable of predicting communication quality with high accuracy using a small number of samples. [Means for solving the problem]

[0011] A communication quality prediction device according to one embodiment of the present disclosure includes a static spatial information generation unit that generates static spatial information in an environment in which no moving objects are present within the region, a dynamic spatial information generation unit that generates dynamic spatial information of moving objects present within the region, a deformation information generation unit that generates deformation information by deforming the dynamic spatial information, a communication quality calculation unit that calculates a first communication quality when the deformation information is activated in the environment of the static spatial information, a first spatial information generation unit that generates first spatial information by combining the static spatial information within the region and the deformation information, a first dataset generation unit that generates a first dataset that is a pair of the first spatial information and the first communication quality, a training unit that uses the first dataset to train a prediction model of communication quality within the region, and a communication quality prediction unit that predicts communication quality within the region using the prediction model trained by the training unit.

[0012] A communication quality prediction method according to one embodiment of the present disclosure includes a static spatial information generation unit that generates static spatial information in a static environment where no moving objects are present within the area, a dynamic spatial information generation unit that generates dynamic spatial information of moving objects present within the area, a deformation information generation unit that generates deformation information by deforming the dynamic spatial information, a communication quality calculation unit that calculates a first communication quality when the deformation information is operated in the environment of the static spatial information, a first spatial information generation unit that generates first spatial information by combining the static spatial information within the area with the deformation information, a first dataset generation unit that generates a first dataset that is a pair of the first spatial information and the first communication quality, a training unit that trains a prediction model of the communication quality using the first dataset, and a communication quality prediction unit that predicts the communication quality within the area using the prediction model trained by the training unit.

[0013] One aspect of the present disclosure is a program for causing a computer to function as the communication quality prediction device. [Effects of the Invention]

[0014] According to the present disclosure, it is possible to predict communication quality with high accuracy using a small number of samples. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is an explanatory diagram showing a communication quality prediction device according to an embodiment and its surrounding environment. [Figure 2] FIG. 2 is a block diagram showing a detailed configuration of the prediction model learning unit 60. As shown in FIG. [Figure 3] FIG. 3 is an explanatory diagram showing a background point cloud, a real space point cloud, a difference point cloud, and a pedestrian point cloud. [Figure 4A] FIG. 4A is an explanatory diagram showing the results of a simulation performed on a simulation mesh using the ray tracing method, and shows the results when there is good visibility between the two communication devices. [Figure 4B] FIG. 4B is an explanatory diagram showing the results of a simulation performed on a simulation mesh using the ray tracing method, and shows the results when some object blocks the path between two communication devices. [Figure 5] FIG. 5 is a table showing line-of-sight determination, signal strength, and minimum path loss in the environments shown in FIGS. 4A and 4B. [Figure 6] FIG. 6 is a flowchart illustrating a processing procedure of the communication quality prediction device according to the embodiment. [Figure 7] Figure 7 is a graph showing the change in communication quality prediction error after 500 [ms] when pre-training was performed using a simulation data set of 105.5 samples. [Figure 8] FIG. 8 is a block diagram showing the hardware configuration of this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0016] Hereinafter, an embodiment will be described with reference to the drawings. Fig. 1 is an explanatory diagram showing a communication quality prediction device 100 according to an embodiment and its surrounding environment. The communication quality prediction device 100 according to this embodiment predicts communication quality that changes within an area R1 shown in Fig. 1, for example, based on the movement status of stationary objects and moving objects such as pedestrians present within this area R1.

[0017] 1, the communication quality predicting device 100 includes a prediction model learning unit 60 and a communication quality predicting unit 70. The communication quality predicting device 100 is connected to a communication quality sensor 51, a point cloud sensor 52, an environmental information collecting unit 53, a point cloud database 55, and a control device 54.

[0018] The communication quality sensor 51 is connected to a communication device AP that performs wireless communication within the area R1, and detects the communication quality when the communication device AP performs communication. The communication quality includes received power, throughput, and the like.

[0019] The point cloud sensor 52 is, for example, an RGB camera or LiDAR, and detects stationary objects and moving objects such as pedestrians p present in the region R1 and collects point cloud data (spatial information) thereof. The point cloud sensor 52 outputs the collected point cloud data to the prediction model learning unit 60 and the point cloud database 55.

[0020] The environmental information collection unit 53 acquires surrounding environmental information based on point cloud data acquired by the LiDAR of the point cloud sensor 52 and images acquired by the RGB camera. The environmental information collection unit 53 acquires environmental information such as weather and seasonal information. The environmental information collection unit 53 outputs the acquired environmental information to the point cloud database 55 and the prediction model learning unit 60.

[0021] The point cloud database 55 stores the point cloud data collected by the point cloud sensor 52 and the environmental information acquired by the environmental information acquisition unit 53 .

[0022] The communication device AP communicates with a plurality of terminals ST1 to STN within the area R1.

[0023] The communication quality sensor 51 measures the communication quality when the communication device AP is performing communication. The communication quality sensor 51 outputs the measured communication quality to the prediction model learning unit 60.

[0024] The control device 54 controls communication by the communication device AP.

[0025] FIG. 2 is a block diagram showing the detailed configuration of the prediction model learning unit 60. As shown in FIG. 2, the prediction model learning unit 60 includes an analysis unit 1, a quality measurement unit 2, a simulation unit 3, and a transfer unit 4. Note that "n" shown on the connecting lines between the blocks in FIG. 2 indicates the number of samples acquired in the real space and used for fine tuning (details will be described later). Also, "m" indicates the number of samples generated by simulation and used for pre-training. The number of samples m is, for example, 10 2 ≦m≦10 5.5 is.

[0026] The analysis unit 1 includes a background point cloud generation unit 11, a background mesh generation unit 12, a real space point cloud acquisition unit 13, a difference point cloud generation unit 14, a pedestrian point cloud generation unit 15, and a storage unit 16.

[0027] The background point cloud generating unit 11 extracts a point cloud (hereinafter referred to as a "background point cloud") when no moving object such as a pedestrian p is present in the region R1, based on the point clouds stored in the point cloud database 55.

[0028] The background point cloud generation unit 11 obtains, for example, the background point cloud shown in Fig. 3(a). The background point cloud is a point cloud observed by the point cloud sensor 52 in an environment where no moving objects such as a pedestrian p exist within the region R1. The background point cloud does not include a point cloud due to moving objects such as the pedestrian p. The background point cloud generation unit 11 functions as a static spatial information generation unit that generates spatial information (point cloud) in an environment where no moving objects such as a pedestrian p exist within the region R1.

[0029] The background mesh generation unit 12 performs a plane detection process based on the background point cloud acquired by the background point cloud generation unit 11 to generate mesh data (hereinafter referred to as a "background mesh"). The background mesh generation unit 12 outputs the generated background mesh to the simulation mesh generation unit 33 of the simulation unit 3.

[0030] The real space point cloud acquisition unit 13 acquires a point cloud detected by the point cloud sensor 52 within the region R1 shown in FIG. 1. This point cloud includes not only the background point cloud described above but also a point cloud due to moving objects such as pedestrians p that exist within the region R1. Hereinafter, a point cloud including the background point cloud and the point cloud due to moving objects will be referred to as a "real space point cloud." The real space point cloud acquisition unit 13 can acquire, for example, a real space point cloud as shown in FIG. 3(b). The real space point cloud acquisition unit 13 has a function as a space information measurement unit that measures space information within the region R1.

[0031] The difference point cloud generation unit 14 generates a difference point cloud based on the difference between the above-mentioned real space point cloud and the background point cloud. The process of calculating the difference point cloud will be described below. The difference point cloud generation unit 14 extracts a point cloud of a moving object existing in the region R1 by calculating the difference between the background point cloud generated by the background point cloud generation unit 11 and the real space point cloud acquired by the point cloud sensor 52. That is, the difference point cloud shown in FIG. 3(c) is generated by subtracting the background point cloud shown in FIG. 3(a) from the real space point cloud shown in FIG. 3(b).

[0032] The Octree difference detection method can be used to calculate the difference point cloud. In addition to the Octree difference detection method, it is also possible to apply a deep learning model such as "PointNet" to the point cloud to perform segmentation and extract point cloud data that affects communication quality, such as people and vehicles.

[0033] The pedestrian point cloud generation unit 15 generates a pedestrian point cloud by clustering the difference point cloud generated by the difference point cloud generation unit 14. Specifically, a point cloud for one pedestrian (pedestrian point cloud) is extracted.

[0034] The pedestrian point cloud generation unit 15 acquires information about the number of pedestrians based on the number of clusters. A wide range of clustering methods, such as "DBSCAN" and "k-means algorithm," can be used. The clustered pedestrian point cloud can be tracked for a short period of time, such as 30 to 60 seconds, to acquire information about the pedestrian's movement range and speed range. That is, the pedestrian point cloud generation unit 15 acquires the number of pedestrians present in the region R1, their movement range, and speed range based on the pedestrian point cloud. As a result, for example, the pedestrian point cloud shown in FIG. 3(d) is generated. The pedestrian point cloud generation unit 15 functions as a dynamic spatial information generation unit that generates spatial information (point cloud) of moving objects such as pedestrian p present in the region R1.

[0035] The memory unit 16 stores, as pedestrian information, the data on the number of pedestrians, their movement range, and their speed range acquired by the pedestrian point cloud generation unit 15. The pedestrian information stored in the memory unit 16 is read by the pedestrian point cloud processing unit 31 of the simulation unit 3.

[0036] The quality measurement unit 2 includes a communication quality acquisition unit 21 and a second data set generation unit 22.

[0037] The communication quality acquisition unit 21 acquires the communication quality within the region R1 based on the communication quality of the communication device AP detected by the communication quality sensor 51 (see FIG. 1). The process of acquiring communication quality in the real space will be described below. The communication quality acquisition unit 21 acquires the communication quality (e.g., power consumption, throughput) of the real space acquired by the communication quality sensor 51 for only a short period of about 30 seconds. The communication quality acquisition unit 21 performs scaling so that the average of the communication quality is "0" and the standard deviation is "1". The ranges of the values ​​of the communication quality in the real space and the communication quality obtained by simulation (processing by the communication quality calculation unit 35, described later) often differ. For this reason, scaling is performed to match the ranges of the values.

[0038] The second dataset generation unit 22 generates a second dataset which is a pair of a real space point cloud (measured value of spatial information) acquired by the point cloud sensor 52 and a communication quality (measured value of communication quality) acquired by the communication quality acquisition unit 21. That is, the second dataset generation unit 22 generates a second dataset which is a pair of a measured value of spatial information and a measured value of communication quality in real space. The second dataset generation unit 22 outputs the generated second dataset to the transfer unit 4. Thereafter, the time series of the real space point cloud is associated with future communication quality values ​​using "temporal difference labeling."

[0039] The above process generates a second data set used to predict future communication quality. This second data set is used for fine tuning performed by the fine tuning unit 41 of the transfer unit 4.

[0040] The simulation unit 3 includes a pedestrian point cloud processing unit 31, a pedestrian mesh generation unit 32, a simulation mesh generation unit 33, a simulation point cloud generation unit 34, a communication quality calculation unit 35, a first dataset generation unit 36, and a training unit 37.

[0041] The pedestrian point cloud processing unit 31 acquires the pedestrian point cloud generated by the pedestrian point cloud generation unit 15 and the pedestrian information stored in the storage unit 16. Based on the pedestrian point cloud and pedestrian information, the pedestrian point cloud processing unit 31 applies geometric processing such as translation and rotation to the pedestrian point cloud to reproduce the movement of pedestrians. By repeating this geometric processing, pedestrian point clouds for various pedestrian movements are acquired. Hereinafter, the pedestrian point cloud that has been subjected to geometric processing will be referred to as a "deformed pedestrian point cloud." The pedestrian point cloud processing unit 31 functions as a deformation information generation unit that generates a transformed pedestrian point cloud (deformation information) by transforming the pedestrian point cloud (dynamic spatial information).

[0042] The pedestrian point cloud processing unit 31 outputs the acquired transformed pedestrian point cloud to the pedestrian mesh generation unit 32 and the simulation point cloud generation unit .

[0043] The pedestrian mesh generation unit 32 applies surface reconstruction processing to the transformed pedestrian point cloud to generate mesh data (hereinafter referred to as "pedestrian mesh"). In other words, the pedestrian mesh refers to data obtained by converting the transformed pedestrian point cloud into mesh data. The pedestrian mesh generation unit 32 outputs the generated pedestrian mesh to the simulation mesh generation unit 33.

[0044] The simulation mesh generation unit 33 acquires the background mesh generated by the background mesh generation unit 12 and the pedestrian mesh generated by the pedestrian mesh generation unit 32, and synthesizes them to generate a simulation mesh. The simulation mesh generation unit 33 outputs the generated simulation mesh to the communication quality calculation unit 35.

[0045] The simulation point cloud generation unit 34 generates a simulation point cloud by combining the background point cloud acquired by the background point cloud generation unit 11 and the transformed pedestrian point cloud. The simulation point cloud generation unit 34 outputs the generated simulation point cloud to the first data set generation unit 36.

[0046] Furthermore, the simulation point cloud generator 34 generates a point cloud that reflects changes in the surrounding environment. In the real space, the number of pedestrians in the area R1 may change over time from the physical space environment acquired over a short period of time, and the environment, such as visibility and roadside trees, may change due to weather and seasonal changes. By generating a simulation scenario in advance regarding these environmental changes over time and using it for data acquisition and learning, flexibility to adapt to environmental changes is improved. Environmental information such as weather and seasonal changes can be acquired from the environmental information collector 53 shown in Figure 1.

[0047] As a specific example, the simulation point cloud generating unit 34 executes the following processes (1) to (3). (1) For objects such as human bodies and vehicles, the objects are replicated using existing 3D models of people and vehicles or those scanned by the analysis unit 1, and placed in the scanned environment. (2) Regarding weather, in the radio wave propagation simulation, a radio wave propagation model according to the weather, for example, a radio wave propagation model for rainy or snowy weather, is applied. (3) For objects that change with the seasons, such as roadside trees, a general point cloud model is prepared for each season, for example, for deciduous trees, with leaves in summer and leaves removed in winter, and placed as an object.

[0048] By employing the above-described processing, it is possible to generate a variety of simulation scenarios that adapt to the state of the real space and correspond to future changes in the real space, and generate a simulation point cloud. The simulation point cloud generator 34 functions as a first space information generator that generates first space information (simulation point cloud) by combining the background point cloud (static space information) within the region R1 and the transformed pedestrian point cloud (transformation information).

[0049] The communication quality calculation unit 35 performs a simulation using a ray tracing method based on the simulation mesh generated by the simulation mesh generation unit 33, and calculates the communication quality when the simulation mesh is used (hereinafter referred to as "first communication quality"). The communication quality calculation unit 35 has a function of calculating the first communication quality when the transformed pedestrian point cloud is operated in the environment of the background point cloud (static space information). Furthermore, the communication quality calculation unit 35 calculates the first communication quality by expanding the transformed pedestrian point cloud (deformation information) within a range in which it can move within the region R1 in which the background point cloud (static space information) is set.

[0050] By performing a ray-tracing simulation on the simulation mesh, labels are assigned to the simulation point cloud. The materials used in the ray-tracing simulation of radio wave propagation can be set according to the results of point cloud processing. Specifically, if a pedestrian is detected as a result of segmentation, the pedestrian is assigned "water," and if a wall is detected, the wall is assigned "concrete," "wood," etc.

[0051] Note that the amount of calculation required for segmentation can be excessive, so if you want to speed up the process, you can simplify it by uniformly setting all materials to "concrete," "wood," etc. In particular, in high-frequency band communications such as sub-6 GHz and millimeter waves, the influence of direct waves is large, while the influence of waves that reach the terminal after multiple diffractions is small, making simplification effective. Also, since simulations take a long time when the space is large, it is possible to shorten the calculation time by setting a sufficiently large threshold for the distance from the terminal or access point and installing walls in those locations.

[0052] 4A and 4B are explanatory diagrams showing the results of a simulation performed using the ray tracing method on a simulation mesh, with Fig. 4A showing the results when there is good visibility between the two communication devices Q1 and Q2, and Fig. 4B showing the results when the space between the two communication devices Q1 and Q2 is blocked by an object (for example, an object obtained by meshing the pedestrian point cloud shown in Fig. 3(d)). Note that the object is not limited to a pedestrian, but also includes animals, vehicles, etc.

[0053] Figure 5 is a table showing the relationship between the visibility judgment result, communication quality, and minimum path loss. As shown in Figure 5, when the visibility between the two communication devices Q1 and Q2 is good, the communication quality is high and the minimum path loss is low.

[0054] Returning to FIG. 2, the communication quality calculation unit 35 outputs data on the calculated communication quality to the first data set generation unit 36.

[0055] The first dataset generation unit 36 ​​generates a first dataset representing a pair of the simulation point cloud generated by the simulation point cloud generation unit 34 and a simulated communication quality obtained by scaling the first communication quality calculated by the communication quality calculation unit 35. For example, the first dataset generation unit 36 ​​generates the first dataset by performing "temporal difference labeling." That is, the first dataset generation unit 36 ​​has a function of generating a first dataset representing a pair of the simulation point cloud (first spatial information) and the first communication quality obtained as a result of the simulation. The first dataset generation unit 36 ​​outputs the generated first dataset to the training unit 37.

[0056] The training unit 37 performs pre-training on the prediction model of communication quality using the first data set to generate a pre-trained model. That is, the training unit 37 has a function of training the prediction model of communication quality within the region R1 using the first data set. The training unit 37 outputs the pre-trained model to the fine-tuning unit 41 of the transfer unit 4.

[0057] The transfer unit 4 includes a fine tuning unit 41 and a deployment unit 42.

[0058] The fine-tuning unit 41 performs fine-tuning on the prediction model trained by the training unit 37, using the second dataset generated by the second dataset generation unit 22. Specifically, the second dataset in the real space is divided into a training set and a validation set, and the training set is used to train the prediction model for communication quality.

[0059] The fine-tuning unit 41 applies inverse scaling to the fine-tuned prediction model based on the real-space communication quality of the training set, thereby enabling it to fit the real-space point cloud data and transfer the prediction model to the real space.

[0060] Inverse scaling using a small number of samples may not realize the true range of communication quality in the real world. Furthermore, communication quality may change over time if the communication environment changes slightly. In such cases, it is best to determine the range of communication quality in advance based on the communication standard. For example, the maximum physical transmission rate in IEEE 802.11ad is set as the maximum throughput, and 0 [bit / s] is set as the minimum throughput. This makes it possible to predict values ​​that are in line with the communication standard.

[0061] In addition, the error of the inversely scaled model is evaluated using the validation set. If the error is large, the hyperparameters used in fine tuning, such as the learning rate and number of training epochs in machine learning, are adjusted and tuning is performed again. If the predictive model still does not achieve the desired accuracy after retuning, the use of the predictive model is temporarily discontinued.

[0062] The communication system continues to operate, and after the number of samples in the second data set in the real world is increased, the above tuning is performed again and verification is performed. As a result, the communication quality prediction model is used for communication control only if the desired performance is met. This makes it possible to avoid degradation of communication performance due to communication control based on inaccurate communication quality predictions.

[0063] The deployment unit 42 deploys the communication quality prediction model trained by the fine tuning unit 41.

[0064] 1 predicts the communication quality in the region R1 by inputting the environmental information acquired by the environmental information collection unit 53 into the prediction model deployed by the deployment unit 42. That is, the communication quality prediction unit 70 has a function of predicting the communication quality in the region R1 using the prediction model trained by the training unit 37.

[0065] Next, the processing procedure of the communication quality prediction device 100 according to this embodiment configured as described above will be described with reference to the flowchart shown in Fig. 6. First, in step S11, the background point cloud generation unit 11 generates a background point cloud (static spatial information) of the region R1 based on the point clouds stored in the point cloud database 55.

[0066] In step S12, the pedestrian point cloud processing unit 31 generates a deformed pedestrian point cloud based on the pedestrian information stored in the storage unit 16 and the pedestrian point cloud generated by the pedestrian point cloud generating unit 15.

[0067] In step S13, the communication quality calculation unit 35 calculates the communication quality (first communication quality) when the deformed pedestrian point cloud is moved within the region R1.

[0068] In step S14, the simulation point cloud generation unit 34 generates a simulation point cloud by combining the deformed pedestrian point cloud and the background point cloud.

[0069] In step S15, the first data set generator 36 generates a first data set that is a pair of a simulation point cloud and a first communication quality.

[0070] In step S16, the training unit 37 trains a prediction model for communication quality using the first data set. The training unit 37 outputs the trained prediction model to the fine tuning unit 41.

[0071] In step S17, the second data set generator 22 generates a second data set that is a pair of the measurement value of the spatial information and the measurement value of the communication quality.

[0072] In step S18, the fine tuning unit 41 fine-tunes the prediction model trained in the process of step S16 using the second data set.

[0073] In step S19, the communication quality prediction unit 70 predicts the communication quality within the region R1 using the fine-tuned prediction model. The communication quality prediction unit 70 outputs the predicted communication quality data to the outside. In this way, the communication quality within the region R1 can be predicted using a small number of samples.

[0074] The inventors conducted experimental evaluation using a LiDAR point cloud and a 60 GHz band millimeter wave throughput data set. 5.5 This is a graph showing the change in communication quality prediction error after 500 [ms] when pre-trained with a sample simulation dataset.

[0075] Curve s1 shows the case where pre-training was not performed, curve s2 shows the case where pre-training was performed using "Sim2Real," and curve s3 shows the case where pre-training was performed using "Real2Sim2Real." As can be seen from curves s1 to s3 in Fig. 7, by adopting this embodiment, it was possible to reduce by 33% the number of fine-tuning samples acquired in real space that were required to reduce the root-mean-squared error (RMSE) below 0.27 [Gbit / s].

[0076] The vertical axis of FIG. 7, 0.27 [Gbit / s], represents the error value reached by learning with 3000 samples without pre-training, and it was shown that by adopting the communication quality prediction device 100 according to this embodiment, high communication quality prediction accuracy can be achieved with a small number of fine-tuning samples.

[0077] As described above, the communication quality predicting device 100 according to this embodiment includes a static spatial information generating unit (background point cloud generating unit 11) that generates spatial information (background point cloud) in an environment where no moving body (e.g., pedestrian) exists in the region R1, a dynamic spatial information generating unit (pedestrian point cloud generating unit 15) that generates spatial information (pedestrian point cloud) of moving bodies that exist in the region R1, a deformation information generating unit (pedestrian point cloud processing unit 31) that generates deformation information (deformed pedestrian point cloud) that has been modified by modifying the dynamic spatial information, and a calculation unit that calculates a first communication quality when the deformation information is operated in an environment of the static spatial information. a first spatial information generation unit (simulation point cloud generation unit 34) that generates first spatial information (simulation point cloud) that combines static spatial information and deformation information within the region R1; a first dataset generation unit 36 ​​that generates a first dataset that is a pair of the first spatial information and the first communication quality; a training unit 37 that trains a prediction model for communication quality within the region R1 using the first dataset; and a communication quality prediction unit 70 that predicts communication quality within the region R1 using the prediction model trained by the training unit 37.

[0078] In this embodiment, a data set consisting of pairs of point cloud data and communication quality is generated using a small number of samples, and a prediction model for communication quality is trained using these data sets. Therefore, it is possible to predict communication quality within the region R1 with high accuracy using a small number of samples. This makes it possible to reduce communication traffic when transmitting data from the point cloud sensor 52 to the communication quality prediction device 100.

[0079] In this embodiment, a first data set for pre-training a communication quality prediction model is generated using the transformed pedestrian point cloud generated by the pedestrian point cloud processing unit 31. That is, the transformed pedestrian point cloud is subjected to geometric processing such as translation and rotation based on the movement range and movement speed acquired from the pedestrian information. As a result, a first data set including various conditions of moving objects within the region R1 can be generated.

[0080] By training a prediction model using the first dataset generated in this way, it is possible to build a prediction model that can predict communication quality with high accuracy. Furthermore, by using the first dataset, it is possible to predict communication quality with high accuracy even when the amount of data acquired in the real space is small.

[0081] In this embodiment, a second data set is generated, which is a pair of spatial information measurements within region R1 and communication quality measurements, and the prediction model trained by training unit 37 is fine-tuned using the second data set, making it possible to construct a prediction model that can predict communication quality with higher accuracy.

[0082] In this embodiment, the first communication quality described above is calculated by a simulation using the ray tracing method, so that it is possible to generate a more reliable first data set.

[0083] In this embodiment, the communication quality calculation unit 35 calculates the first communication quality by expanding the deformed pedestrian point cloud (deformation information) within the range of possible movement within the region R1 where the static space information is set, thereby making it possible to generate the first data set making maximum use of the background point cloud within the region R1. As a result, it becomes possible to further improve the prediction accuracy of the communication quality.

[0084] The communication quality predicting device 100 of the present embodiment described above can be implemented as a general-purpose computer system including, for example, a CPU (Central Processing Unit, processor) 901, a memory 902, a storage 903 (HDD: Hard Disk Drive, SSD: Solid State Drive), a communication device 904, an input device 905, and an output device 906, as shown in Fig. 8. The memory 902 and the storage 903 are storage devices. In this computer system, the CPU 901 executes a predetermined program loaded on the memory 902, thereby realizing each function of the communication quality predicting device 100.

[0085] The communication quality predicting device 100 may be implemented in one computer or in multiple computers, or may be a virtual machine implemented in a computer.

[0086] The program for the communication quality prediction device 100 can be stored in a computer-readable recording medium such as an HDD, an SSD, a USB (Universal Serial Bus) memory, a CD (Compact Disc), or a DVD (Digital Versatile Disc), or can be distributed via a network. The computer-readable recording medium is, for example, a non-transitory recording medium.

[0087] The present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the present disclosure. [Explanation of symbols]

[0088] 1 Analysis section 2 Quality measurement department 3 Simulation Section 4 Transition zone 11 Background point cloud generation unit (static spatial information generation unit) 12 Background mesh generation section 13 Real space point cloud acquisition unit 14 Difference point cloud generator 15 Pedestrian point cloud generation unit 16 Memory section 21 Communication quality acquisition unit 22 Second Data Set Generation Unit 31 Pedestrian point cloud processing unit (deformation information generation unit) 32 Pedestrian mesh generation unit 33 Simulation mesh generation unit 34 Simulation point cloud generation unit (first spatial information generation unit) 35 Communication quality calculation unit 36 First Data Set Generation Unit 37 Training Department 41 Fine Tuning Section 42 Deployment Department 51 Communication Quality Sensor 52 Point Cloud Sensor 53 Environmental Information Collection Department 54 Control device 55 Point Cloud Database 60 Prediction model learning unit 70 Communication Quality Prediction Department 100 Communication quality prediction device AP communication device

Claims

1. a static space information generating unit that generates static space information in an environment in which no moving object exists within the area; a dynamic space information generating unit that generates dynamic space information of a moving object present in the area; a deformation information generation unit that generates deformation information by modifying the dynamic spatial information; a communication quality calculation unit that calculates a first communication quality when the deformation information is operated in an environment of the static space information; a first spatial information generation unit that generates first spatial information by combining the static spatial information within the region and the deformation information; a first data set generator that generates a first data set that is a pair of the first spatial information and the first communication quality; a training unit that uses the first data set to train a prediction model of communication quality within the region; a communication quality prediction unit that predicts communication quality within the area using the prediction model trained by the training unit; A communication quality prediction device comprising:

2. a spatial information measurement unit that measures spatial information within the region; a communication quality acquisition unit that acquires communication quality within the area; a second data set generation unit that generates a second data set that is a pair of the measurement value of the spatial information and the measurement value of the communication quality; a fine-tuning unit that fine-tunes the prediction model trained by the training unit using a second data set; The communication quality prediction device according to claim 1 , further comprising:

3. The communication quality calculation unit calculates the first communication quality using a ray tracing method. The communication quality prediction device according to claim 1 or 2.

4. The communication quality calculation unit calculates the first communication quality by expanding the deformation information within a range of movement possible within the area in which static space information is set. The communication quality prediction device according to claim 1 or 2.

5. a static space information generating unit generating static space information in a static environment in which no moving object exists within the area; a dynamic space information generating unit generating dynamic space information of a moving object existing within the area; a deformation information generation unit that generates deformation information by modifying the dynamic spatial information; a communication quality calculation unit calculating a first communication quality when the deformation information is operated in an environment of the static space information; a first spatial information generation unit generating first spatial information by combining the static spatial information and the deformation information within the region; a first data set generation unit generates a first data set that is a pair of the first spatial information and the first communication quality; a training unit that uses the first data set to train the communication quality prediction model; A communication quality prediction unit predicts communication quality within the area using the prediction model trained by the training unit. Communication quality prediction method.

6. A program that causes a computer to function as the communication quality prediction device according to claim 1.

Images