JP7366790B2 - Communication status prediction device, communication status prediction method, and program - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act On December 20, 2020, IEEE-INTERNATIONAL CONFERENCE ON IOT EMBEDDED SYSTEMS AND COMMUNICATIONS (IINTEC) 2019 IINTEC'19 1570 Published at 604245

The present invention relates to communication status of wireless communication and technology for prediction processing of the communication state.

In small spaces such as factories, unexpected communication quality (QoS: Quality of Service) deterioration occurs due to changes in the propagation environment due to the movement of machine tools, products, workers, etc., and interference between multiple wireless applications. Productivity will decrease.

Therefore, it is necessary to take avoidance action before extreme deterioration of QoS (communication failure) occurs, and to achieve this, it is necessary to predict the occurrence of communication failure. By predicting communication failures, it is possible to prevent a decline in productivity by taking the following measures, for example.
(1) When a sign of a communication failure is detected in the communication link of a certain wireless terminal, the reduction in transmission throughput can be prevented by stopping other wireless terminals using the same wireless channel or changing the usage band. To prevent.
(2) When a sign of a communication failure is detected on a certain communication link, the band of the terminal causing the interference is automatically limited to prevent a drop in transmission throughput or the occurrence of large latency. .

For example, Patent Documents 1 and 2 disclose techniques for predicting communication quality at a future time based on past information.

Japanese Patent Application Publication No. 2016-91271 JP 2017-5473 Publication

When using the techniques of Patent Documents 1 and 2 to detect signs of communication failure, a two-choice determination is made as to whether or not the predicted value satisfies the required QoS value, or the degree of similarity between the required QoS value and the predicted value (e.g. , the Euclidean distance between the required QoS value and the predicted value) to quantify and determine the risk of communication failure. When making predictions based on past information in this way, it is inevitable that frequent input patterns can be predicted with high accuracy, but the prediction accuracy deteriorates for rare input patterns or unlearned input patterns. For this reason, when the above-mentioned technique is used, a problem arises in that communication failures may be undetected or erroneously detected.

SUMMARY OF THE INVENTION In view of the above problems, the present invention provides a communication status prediction device and communication status prediction device that performs highly accurate communication failure prediction processing even when a rare input pattern or an unlearned input pattern occurs. The purpose is to realize a method and a program.

In order to solve the above problems, a first invention includes a learning data generation processing section, a learning section, a storage section, a predicted value acquisition section, a prediction error distribution acquisition section, and a risk value acquisition section. This is a communication status prediction device equipped with a communication status prediction device.

The learning data generation processing unit acquires the degree of abnormality of the time-series pattern of the communication quality data, identifies excluded data that is data to be excluded from the learning process according to the acquired degree of abnormality, and generates the communication quality data. Data obtained by excluding the excluded data from the time-series pattern is obtained as data to be subjected to learning processing.

The learning section performs learning processing using the learning processing target data acquired by the learning data generation processing section. The learning unit then executes a learning process using the generation pattern of the time-series pattern of communication quality data and the communication quality data at a future time after the time-series pattern occurs.

The storage unit stores a rarity calculation data set that includes occurrence patterns of time-series patterns of the communication quality data acquired by the learning unit, and communication quality data at a future time that is a predetermined time after the occurrence of the time-series patterns. Store trained data including a label data set including

The predicted value acquisition unit acquires a predicted value of communication quality data after a predetermined time from a time-series pattern of communication quality data input at the current time based on learned data.

The prediction error distribution acquisition unit acquires a prediction error distribution that is a distribution of prediction error values that are the difference between the predicted value of the communication quality data acquired by the predicted value acquisition unit and the true value of the predicted value.

The risk value acquisition unit acquires a risk value indicating the possibility that a communication failure will occur at a future time based on the prediction error distribution and the prediction value.

This communication status prediction device acquires a risk value indicating the possibility of a communication failure occurring at a future time based on the prediction error distribution and the predicted value. The possibility of occurrence can be appropriately grasped. In addition, with this communication status prediction device, by processing as described above, it is possible to obtain risk values as continuous values or multivalued data, so unlike the conventional technology, it is possible to obtain risk values as binary data indicating the presence or absence of a communication failure. It is possible to achieve a much more accurate prediction process than the determination process using .

Furthermore, this radio situation prediction device detects abnormal patterns in observed data (for example, QoS values) based on the degree of abnormality, and generates learning data excluding the detected abnormal patterns. Since this radio situation prediction device executes learning processing using the learning data, appropriate learning processing is executed and an appropriate error distribution of QoS values can be obtained. As a result, this radio situation prediction device 100 can appropriately avoid outputting abnormal data as a prediction result.

A second invention is the first invention, in which the learning data generation processing unit sets an average vector acquisition period, sets a unit section including a predetermined number of communication quality data, and sets a unit section including a predetermined number of communication quality data, When communication quality data is time series data a _i (i: natural number), a vector v_a _i whose elements are a plurality of time series data a _i and an average vector v_a _ave of time series data included in the average vector acquisition period. The norm of the difference vector between the vector v_a _i and the average vector v_a _ave is obtained as the degree of abnormality.

Thereby, this communication status prediction device can obtain the norm of the difference vector between the vector v_a _i and the average vector v_a _ave as the degree of abnormality.

A third invention is the first or second invention, in which the learning data generation processing unit compares the degree of abnormality with a predetermined threshold, and when the degree of abnormality is greater than the threshold, the degree of abnormality is A time-series pattern of corresponding communication quality data is identified as excluded data.

Thereby, in this communication status prediction device, it is possible to determine a time series pattern whose degree of abnormality is larger than a predetermined threshold value as an abnormal pattern, and to obtain learning data excluding the time series pattern.

A fourth invention is the first or second invention, wherein the learning data generation processing unit sets an average vector acquisition period and performs a sorting process on the abnormality degrees acquired in the average vector acquisition period. The time-series pattern of the communication quality data corresponding to the abnormality degree included in the top x% (x: real number) from the highest abnormality degree value is specified as excluded data.

As a result, this communication status prediction device uses learning data that excludes time-series patterns of communication quality data corresponding to abnormalities included in the top x% (x: real number) from the highest abnormality degree value. can be obtained.

Note that x in "Top x%" is preferably a real number of 10 or less, for example.

A fifth invention is a communication status prediction method comprising a learning data generation processing step, a learning step, a storage step, a predicted value acquisition step, a prediction error distribution acquisition step, and a risk value acquisition step.

The learning data generation processing step acquires the degree of abnormality of the time-series pattern of the communication quality data, identifies excluded data that is data to be excluded from the learning process according to the acquired degree of abnormality, and generates the communication quality data. Data obtained by excluding the excluded data from the time-series pattern is obtained as data to be subjected to learning processing.

In the learning step, learning processing is performed using the learning processing target data acquired in the learning data generation processing step. In the learning step, a learning process is executed using the generation pattern of the time-series pattern of communication quality data and the communication quality data at a future time after the time-series pattern occurs.

The storage step includes a rarity calculation data set that includes the generation pattern of the time series pattern of the communication quality data acquired in the learning step, and communication quality data at a future time that is a predetermined time after the occurrence of the time series pattern. Store trained data including a label data set including

The predicted value acquisition step acquires a predicted value of communication quality data after a predetermined time from a time-series pattern of communication quality data input at the current time based on learned data.

The prediction error distribution acquisition step acquires a prediction error distribution that is a distribution of prediction error values that are the difference between the predicted value of the communication quality data acquired in the predicted value acquisition step and the true value of the predicted value.

The risk value acquisition step acquires a risk value indicating the possibility that a communication failure will occur at a future time based on the prediction error distribution and the prediction value.

Thereby, it is possible to realize a communication status prediction method that has the same effects as the first invention.

A sixth invention is a program for causing a computer to execute the communication situation prediction method according to the fifth invention.

Thereby, it is possible to realize a program for causing a computer to execute a communication situation prediction method that has the same effects as the first invention.

According to the present invention, there is provided a communication status prediction device and a communication status prediction method that perform highly accurate communication failure prediction processing even when a rare input pattern or an unlearned input pattern occurs. program can be realized.

FIG. 1 is a schematic configuration diagram of a radio situation prediction device 100 (an example of a communication situation prediction device) according to a first embodiment. FIG. 1 is a schematic configuration diagram of a learning data generation processing unit 1 of a radio situation prediction device 100 according to a first embodiment. FIG. 3 is a schematic configuration diagram of a prediction determination processing section 3 of the radio situation prediction device 100 according to the first embodiment. FIG. 3 is a diagram showing time series data (an example) of QoS values acquired by the radio situation prediction device 100; FIG. 3 is a diagram for explaining average vector acquisition processing executed by the radio situation prediction device 100. FIG. 3 is a diagram for explaining an abnormality degree calculation process and a learning data generation process executed by the radio situation prediction device 100. 3 is a diagram showing learning data D_train (an example) acquired by the learning data generation processing unit 1 of the radio situation prediction device 100. FIG. 2 is a diagram showing observation data Din (an example) of time-series data of QoS values acquired by the radio situation prediction device 100 and learning data D_train (an example) acquired by the learning data generation processing unit 1. FIG. FIG. 3 is a diagram showing time series data (an example) of QoS values acquired by the radio situation prediction device 100; FIG. 2 is a diagram for explaining a first step process executed during a learning period in the radio situation prediction device 100. FIG. FIG. 7 is a diagram for explaining a second step process executed during a learning period in the radio situation prediction device 100. FIG. 7 is a diagram for explaining a third step of processing executed during a learning period in the radio situation prediction device 100. 6 is a diagram for explaining a prediction process executed during a prediction process execution period in the radio situation prediction device 100. FIG. FIG. 3 is a diagram for explaining prediction processing (communication failure prediction execution processing). FIG. 3 is a diagram showing a CPU bus configuration.

[First embodiment]
The first embodiment will be described below with reference to the drawings.
<1.1: Configuration of radio situation prediction device>
FIG. 1 is a schematic configuration diagram of a radio situation prediction device 100 (an example of a communication situation prediction device) according to the first embodiment.

FIG. 2 is a schematic configuration diagram of the learning data generation processing unit 1 of the radio situation prediction device 100 according to the first embodiment.

FIG. 3 is a schematic configuration diagram of the prediction determination processing section 3 of the radio situation prediction device 100 according to the first embodiment.

As shown in FIG. 1, the radio situation prediction device 100 includes an antenna Ant1, a demodulation processing unit Dm1, an observation data acquisition unit FE1, a control unit CPU1, a selector SEL0, a learning data generation processing unit 1, and a learning data generation processing unit 1. unit 2, a learned data storage unit St1, and a prediction determination processing unit 3.

Antenna Ant1 is an antenna for receiving an external RF signal (for example, a carrier band OFDM signal). The antenna Ant1 outputs the received RF signal (for example, carrier band OFDM signal) to the demodulation processing unit Dm1 as a signal RFin. Note that the antenna Ant1 may be an antenna (transmission/reception antenna) that also functions as a transmission antenna.

The demodulation processing section Dm1 includes an RF processing section Dm11 and a BB processing section Dm12.

The RF processing unit Dm11 performs RF demodulation processing on the signal RFin and obtains a baseband signal.

The BB processing unit Dm12 executes baseband demodulation processing on the baseband signal demodulated by the RF processing unit Dm11, obtains the signal (data) after the baseband demodulation processing as data D0, and obtains the obtained data D0. is output to the observation data acquisition unit FE1.

The observation data acquisition unit FE1 inputs the data D0 output from the demodulation processing unit Dm1, and acquires a QoS value, which is data indicating the QoS of the wireless communication environment in which the radio situation prediction device 100 is placed, from the data D0. do. Then, the observation data acquisition unit FE1 outputs the data including the acquired QoS value as data Din0 to the learning data generation processing unit 1 and the selector SEL0.

Note that the QoS value is an index value indicating communication quality such as throughput, number of (continuous) packet losses, amount of delay, etc.

The control unit CPU1 is a functional unit for controlling each functional unit of the radio situation prediction device 100. The control unit CPU 1 transmits a mode signal Mode including data indicating the processing mode (for example, (1) learning mode, (2) prediction processing execution mode) of the radio situation prediction device 100 to the learning data generation processing unit 1 and the learning unit 2. , the prediction determination processing section 3, and the selector SEL0. Further, the control unit CPU1 outputs a control signal ctl1 to the prediction determination processing unit 3.

As shown in FIG. 2, the learning data generation processing section 1 includes a data holding section 11, an average vector acquisition section 12, an abnormality degree calculation section 13, and a learning data generation section 14. The learning data generation processing section 1 receives a mode signal output from the control section CPU1, and executes processing when the mode signal is a signal indicating a learning mode.

The data holding unit 11 receives data Din0 from the observation data acquisition unit FE1. The data holding unit 11 stores and holds the input data Din0 for a predetermined period of time. The data holding unit 11 outputs time-series data of the data Din0, which is the data for which the average vector is calculated, to the average vector acquisition unit 12 as data D_ptn. If data Din0 at time i is expressed as a _i , data D_ptn is, for example, Na (Na: natural number) continuous time series data D_ptn={a _i , a _i+1 , ..., a _i+Na-1 } is set as

In addition, the data holding unit 11 uses the set data v_a _j consisting of a plurality of pieces of time series data a _i to be subjected to the abnormality degree calculation process as data D_ptn (v_a j ), and uses the data D_ptn (v_a _j ) for the abnormality degree calculation unit 13 and the learning data generation. output to section 14. In addition, the set data v_a _j is
v_a _j ={a _j , a _j+1 ,..., a _j+Nb-1 }
Nb: Set as the number of time series data a _i to be subjected to the abnormality degree calculation process.

The average vector acquisition unit 12 receives the data D_ptn output from the data holding unit 11 and acquires an average vector for the data D_ptn (details will be described later). Then, the average vector acquisition unit 12 outputs the acquired average vector to the abnormality degree calculation unit 13 as data v_a _ave .

The abnormality degree calculation unit 13 receives data D_ptn(v_a _j ) output from the data holding unit 11 and data v_a _ave output from the average vector acquisition unit 12. The abnormality degree _calculation unit 13 calculates the abnormality degree of the data v_a j using the data D_ptn(v_a _j ) and the data v_a _ave (details will be described later). Then, the abnormality degree calculation unit 13 outputs the calculated data including the abnormality degree for the data v_a _j to the learning data generation unit 14 as data deg_abnml(v_a _j ).

The learning data generation section 14 receives data D_ptn (v_a _j ) output from the data holding section 11 and data deg_abnml (v_a _j ) output from the abnormality degree calculation section 13 . The learning data generating unit 14 determines whether or not to include the data D_ptn(v_a _j ) in the learning data based on the degree of abnormality deg_abnml(v_a _j ) of the data D_ptn(v_a _j ). Then, the learning data generation unit 14 generates data (collective data of time series data a _i ) that includes only the data determined to include the data D_ptn (v_a _j ) in the learning data, and uses the generated data. The data is output as learning data D_train to the learning section 2 and selector SEL0.

The selector SEL0 is a selector with two inputs and one output, and selects and outputs one of the two inputs according to the mode signal Mode output from the control unit CPU1. Specifically, when the mode signal Mode is a signal indicating the learning mode, the selector SEL0 selects the data D_train output from the learning data generation processing section 1, and outputs it to the prediction determination processing section 3 as data Din. do. On the other hand, when the mode signal Mode is a signal indicating the prediction processing execution mode, the selector SEL0 selects the data Din0 output from the observation data acquisition section FE1, and outputs it to the prediction determination processing section 3 as data Din.

The learning unit 2 inputs data D_train output from the learning data generating unit 14. The learning unit 2 executes a learning process for executing a pattern matching prediction process using a PNN (Probabilistic Neural Network) or the like on the data D_train, that is, time series data of QoS values, and obtains learned data. . As learned data, for example, (1) a data set for Rarity calculation and (2) a data set for future value prediction are acquired (details will be described later). The learning unit 2 stores the acquired trained data as data D_trained in the trained data storage unit St1.

The learned data storage section St1 stores learned data output from the learning section 2. Further, the trained data storage section St1 reads data in accordance with a request from the prediction determination processing section 3, and outputs the read data (for example, data D_trained) to the prediction determination processing section 3.

As shown in FIG. 3, the prediction determination processing section 3 includes a data holding section 31, a rarity acquisition section 32, a predicted value acquisition section 33, a subtractor 34, a prediction error distribution acquisition section 35, and a risk value acquisition section. 36.

The data holding unit 31 receives data Din from the selector SEL0. The data holding unit 31 stores and holds the input data Din for a predetermined period of time. The data holding unit 31 outputs the time series data of the data Din to the rarity acquisition unit 32 and the predicted value acquisition unit 33 as data D1s. If data Din at time i is expressed as D _i , then data D1s is set as, for example, three consecutive time series data D1s={D _i , D _i+1 , D _i+2 }. Note that the number of data of the time series data D1s may be any number other than 3 (k1 (k1: natural number) continuous time series data (D1s={D _i , D _i+1 , . . . D _i+k1-1 }).

Furthermore, assuming that the predicted time is time i+k (k: natural number), the data holding unit 31 outputs actual data Din (=D _i+k ) (true value) at time i+k to the subtracter 34 as data D1_true.

As shown in FIG. 3, the rarity acquisition unit 32 includes a norm acquisition unit 321 and a rarity calculation unit 322.

The norm acquisition unit 321 receives the learned data D_trained outputted from the trained data storage unit St1 and the data D1s outputted from the data holding unit 31. The norm acquisition unit 321 calculates a norm (for example, Euclidean distance) using the learned data D_trained and the data D1s, and outputs data including the calculated norm to the Rarity calculation unit 322 as data D_norm. Note that when there are n1 calculated norms (n1: natural number), the data D_norm is data of a set including n1 norms (d ₀ , d ₁ , . . . , d _n1-1 ). At this time, the data D_norm is
D_norm={d ₀ , d ₁ ,..., d _n1-1 }
It is written as.

The Rarity calculation unit 322 inputs the data D_norm output from the norm acquisition unit 321, and calculates the rarity (Rarity) indicating the degree of rarity of appearance of the time series data D1s from the data D_norm. Then, the Rarity calculation unit 322 outputs data including the calculated rarity to the prediction error distribution acquisition unit 35 as data D2.

The predicted value acquisition unit 33 inputs the trained data D_trained outputted from the trained data storage unit St1 and the data D1s outputted from the data holding unit 31. The predicted value acquisition unit 33 performs a prediction process (for example, a prediction process using PNN) using the learned data D_trained and the data D1s, and acquires the value of the predicted data at a future time as data D_predict. Then, the predicted value acquisition unit 33 outputs the acquired data D_predict to the subtractor 34 and the risk value acquisition unit 36.

The subtracter 34 inputs the data D1_true output from the data holding section 31 and the data D_predict output from the predicted value acquisition section 33. The subtractor 34 is
D_prediction_error=D1_true-D_predict
A process corresponding to (subtraction process) is executed to obtain prediction error data D_prediction_error. Then, the subtracter 34 outputs the acquired prediction error data D_prediction_error to the prediction error distribution acquisition unit 35.

As shown in FIG. 3, the prediction error distribution acquisition unit 35 includes a selector SEL1, a Rarity distribution acquisition unit 351, a class classification processing unit 352, a first storage unit St2, a Rarity class classification unit 353, and a prediction error distribution. It includes a data acquisition section 354, a second storage section St3, and a class determination section 355.

The selector SEL1 is a 1-input, 3-output selector, and inputs the data D2 output from the rarity calculation unit 322 of the rarity acquisition unit 32 and the control signal ctl1 output from the control unit CPU1. Selector SEL1 outputs input data D2 to any one of (1) Rarity distribution acquisition section 351, (2) Rarity class classification section 353, and (3) class determination section 355 according to control signal ctl1.
The Rarity distribution acquisition unit 351 receives data D2 output from the selector SEL1. The Rarity distribution acquisition unit 351 acquires a Rarity distribution from the data D2, and outputs the acquired Rarity distribution data to the class classification processing unit 352 as data Dist_Rarity.

The class classification processing unit 352 inputs the data Dist_Rarity output from the rarity distribution acquisition unit 351. The class classification processing unit 352 executes the class classification process based on the data Dist_Rarity, acquires the set data Th of threshold values for classifying the class, and stores the acquired set data Th of the threshold values in the first storage unit St2. .

In addition, when the number of thresholds required for class classification is n2 (n2: natural number), the set data Th of thresholds is calculated by dividing n2 thresholds (th ₀ , th ₁ , ..., th _n2-1 ). This is the data of the containing set. At this time, the threshold set data Th is
Th={th ₀ , th ₁ ,..., th _n2-1 }
It is written as.

The first storage unit St2 stores the threshold set data Th output from the class classification processing unit 352. Further, the first storage unit St2 reads data in accordance with a data read request from the Rarity class classification unit 353, and outputs the data to the Rarity class classification unit 353. Further, the first storage unit St2 reads data in accordance with a data read request from the class determination unit 355, and outputs the data to the class determination unit 355.

The Rarity class classification unit 353 receives the data D2 output from the selector SEL1 and the threshold set data Th output from the first storage unit St2. The Rarity class classification unit 353 classifies the data D2 based on the threshold set data Th, and outputs data including the number j of the class to which the data D2 is assigned as data D3 to the prediction error distribution data acquisition unit 354. do.

The prediction error distribution data acquisition unit 354 inputs the data D_prediction_error output from the subtracter 34 and the data D3 output from the Rarity class classification unit 353. The prediction error distribution data acquisition unit 354 acquires prediction error distribution data by sorting the data D_prediction_error into classes based on the data D3 and acquiring the occurrence frequency distribution of the data D_prediction_error for each sorted class. Then, the prediction error distribution data acquisition unit 354 stores data including the acquired prediction error distribution data Err_dist _j as data D4 in the second storage unit St3. Note that prediction error distribution data of prediction errors classified into the class of class number j is expressed as Err_dist _j .

The second storage unit St3 stores data D4 (=Err_dist _j ) output from the prediction error distribution data acquisition unit 354. Further, the second storage unit St3 reads data in accordance with a data read request from the risk value acquisition unit 36, and outputs the data to the risk value acquisition unit 36.

The class determination unit 355 receives data D2 output from the selector SEL1 and threshold set data Th output from the first storage unit St2. The class determination unit 355 classifies the data D2 based on the threshold set data Th, and classifies the class number j (=class (r _i ), class (x) is a function for acquiring the class number of x) is output to the risk value acquisition unit 36.

The risk value acquisition unit 36 receives the data D_predict output from the predicted value acquisition unit 33, the class number j output from the class determination unit 355, and the data D4 (=Err_dist _j ) output from the second storage unit St3. , data QoS_req indicating a required reference value of communication quality data, and a mode signal Mode output from the control unit CPU1 are input. When the mode signal Mode is a signal indicating "prediction processing execution mode", the risk value acquisition unit 36 determines the time (in the future) at which the true value of the data D_predict will be acquired based on the prediction error distribution data D4 (=Err_dist _j ). obtain a risk value indicating the possibility of a communication failure occurring at the specified time). Then, the risk value acquisition unit 36 outputs data including the acquired risk value as data Dout.

Note that the data QoS_req indicating the request standard value of communication quality data is data that is preset in the radio situation prediction device 100 (for example, data that is preset in a storage unit (not shown) in the radio situation prediction device 100 and stored and held). data) or data set from outside the radio situation prediction device 100.

<1.2: Operation of radio situation prediction device>
The operation of the radio situation prediction device 100 configured as above will be described below.

FIG. 4 is a diagram illustrating observation data Din (an example) of time-series data of QoS values acquired by the radio situation prediction device 100. In the graph of FIG. 4, the horizontal axis is time, and the vertical axis is QoS value.

FIG. 5 is a diagram for explaining the average vector acquisition process executed by the radio situation prediction device 100.

FIG. 6 is a diagram for explaining the abnormality degree calculation process and the learning data generation process executed by the radio situation prediction device 100.

FIG. 7 is a diagram showing learning data D_train (an example) acquired by the learning data generation processing unit 1 of the radio situation prediction device 100. In the graph of FIG. 4, the horizontal axis is time, and the vertical axis is QoS value.

FIG. 8 is a diagram showing observation data Din (an example) of time-series data of QoS values acquired by the radio situation prediction device 100 and learning data D_train (an example) acquired by the learning data generation processing unit 1. It is. In the graph of FIG. 8, the horizontal axis is time, and the vertical axis is QoS value.

FIG. 9 is a diagram showing time-series data (an example) of QoS values acquired by the radio situation prediction device 100. In the graph of FIG. 9, the horizontal axis is time and the vertical axis is QoS value.

FIG. 10 is a diagram for explaining the first step process executed during the learning period in the radio situation prediction device 100.

FIG. 11 is a diagram for explaining the second step process executed during the learning period in the radio situation prediction device 100.

FIG. 12 is a diagram for explaining the third step process executed during the learning period in the radio situation prediction device 100.

FIG. 13 is a diagram for explaining the prediction process executed in the prediction process execution period in the radio situation prediction device 100.

(1.2.1: Learning data generation process)
First, the learning data generation process will be explained.

The demodulation processing unit Dm1 of the radio situation prediction device 100 performs demodulation processing on the radio signal received via the antenna Ant1, and obtains an RF demodulated signal and a BB demodulated signal. The observation data acquisition unit FE1 extracts QoS values (indicators indicating communication quality such as throughput, number of (continuous) packet losses, amount of delay, etc.) from the RF demodulated signal and/or the BB demodulated signal acquired by the demodulation processing unit Dm1. value). Then, the observation data acquisition unit FE1 outputs data including the acquired QoS value to the learning data generation processing unit 1 as data Din0.

The control unit CPU1 sets the mode signal Mode to a signal value indicating “learning mode” and outputs it to the learning data generation processing unit 1.

The data holding unit 11 of the learning data generation processing unit 1 receives data Din0 from the observation data acquisition unit FE1. The data holding unit 11 stores and holds the input data Din0 for a predetermined period of time. The data holding unit 11 sets the time-series data of the data Din0, which is the data for which the average vector is calculated, as data D_ptn (={a _i , a _i+1 , ..., a _i+Na-1 }), It is output to the average vector acquisition section 12.
In addition, the data holding unit 11 uses the set data v_a _j consisting of a plurality of pieces of time series data a _i to be subjected to the abnormality degree calculation process as data D_ptn (v_a j ), and uses the data D_ptn (v_a _j ) for the abnormality degree calculation unit 13 and the learning data generation. output to section 14. In addition, the set data v_a _j is
v_a _j ={a _j , a _j+1 ,..., a _j+Nb-1 }
Nb: Set as the number of time series data a _i to be subjected to the abnormality degree calculation process.

The average vector acquisition unit 12 receives the data D_ptn output from the data holding unit 11 and acquires an average vector for the data D_ptn. Specifically, the average vector acquisition unit 12 executes the process as follows and acquires the average vector for the data D_ptn.
(1) Obtain k1 time series observation value series (time series data of QoS values) from time i (corresponding to time t0 in Figure 5) (k1 = 3 in Figure 5), and A vector v_a _i whose elements are the QoS values of is obtained. Note that for convenience of explanation, the following explanation will be based on the assumption that k1=3, but k1 is not limited to "3" and may be any other natural number.

If k1=3,
v_a _i =(a _i , a _i+1 , a _i+2 )
It is. For example, the average vector acquisition unit 12 sets the period Prd1 (period from time t0 to t1) in FIGS. When the time series data of values) are {a ₀ , a ₁ , ..., a _Na-1 , a _Na , a _Na+1 }, the vectors v_a ₀ to v_a _Na-1 for calculating the average vector are shown in FIG. As shown in
v_a ₀ = (a ₀ , a ₁ , a ₂ )
v_a ₁ = (a ₁ , a ₂ , a ₃ )
v_a ₂ = (a ₂ , a ₃ , a ₄ )
...
v_a _Na-1 = (a _Na-1 , a _Na , a _Na+1 )
becomes.

｜Ｖ｜：集合Ｖの位数（要素数）
なお、集合Ｖは、ＱｏＳ値の時系列データａ_ｉ添え字ｉについての集合であり、図５の場合、集合Ｖ＝｛０，１，・・・，Ｎａ－１｝（ｉのとり得る数の集合）であり、その要素数は、Ｎａである。 Then, the average vector acquisition unit 12 acquires the average vector v_a _ave by averaging the vectors v_a ₀ to v_a _Na−1 for calculating the average vector, that is, by executing the process corresponding to the following formula.

|V|: Order of set V (number of elements)
The set V is a set of QoS value time series data a _i subscript i, and in the case of FIG. ), and the number of elements thereof is Na.

The average vector acquisition unit 12 outputs the average vector v_a _ave acquired through the above process to the abnormality degree calculation unit 13.

The abnormality degree calculation unit 13 receives data D_ptn(v_a _j ) output from the data holding unit 11 and data v_a _ave output from the average vector acquisition unit 12. The abnormality degree calculating _unit 13 calculates the abnormality degree of the data v_a j using the data D_ptn(v_a _j ) and the data v_a _ave . Specifically, the abnormality degree calculation unit 13 calculates the norm of the difference between the vector v_a _j and the average vector v_a _ave (Euclidean distance between the vector v_a _j and the average vector v_a _ave ) s _j . In other words, the abnormality degree calculation unit 13
s _j = Norm(v_a _j −V_a _ave )
Norm(x): The norm of the difference between the vector v_a _j and the average vector v_a _ave (Euclidean distance between the vector v_a _j and the average vector v_a _ave ) s _j by a process equivalent to a function that obtains the norm of the vector x. Calculate.

Then, the abnormality degree calculation unit 13 calculates the norm of the difference between the calculated vector v_a _j and the average vector v_a _ave (Euclidean distance between the vector v_a _j and the average vector v_a _ave ) s _j as an abnormality value for the data v_a _j . degree.

Then, the abnormality degree calculation unit 13 outputs data including the calculated abnormality degree s _j for the data v_a _j to the learning data generation unit 14 as data deg_abnml (v_a _j ).

The learning data generation section 14 receives data D_ptn (v_a _j ) output from the data holding section 11 and data deg_abnml (v_a _j ) output from the abnormality degree calculation section 13 . The learning data generating unit 14 determines whether or not to include the data D_ptn(v_a _j ) in the learning data based on the degree of abnormality deg_abnml(v_a _j ) of the data D_ptn(v_a _j ). Specifically, the learning data generation unit 14 compares the degree of abnormality s _j for the data v_a _j with a predetermined threshold Th0, and if (1) s _j ≦Th0, the data v_a _j is set as the learning data If it is determined that the data v_a j is included in the data, and (2) s _j >Th0, it is determined that the data v_a _j is not included in the learning data. For example, in the case of FIG. 6, the abnormality degree s _j of the observed value series (QoS data values) included in the region R0 (the region containing the observed value series with a downwardly convex spike-like waveform pattern) is lower than the threshold Th0. It's also big. In other words, as shown in Figure 6,
Norm(v_a _j+1 -v_a _ave )=s _j+1 >Th0
Norm(v_a _j+2 -v_a _ave )=s _j+2 >Th0
The abnormality degree s _i for the vector v_a _i other than the above is less than the threshold Th0. Therefore, the learning data generation unit 14 generates the time series data {a _{j+1 ,} a _{j+2 ,} a j _{+ 3} }, and {a _j+2 , a _{j+ 3} , a _{j+ 4 }} is determined to be an abnormal pattern, and the time series data {a j _{+ 1} , a _j+2 , a _j+3 , a _j+4 } are not included in the learning data. Then, the learning data generation unit 14 includes only the time series data determined not to be an abnormal pattern in the learning data.

In other words, the learning data generation unit 14 uses the time series data of period Prd1 in FIG. 6 excluding the time series data {a _j+1 , a _j+2 , a _j+3 , a _j+4 } as learning data. .

That is, the learning data generation unit 14 generates data (collective data of time series data a _i ) that includes only data determined to include data D_ptn(v_a _j ) in the learning data, and uses the generated data to It is output to the learning unit 2 as learning data D_train (for period Prd1, time series data shown in FIG. 7).

In the learning data generation processing unit 1, the same process as described above is executed for the periods Prd2 and Prd3, and the learning data D_train is acquired.

For example, the above process is executed by the learning data generation processing unit 1 from the data Din0 input to the learning data generation unit 14 shown in the upper diagram of FIG. 8, so that the learning data shown in the lower diagram of FIG. D_train is obtained. As can be seen from FIG. 8, the data obtained by deleting the data in the region having the downwardly convex spike-like waveform pattern is the learning data D_train.

In this way, by using data that excludes highly abnormal patterns (time-series patterns of QoS values) as learning data, the radio situation prediction device 100 can perform appropriate learning processing and obtain appropriate QoS values. Error distribution can be obtained. As a result, the radio situation prediction device 100 can avoid outputting abnormal data as a prediction result.

(1.2.2: Learning process)
Next, learning processing will be explained. The learning process is executed divided into (1) first step process, (2) second step process, and (3) third step process.
(1) In the first step, learning vector and label acquisition processing is performed from time-series data (observed value series) of QoS values.
(2) In the second step, a threshold for class determination is calculated based on the Rarity distribution.
(3) In the third step, data for calculating the prediction error distribution of QoS values is acquired.

Hereinafter, the process will be explained separately into (1) first step processing, (2) second step processing, and (3) third step processing.

(1.2.2.1: First step processing)
In the process of the first step, the radio situation prediction device 100 generates learning data D_train (QoS value time series data (observed value series) obtained by excluding abnormal patterns from the QoS value time series data (observed value series)). The learning vector and label acquisition process is executed from the training data (used for learning).
The control unit CPU1 sets the mode signal Mode to a signal value indicating “learning mode” and outputs it to the learning unit 2.

The learning unit 2 performs a learning process on the learning data D_train to perform a pattern matching prediction process using a PNN (Probabilistic Neural Network) or the like, and obtains learned data. Specifically, the learning unit 2 processes as shown in (1) to (3) below.
(1) Obtain k1 learning data D_train (time series data of QoS values) from time i (corresponding to time t0 in FIG. 10), and obtain a learning vector. Furthermore, the future value data (correct data) corresponding to the learning vector is set as the observed value (QoS value) at time i+k2 in the label associated with the learning vector. For example, in the case of FIG. 10, i=0, k1=3, and the time series data D1s of the QoS value is
D1s={D _i , D _i+1 , D _i+2 }
={a _i , a _i+1 , a _i+2 }={a ₀ , a ₁ , a ₂ }
The above data D1s is used as a learning vector. In the case of FIG. 10, since k2=4, the label associated with the learning vector is set to the observed value (QoS value) a i+ ₄ at time i+k2 (=i+4). Note that the target data for the above processing in the case of FIG. 10 is data included in the section of the window Win1 (the section from time t0 to tw) in FIG.
(2) Next, advance the time (time step) by one. That is, the section of the window Win1 in FIG. 10 is shifted by one time step (time corresponding to one data sampling) in the time axis direction Dir1, and the shifted section of the window Win1 is set as the data to be processed.

Similarly to process (1), k1 pieces of learning data D_train (time-series observed value series excluding abnormal patterns (time-series data of QoS values)) are acquired from time i+1, and learning vectors are acquired. Further, future value data (correct data) corresponding to the learning vector is set as an observed value (QoS value) at time i+k2+1 in a label associated with the learning vector. For example, in the case of FIG. 10, i=0, k1=3, and the time series data D1s of the QoS value is
D1s={D _i+1 , D _i+2 , D _i+3 }
= {a _i+1 , a _i+2 , a _i+3 }={a ₁ , a ₂ , a ₃ }
The above data D1s is used as a learning vector. In the case of FIG. 10, since k2=4, the label associated with the learning vector is set to the observed value (QoS value) a i+ ₄ at time i+k2 (=i+4).
(3) The learning unit 2 repeatedly executes the processes of (1) and (2) above.

At time i+n1-1 (n1: natural number), similarly to process (1), the learning data D_train (time-series observed value series excluding abnormal patterns (time-series data of QoS values)) is transferred to time i+n1-1. k1 pieces are obtained from , and a learning vector is obtained. Further, future value data (correct data) corresponding to the learning vector is set as an observed value (QoS value) at time i+n1-1+k2 in a label associated with the learning vector. For example, in the case of FIG. 10, i=0, k1=3, and the time series data D1s of the QoS value is
D1s={D _i+n1-1 , D _i+n1 , D _i+n1+1 }
= {a _i+n1-1 , a _i+n1 , a _i+n1+1 }={a _n1-1 , a _n1 , a _n1+1 }
The above data D1s is used as a learning vector. In the case of FIG. 10, since k2=4, the label associated with the learning vector is set to the observed value (QoS value) a _n1+3 at time i+n1-1+k2 (=i+n1+3).

The learning unit 2 obtains a future value prediction data set Dset1, a Rarity calculation data set Dset2, and a label (correct data) data set Dset_Label from the data obtained above.

For convenience of explanation, the following description will be made assuming that k1=3, k2=2, and the number of label data corresponding to one learning vector is one (corresponding to the case shown in FIG. 10). Further, k1, k2, and the number of label data corresponding to one learning vector are not limited to the above numbers.

なお、上記のＲａｒｉｔｙ計算用データセットＤｓｅｔ２の行列において、各行ベクトルが、時刻ｉの学習ベクトルに対応する。 In the above case (corresponding to the case shown in FIG. 10), the future value prediction data set Dset1, the Rarity calculation data set Dset2, and the label (correct data) data set Dset_Label are expressed as follows using a matrix. I can express it.

Note that in the matrix of the above-mentioned Rarity calculation data set Dset2, each row vector corresponds to a learning vector at time i.

The learning unit 2 uses the above data set to perform, for example, a learning process using PNN, and obtains a learned model (for example, a learning model using PNN). Note that as the processing using the PNN, for example, techniques disclosed in Japanese Patent Application No. 2017-232442 and Japanese Patent Application No. 2018-042747 may be used.

The learning unit 2 then uses the data acquired as described above ((1) future value prediction data set Dset1, (2) Rarity calculation data set Dset2, and (3) label (correct data) data set. Dset_Label) is stored in the trained data storage unit St1 as trained data D_trained.

(1.2.2.2: Second step processing)
In the process of the second step, the radio situation prediction device 100 acquires a Rarity distribution from the learning data D_train (time series data (observed value series) of QoS values excluding abnormal patterns), and based on the Rarity distribution, , executes processing to calculate a threshold value for class determination.
The control unit CPU1 sets the mode signal Mode to a signal value indicating “learning mode” and outputs it to the prediction determination processing unit 3. Further, the control unit CPU1 sets the control signal ctl1 to a signal value for executing the “second step process” (a signal value at which terminal 0 of the selector SEL1 is selected), and outputs it to the prediction determination processing unit 3. .

The data holding unit 31 stores and holds the data Din (=D_train) from the selector SEL0 for a predetermined period. The data holding unit 31 outputs the time series data of the data Din (=D_train) to the rarity acquisition unit 32 and the predicted value acquisition unit 33 as data D1s. If data Din at time i is expressed as D _i , then data D1s is set as, for example, three consecutive time series data D1s={D _i , D _i+1 , D _i+2 }. For convenience of explanation, as shown in FIG. 11, the observed value series of QoS values after time t1 (time series data included in the learning data D_train) are expressed as b ₀ , b ₁ , b ₂ , . . . shall be. In this case, the data holding unit 31 outputs data D1s={b ₀ , b ₁ , b ₂ } to the rarity acquisition unit 32 .

The norm acquisition unit 321 of the rarity acquisition unit 32 reads out the rarity calculation data set Dset2 of the trained data D_trained from the trained data storage unit St1, and inputs the data D1s output from the data holding unit 31. Note that here, a case will be described in which the norm acquisition unit 321 processes data D1s={b ₀ , b ₁ , b ₂ }.

The norm acquisition unit 321 calculates a norm (for example, Euclidean distance) using each row vector of the Rarity calculation data set Dset2 and data D1s (={b ₀ , b ₁ , b ₂ }). In other words, the norm acquisition unit 321 assumes that the m-th row vector of the Rarity calculation data set Dset2 is _{Dset2_rm} .
d ₀ = Norm (Dset2_r ₀ , D1s)
d ₁ = Norm(Dset2_r ₁ , D1s)
d ₂ = Norm(Dset2_r ₂ , D1s)
...
d _n1-1 = Norm(Dset2_r _n1-1 , D1s)
Norm (v1, v2): A function that obtains the norm (for example, Euclidean distance) of vector v1 and vector v2.
By the process corresponding to , norm set data {d ₀ , d ₁ , . . . , d _n1-1 } is obtained. Then, the norm acquisition unit 321 outputs data including the acquired norm set data {d ₀ , d ₁ , . . . , d _n1-1 } to the Rarity calculation unit 322 as data D_norm.

The Rarity calculation unit 322 calculates the rarity (Rarity) indicating the degree of rarity of the appearance of the time series data D1s from the data D_norm output from the norm acquisition unit 321. Specifically, the Rarity calculation unit 322
r _i =min(d _i , d _i+1 ,..., d _i+n1-1 )
min( ): A rarity value r _i corresponding to the time series data of QoS values {D _i , D _i+1 , D _i+2 } is obtained by executing a process equivalent to a function that obtains the minimum value of an element. Note that here, since the time series data of QoS values {D _i , D _i+1 , D _i+2 }={b ₀ , b ₁ , b ₂ }, the Rarity calculation unit 322
r ₀ = min (d ₀ , d ₁ ,..., d _n1-1 )
get.

Then, the Rarity calculation unit 322 outputs data including the calculated rarity r _i to the selector SEL1 of the prediction error distribution acquisition unit 35 as data D2.

The selector SEL1 selects the terminal 0 according to the control signal ctl1 set to the signal value for executing the "second step process" (the signal value that selects the terminal 0 of the selector SEL1), and selects the input signal. The data D2 is output to the rarity distribution acquisition section 351.

The Rarity distribution acquisition unit 351 acquires the Rarity distribution from the data D2 output from the selector SEL1. Specifically, as shown in FIG. 11, the Rarity distribution acquisition unit 351 acquires the rarity value _ri acquired at each time (each sampling time) for a predetermined period, and creates a frequency distribution diagram of the rarity value ri ( histogram) (corresponding to the lower right diagram in FIG. 11). The Rarity distribution acquisition unit 351 outputs the data specifying the frequency distribution diagram (histogram) of the rarity value ri acquired as described above to the class classification processing unit 352 as Rarity distribution data Dist_Rarity.

The class classification processing unit 352 executes class classification processing based on the data Dist_Rarity output from the rarity distribution acquisition unit 351. Specifically, a clustering method such as the K-means method or a method using SOM (Self-organizing map) is applied to classify Rarity into a predetermined number of classes, and a threshold Th (={th ₀ , th ₁ , ..., th _n2-1 }).

Here, a case will be described in which the K-means method is adopted.

When classifying each rarity value r _i into a preset number of classes (number of clusters) kcl (kcl: a natural number of 2 or more) based on the rarity distribution data Dist_Rarity, the class classification processing unit 352 uses the K-means method. , calculate the norm (e.g., Euclidean distance) between each rarity value r _i and the cluster center, and assign each rarity value r _i to the cluster with the nearest cluster center. The class classification processing unit 352 repeatedly executes this process, and when the cluster allocation situation does not change or the amount of change becomes less than a predetermined value, it determines that convergence has occurred and ends the class classification process. Then, when the class classification processing is completed, the class classification processing unit 352 sets a threshold Th (={th ₀ , th ₁ , ..., th _n2-1 }) for separating each class (each cluster). Set. It is assumed that the number of classes (number of clusters) kcl is set in advance in the class classification processing section 352.

The class classification processing unit 352 stores the threshold set data Th (={th ₀ , th ₁ , ..., th _n2-1 }) for classifying the classes acquired as described above in the first storage unit St2. Remember.

In the following, for convenience of explanation, as shown in the lower right diagram of FIG. 11, the number of classes is assumed to be "3", each class is defined as class Class_0, class Class_1, and class Class_2, and threshold value Th={th ₀ , th ₁ } will be explained.

(1.2.2.3: Third step processing)
In the third step, the radio condition prediction device 100 acquires data for calculating a prediction error distribution of QoS values.
The control unit CPU1 sets the mode signal Mode to a signal value indicating “learning mode” and outputs it to the prediction determination processing unit 3. Further, the control unit CPU1 sets the control signal ctl1 to a signal value for executing the “third step process” (a signal value at which terminal 1 of the selector SEL1 is selected), and outputs it to the prediction determination processing unit 3. .

The data holding unit 31 stores and holds the data Din (=D_train) from the selector SEL0 for a predetermined period. The data holding unit 31 outputs the time series data of the data Din (=D_train) to the rarity acquisition unit 32 and the predicted value acquisition unit 33 as data D1s. If data Din at time i is expressed as D _i , then data D1s is set as, for example, three consecutive time series data D1s={D _i , D _i+1 , D _i+2 }. For convenience of explanation, as shown in FIG. 12, the observed value series of QoS values after time t2 (time series data included in the learning data D_train) are expressed as c ₀ , c ₁ , c ₂ , . . . shall be. In this case, the data holding unit 31 outputs data D1s={c ₀ , c ₁ , c ₂ } to the rarity acquisition unit 32 and predicted value acquisition unit 33 .

Further, the data holding unit 31 outputs the actual data (true value) of the time to be predicted by the predicted value acquisition unit 33 to the subtracter 34 as data D1_true (=D _{i + k} (c ₄ in the case of FIG. 12)). do.

The norm acquisition unit 321 of the rarity acquisition unit 32 reads out the rarity calculation data set Dset2 of the trained data D_trained from the trained data storage unit St1, and inputs the data D1s output from the data holding unit 31. Note that here, a case will be described in which the norm acquisition unit 321 processes data D1s={c ₀ , c ₁ , c ₂ }.

The norm acquisition unit 321 calculates a norm (for example, Euclidean distance) using each row vector of the Rarity calculation data set Dset2 and data D1s (={c ₀ , c ₁ , c ₂ }). In other words, the norm acquisition unit 321 assumes that the m-th row vector of the Rarity calculation data set Dset2 is _{Dset2_rm} .
d ₀ = Norm (Dset2_r ₀ , D1s)
d ₁ = Norm(Dset2_r ₁ , D1s)
d ₂ = Norm(Dset2_r ₂ , D1s)
...
d _n1-1 = Norm(Dset2_r _n1-1 , D1s)
Norm (v1, v2): A function that obtains the norm (for example, Euclidean distance) of vector v1 and vector v2.
By the process corresponding to , norm set data {d ₀ , d ₁ , . . . , d _n1-1 } is obtained. Then, the norm acquisition unit 321 outputs data including the acquired norm set data {d ₀ , d ₁ , . . . , d _n1-1 } to the Rarity calculation unit 322 as data D_norm.

The Rarity calculation unit 322 calculates the rarity (Rarity) indicating the degree of rarity of the appearance of the time series data D1s from the data D_norm output from the norm acquisition unit 321. Specifically, the Rarity calculation unit 322
r _i =min(d _i , d _i+1 ,..., d _i+n1-1 )
min( ): A rarity value r _i corresponding to the time series data of QoS values {D _i , D _i+1 , D _i+2 } is obtained by executing a process equivalent to a function that obtains the minimum value of an element. Note that here, since the time series data of QoS values {D _i , D _i+1 , D _i+2 }={c ₀ , c ₁ , c ₂ }, the Rarity calculation unit 322
r ₀ = min (d ₀ , d ₁ ,..., d _n1-1 )
get.

Then, the Rarity calculation unit 322 outputs data including the calculated rarity r _i to the selector SEL1 of the prediction error distribution acquisition unit 35 as data D2.

The selector SEL1 selects the terminal 1 according to the control signal ctl1 set to the signal value for executing the "third step process" (the signal value that selects the terminal 1 of the selector SEL1), and selects the input signal. The data D2 is output to the Rarity class classification section 353.

The Rarity class classification unit 353 reads the threshold set data Th (={th ₀ , th ₁ }) from the first storage unit St2, and based on the threshold set data Th (={th ₀ , th ₁ }), The data D2 is classified into classes, and data including the number j of the class to which the data D2 has been assigned is output to the prediction error distribution data acquisition unit 354 as data D3.

The predicted value acquisition unit 33 reads trained data D_trained (Rarity calculation data set Dset2 and label (correct data) data set Dset_Label) from the trained data storage unit St1. Similar to the process executed by the norm acquisition unit 321, the predicted value acquisition unit 33 acquires the QoS value time series data D1s (={c ₀ , c ₁ , c ₂ }) and each row vector of the Rarity calculation data set Dset2. The norm (for example, Euclidean distance) is calculated, the row vector having the minimum value is detected, and the label (correct data) corresponding to the detected row vector is specified using the data set Dset_Label. Thereby, the predicted value acquisition unit 33 acquires the label corresponding to the row vector having the minimum value of the norm as the predicted data value D_predict at a future time. For example, in the case shown in FIG. 12, the time series data D1s (={c ₀ , c ₁ , c ₂ }) of QoS values and the vector ({a ₂ , a ₃ , a ₄ }) takes the minimum value, the predicted value acquisition unit 33 corresponds to the vector ({a ₂ , a ₃ , a ₄ }) in the third row of the Rarity calculation data set Dset2. The label value a ₆ (data in the third row of the data set Dset_Label of labels (correct data)) is specified as the predicted data value D_predict at a future time (if current time i is future time i+4).

Then, the predicted value acquisition unit 33 outputs the predicted data value D_predict (=D' _i+k =a ₆ ) of the future time (if the current time i is the future time i+4) acquired as described above to the subtracter 34.

The subtracter 34 inputs the data D1_true (=D _i+k =c ₄ ) output from the data holding section 31 and the data D_predict (=D' _i+k =a ₆ ) output from the predicted value acquisition section 33. The subtractor 34 is
D_prediction_error=D1_true-D_predict
A process corresponding to (subtraction process) is executed to obtain prediction error data D_prediction_error (=Err _i+k ). Then, the subtracter 34 outputs the acquired prediction error data D_prediction_error (=Err _i+k ) to the prediction error distribution data acquisition section 354 of the prediction error distribution acquisition section 35 .

The prediction error distribution data acquisition unit 354 uses the data D_prediction_error output from the subtracter 34 and the data D3 (=j=class( _ri ), class(x) is the Rarity value x function to get the number of the class to which the class belongs). The prediction error distribution data acquisition unit 354 distributes the data D_prediction_error into classes based on the data D3, and acquires the occurrence frequency distribution of the data D_prediction_error for each classified class (for a group of data assigned to the same class). Then, prediction error distribution data (Err_dist _j ) is obtained. Then, the prediction error distribution data acquisition unit 354 stores data D4 including the acquired prediction error distribution data Err_dist _j in the second storage unit St3.

In general, prediction errors classified into class 0 (Class_0) correspond to normal input patterns (time series data of frequently occurring QoS values), so prediction errors classified into class 0 (Class_0) have small values. The variance (or standard deviation) of the distribution of the error data group is small.

On the other hand, the prediction error classified into class 1 (Class_1) has a lower appearance frequency than the input pattern of class 0, so its value is larger than the prediction error classified into class 0, and the value is larger than the prediction error classified into class 0. ) is larger than the variance (or standard deviation) of the distribution of the prediction error data group classified into class 0 (Class_0).

In addition, the prediction error classified into class 2 (Class_2) has a lower appearance frequency than the input pattern of class 1, so its value is larger than the prediction error classified into class 1, and the value of the prediction error classified into class 2 (Class_2) is lower than that of the input pattern of class 1. The variance (or standard deviation) of the distribution of the classified prediction error data group is larger than the variance (or standard deviation) of the distribution of the prediction error data group classified into class 1 (Class_1).

Because of the above characteristics, the prediction error distribution data Err_dist ₀ of the prediction errors classified into class 0 (Class_0), the prediction error distribution data Err_dist ₁ of the prediction errors classified into class 1 (Class_1), and the class The prediction error distribution data Err_dist ₂ of the prediction errors classified as 2 (Class_2) has a distribution as shown in the middle of the right end of FIG. 12, for example.

(1.2.3: Prediction processing)
Next, the prediction process (communication failure prediction execution process) will be explained.

FIG. 13 is a diagram for explaining prediction processing (communication failure prediction execution processing).

In the prediction process (communication failure prediction execution process), the radio situation prediction device 100 calculates the predicted value D' i _{+ k} of the QoS value at a future time and the prediction error distribution data D4 (= Err_dist _j ), a risk value indicating the possibility of a communication failure occurring at the time (future time) when the true value of the data D_predict is obtained is obtained.

The demodulation processing unit Dm1 of the radio situation prediction device 100 performs demodulation processing on the radio signal received via the antenna Ant1, and obtains an RF demodulated signal and a BB demodulated signal. The observation data acquisition unit FE1 acquires a QoS value from the RF demodulation signal and/or the BB demodulation signal acquired by the demodulation processing unit Dm1. Then, the observation data acquisition unit FE1 outputs data including the acquired QoS value to the selector SEL0 as data Din0. The selector SEL0 selects the data Din0 output from the observation data acquisition section FE1, and outputs it to the data holding section 31 of the prediction determination processing section 3 as data Din (=Din0).

The control unit CPU1 sets the mode signal Mode to a signal value indicating the “prediction processing execution mode” and outputs it to the prediction determination processing unit 3. Further, the control unit CPU1 sets the control signal ctl1 to a signal value for executing the “prediction process” (a signal value at which terminal 2 of the selector SEL1 is selected), and outputs it to the prediction determination processing unit 3.

The data holding unit 31 stores and holds data Din (=Din0) from the selector SEL0 for a predetermined period. The data holding unit 31 outputs the time series data of the data Din to the rarity acquisition unit 32 and the predicted value acquisition unit 33 as data D1s. If data Din at time i is expressed as D _i , then data D1s is set as, for example, three consecutive time series data D1s={D _i , D _i+1 , D _i+2 }. For convenience of explanation, the observed value series (time series data) of QoS values after time t3 are assumed to be q ₀ , q ₁ , q ₂ , . . . as shown in FIG. 13 . In this case, the data holding unit 31 outputs data D1s={q ₀ , q ₁ , q ₂ } to the rarity acquisition unit 32 and predicted value acquisition unit 33 .

The norm acquisition unit 321 of the rarity acquisition unit 32 reads out the rarity calculation data set Dset2 of the trained data D_trained from the trained data storage unit St1, and inputs the data D1s output from the data holding unit 31. Note that here, a case will be described in which the norm acquisition unit 321 processes data D1s={q ₀ , q ₁ , q ₂ }.

The norm acquisition unit 321 calculates a norm (for example, Euclidean distance) using each row vector of the Rarity calculation data set Dset2 and the data D1s (={q ₀ , q ₁ , q ₂ }). In other words, the norm acquisition unit 321 assumes that the m-th row vector of the Rarity calculation data set Dset2 is _{Dset2_rm} .
d ₀ = Norm (Dset2_r ₀ , D1s)
d ₁ = Norm(Dset2_r ₁ , D1s)
d ₂ = Norm(Dset2_r ₂ , D1s)
...
d _n1-1 = Norm(Dset2_r _n1-1 , D1s)
Norm (v1, v2): A function that obtains the norm (for example, Euclidean distance) of vector v1 and vector v2.
By the process corresponding to , norm set data {d ₀ , d ₁ , . . . , d _n1-1 } is obtained. Then, the norm acquisition unit 321 outputs data including the acquired norm set data {d ₀ , d ₁ , . . . , d _n1-1 } to the Rarity calculation unit 322 as data D_norm.

The Rarity calculation unit 322 calculates the rarity (Rarity) indicating the degree of rarity of the appearance of the time series data D1s from the data D_norm output from the norm acquisition unit 321. Specifically, the Rarity calculation unit 322
r _i =min(d _i , d _i+1 ,..., d _i+n1-1 )
min( ): A rarity value r _i corresponding to the time series data of QoS values {D _i , D _i+1 , D _i+2 } is obtained by executing a process equivalent to a function that obtains the minimum value of an element. Note that here, since the time series data of QoS values {D _i , D _i+1 , D _i+2 }={c ₀ , c ₁ , c ₂ }, the Rarity calculation unit 322
r ₀ = min (d ₀ , d ₁ ,..., d _n1-1 )
get.

Then, the Rarity calculation unit 322 outputs data including the calculated rarity r _i to the selector SEL1 of the prediction error distribution acquisition unit 35 as data D2.

The selector SEL1 selects the terminal 2 according to the control signal ctl1 set to the signal value for executing the "prediction process" (the signal value at which the terminal 2 of the selector SEL1 is selected), and selects the input data D2 ( =r _i ) is output to the class determination unit 355.

The class determination unit 355 receives data D2 (=r _i ) output from the selector SEL1 and threshold set data Th output from the first storage unit St2. The class determination unit 355 classifies the data D2 (=r _i ) based on the threshold set data Th, and classifies the class to which the data D2 (=r _i ) is assigned (for example, class number j=class(r _i ) ). Then, the class determination unit 355 outputs the specified class number j (=class( _ri )) to the risk value acquisition unit 36.

The predicted value acquisition unit 33 reads trained data D_trained (Rarity calculation data set Dset2 and label (correct data) data set Dset_Label) from the trained data storage unit St1. Similar to the process executed by the norm acquisition unit 321, the predicted value acquisition unit 33 acquires the QoS value time series data D1s (={q ₀ , q ₁ , q ₂ }) and each row vector of the Rarity calculation data set Dset2. The norm (for example, Euclidean distance) is calculated, the row vector having the minimum value is detected, and the label (correct data) corresponding to the detected row vector is specified using the data set Dset_Label. Thereby, the predicted value acquisition unit 33 acquires the label corresponding to the row vector having the minimum value of the norm as the predicted data value D_predict at a future time. For example, in the case shown in FIG. 13, the QoS value time series data D1s (={q ₀ , q ₁ , q ₂ }) and the vector ({a ₂ , a ₃ , a ₄ }) takes the minimum value, the predicted value acquisition unit 33 corresponds to the vector ({a ₂ , a ₃ , a ₄ }) in the third row of the Rarity calculation data set Dset2. The value a ₆ of the label (data in the third row of the data set Dset_Label of labels (correct data)) is acquired as the predicted data value D_predict at a future time (if current time i is future time i+4).

Then, the predicted value acquisition unit 33 outputs the predicted data value D_predict (=D' _i+k =a ₆ ) at the future time (if the current time i is future time i+4) acquired as described above to the risk value acquisition unit 36. do.

The risk value acquisition unit 36 receives the data D_predict (=D' _i+k =a ₆ ) output from the predicted value acquisition unit 33, the class number j output from the class determination unit 355, and the data output from the second storage unit St3. data D4 (=Err_dist _j ), data QoS_req indicating a required reference value of communication quality data, and a mode signal Mode output from the control unit CPU1 are input. When the mode signal Mode is a signal indicating "prediction processing execution mode", the risk value acquisition unit 36 calculates the data D_predict (=D' _i+k = a ₆ ) based on the prediction error distribution data D4 (=Err_dist _j ). At the time (future time) when the true value (q ₄ ) is obtained, a risk value indicating the possibility that a communication failure will occur is obtained. Specifically, the risk value acquisition unit 36 reads prediction error distribution data Err_dist _j of class j determined when acquiring the predicted value D_predict (=D′ _i+k ) from the second storage unit St3, and obtains the predicted value. A prediction error distribution centered at D_predict (=D' _i+k ) is obtained. Then, in the prediction error distribution centered on the predicted value D_predict (=D' _i+k ), the prediction error value in the area below the required reference value QoS_req of communication quality data is integrated (or integrated), and the integrated value (integrated value) is calculated. Obtain as the risk value. For example, the area of the gray area (area below the required reference value QoS_req of communication quality data) in the diagram shown near the center of FIG. 13 becomes the risk value.

Then, the risk value acquisition unit 36 outputs the risk value acquired as described above as data Dout. By considering the risk value acquired by the risk value acquisition unit 36, it is possible to grasp the possibility of occurrence of a communication failure at a future time, and as a result, highly accurate prediction processing of the occurrence of a communication failure can be realized. .

As described above, the radio situation prediction device 100 executes the prediction process.

Here, using FIG. 14, when an input pattern (an input pattern with a low Rarity value) in which the occurrence pattern of time-series data of observed data (QoS value) frequently appears is input to the radio situation prediction device 100 (FIG. a)) and when an input pattern (input pattern with a high Rarity value) in which the occurrence pattern of time-series data of observed data (QoS value) is rare is input to the radio situation prediction device 100 (FIG. 14(b)). The acquisition of risk values will be explained.

In the case of Fig. 14(a), the occurrence pattern of time-series data of observed data (QoS value) is a frequently occurring input pattern (input pattern with a low Rarity value), and the Rarity value is low and the frequency of occurrence of the input pattern is high. , the dispersion (standard deviation, variance) of the prediction error distribution data Err_dist _j of the class determined by the Rarity value calculated from the input pattern is small. In this case, for example, the class with class number j=0 (Class_0) is selected, and the prediction error distribution data Err_dist ₀ is selected. Then, the risk value acquisition unit 36 acquires a prediction error distribution in which the center of the prediction error distribution data Err_dist ₀ is the predicted value D′ _i+k (D′ _i+k : predicted value of the QoS value at time t _i+k ). In the prediction error distribution, an integral value (integrated value) of a region where the QoS value is equal to or less than the value QoS_req (region AR1 in the case of FIG. 14(a)) is acquired as a risk value. In the case of FIG. 14(a), since the area of the region AR1 is quite small, the risk value is also a small value. That is, in the case of FIG. 14A, it is appropriately determined that the possibility of a communication failure occurring at time t _i+k is quite small.

On the other hand, in the case of Fig. 14(b), the occurrence pattern of the time series data of the observed data (QoS value) is a rare input pattern (input pattern with a high Rarity value), and the Rarity value is high and the frequency of occurrence of the input pattern is Since it is low, the dispersion (standard deviation, variance) of the prediction error distribution data Err_dist _j of the class determined by the Rarity value calculated from the input pattern is large. In this case, for example, the class with class number j=2 (Class_2) is selected, and the prediction error distribution data Err_dist ₂ is selected. Then, the risk value acquisition unit 36 acquires a prediction error distribution in which the center of the prediction error distribution data Err_dist ₂ is the predicted value D′ _i+k (D′ _i+k : predicted value of the QoS value at time t _i+k ). In the prediction error distribution, an integral value (integrated value) of a region where the QoS value is equal to or less than the value QoS_req (region AR2 in the case of FIG. 14(b)) is acquired as a risk value. In the case of FIG. 14(b), since the area of the region AR1 is large, the risk value also becomes a large value. That is, in the case of FIG. 14A, it is appropriately determined that there is a high possibility that a communication failure will occur at time t _i+k .

By processing in this way, the radio situation prediction device 100 can appropriately prevent the occurrence of communication failure even when an input pattern with a rare occurrence pattern of time-series data of observed data (QoS value) is input. It is possible to grasp the possibility of the presence or absence of a communication failure and perform highly accurate prediction processing of the occurrence of a communication failure.

As described above, in the radio situation prediction device 100, by learning the occurrence pattern of time series data of observation data (QoS value) during learning, the occurrence pattern of time series data of frequently occurring observation data (QoS value) Understand. In addition, the radio situation prediction device 100 calculates the similarity between the occurrence pattern of time-series data of observed data (QoS value) acquired during learning and the time-series data of actual observed data (QoS value), for example, by norm ( For example, the determination is made by calculating the Euclidean distance), and a rarity value indicating the degree of rarity of the appearance of time series data of observed data (QoS value) is calculated. Then, the radio situation prediction device 100 obtains a rarity distribution, which is a distribution of rarity values, and classifies the data group of rarity values into a predetermined class using a threshold value based on the rarity distribution.

Then, at the time of prediction, the radio situation prediction device 100 identifies the class to be classified based on the rarity value r _i of the time series data of the observation data (QoS value) to be processed, and determines the prediction error corresponding to the class. Based on the distribution data Err_dist _j and the predicted value of the future time, a risk value indicating the probability of occurrence of a communication failure at the future time is acquired. Then, the radio situation prediction device 100 can appropriately grasp the possibility that a communication failure will occur at a future time based on the acquired risk value. By processing as described above, the radio situation prediction device 100 can obtain the risk value as a continuous value or multivalued data, so it is not possible to make a determination based on binary data indicating the presence or absence of a communication failure as in the conventional technology. It is possible to achieve much more accurate prediction processing than conventional processing.

In this way, the radio situation prediction device 100 takes into account the rarity (rarity) of the input pattern (time-series data of observed data (QoS value)), which is an important factor when detecting a communication failure. Since a risk value indicating the possibility of occurrence of a failure can be obtained, highly accurate prediction processing for the occurrence of a communication failure can be realized.

In addition, the radio situation prediction device 100 executes prediction processing for the occurrence of communication failures in consideration of the rarity (rarity) of input patterns (time series data of observed data (QoS values)). Alternatively, even if an unlearned input pattern occurs, it is possible to appropriately obtain a risk value that indicates the possibility of a communication failure occurring, and as a result, highly accurate communication failure prediction processing can be achieved. .

Furthermore, the radio situation prediction device 100 detects an abnormal pattern in observation data (QoS value) based on the degree of abnormality, and generates learning data excluding the detected abnormal pattern. Then, the radio situation prediction device 100 executes a learning process using the learning data, so that an appropriate learning process is executed and an appropriate error distribution of QoS values can be obtained. As a result, the radio situation prediction device 100 can appropriately avoid outputting abnormal data as a prediction result. If a learning process is executed without removing an abnormal pattern based on the degree of abnormality, and a prediction process is executed after the learning process, even though the probability of an abnormal pattern occurring is low, QoS values similar to the abnormal pattern will be It is common to make incorrect predictions when time series patterns occur. The radio situation prediction device 100 executes a learning process using learning data from which abnormal patterns of observed data (QoS values) are excluded based on the degree of abnormality, and performs a prediction process after the learning process. Note that even though the probability of an abnormal pattern occurring is low, it is not erroneously predicted that a time-series pattern of QoS values similar to the abnormal pattern will occur. That is, the radio situation prediction device 100 can realize highly accurate communication failure prediction processing.

In addition, although the case where the learning data generation processing unit 1 determines the time series pattern of QoS values to be excluded by comparing the degree of abnormality s _j and the threshold Th0 has been described above, the present invention is not limited to this. There isn't. For example, the learning data generation processing unit 1 calculates the degree of abnormality s _j in a predetermined period (for example, each period Prd1, Prd2, Prd3), and performs sorting processing on the degree of abnormality s _j in the period. The pattern (time series pattern of QoS values) corresponding to the abnormality degree s _j included in the top x% (x: real number (for example, x is a real number of 10 or less)) from the largest value is an abnormal pattern. The time series data from which the abnormal pattern has been determined and the abnormal pattern has been excluded may be used as the learning data D_train.

[Other embodiments]
In the above embodiment, a case has been described in which the observed value is a QoS value, but the observed value is not limited to this. Data may also be used.

Further, the learned data storage section St1, the first storage section St2, the second storage section St3, the first learned data storage section St1A, and the second learned data storage section St1B are It may be provided externally.

Furthermore, in the above embodiment, a case has been described in which prediction error distribution data is acquired by acquiring the occurrence frequency distribution of data D_prediction_error, and a risk value is acquired using the acquired prediction error distribution data. For example, the processing may be performed as follows without limitation. In the radio situation prediction device 100, for each classified class, from the occurrence frequency distribution of data D_prediction_error, the average value μ _j of data D_prediction_error (μ _j : average value of prediction error values classified into the class with class number j), Calculate the standard deviation σ _j (σ _j : standard deviation of prediction error values classified into the class with class number j), and determine that the prediction error distribution data Err_dist _j is a distribution that follows the normal distribution N (μ _j , σ _j ² ) It is also possible to perform processing to obtain the risk value by assuming (approximating) the risk value.

Further, in the above embodiment, a case has been described in which a communication failure in a wireless environment is predicted, but the present invention is not limited to this. QoS information may be acquired from a hub in the backbone of a mixed network to predict communication failures in the communication environment.

Furthermore, in the radio situation prediction device 100 described in the above embodiment, each block may be individually formed into one chip using a semiconductor device such as an LSI, or may be formed into one chip so as to include a part or all of the blocks. good.

Although it is referred to as an LSI here, it may also be called an IC, system LSI, super LSI, or ultra LSI depending on the degree of integration.

Moreover, the method of circuit integration is not limited to LSI, and may be implemented using a dedicated circuit or a general-purpose processor. An FPGA (Field Programmable Gate Array) that can be programmed after the LSI is manufactured or a reconfigurable processor that can reconfigure the connections and settings of circuit cells inside the LSI may be used.

Furthermore, part or all of the processing of each functional block in each of the above embodiments may be realized by a program. Part or all of the processing of each functional block in each of the above embodiments is performed by a central processing unit (CPU) in a computer. Further, programs for performing each process are stored in a storage device such as a hard disk or ROM, and are read out to the ROM or RAM and executed.

Further, each process of the above embodiments may be realized by hardware, or by software (including cases where it is realized together with an OS (operating system), middleware, or a predetermined library). Furthermore, it may be realized by mixed processing of software and hardware.

For example, when each functional unit of the above embodiment is realized by software, the hardware configuration shown in FIG. 15 (for example, a hardware configuration in which a CPU, ROM, RAM, input unit, output unit, etc. are connected by a bus) Each functional unit may be realized by software processing using .

Further, the execution order of the processing method in the above embodiment is not necessarily limited to the description of the above embodiment, and the execution order can be changed without departing from the gist of the invention.

A computer program that causes a computer to execute the method described above and a computer-readable recording medium on which the program is recorded are included within the scope of the present invention. Here, examples of computer-readable recording media include flexible disks, hard disks, CD-ROMs, MOs, DVDs, DVD-ROMs, DVD-RAMs, large-capacity DVDs, next-generation DVDs, and semiconductor memories. .

The computer program is not limited to one recorded on the recording medium, but may be transmitted via a telecommunication line, a wireless or wired communication line, a network typified by the Internet, or the like.

Note that the specific configuration of the present invention is not limited to the above-described embodiments, and various changes and modifications can be made without departing from the gist of the invention.

100 Radio situation prediction device 1 Learning data generation processing section 2 Learning section 3 Prediction judgment processing section 32 Rarity acquisition section 33 Predicted value acquisition section 34 Subtractor (subtraction section)
35 Prediction error distribution acquisition unit 36 Risk value acquisition unit St1 Learned data storage unit

Claims (6)

The degree of abnormality of the time-series pattern of communication quality data is acquired, and according to the acquired degree of abnormality, excluded data, which is data to be excluded from the learning process, is determined. a learning data generation processing unit that acquires data from which excluded data has been excluded as learning processing target data;
A learning unit that performs a learning process using the learning processing target data acquired by the learning data generation processing unit, the learning unit comprising: a generation pattern of a time series pattern of the communication quality data; and a time series pattern after the time series pattern is generated; the learning unit that executes the learning process using the communication quality data at a future time that is after a predetermined time ;
a rarity calculation dataset including the occurrence pattern of the time-series pattern of the communication quality data acquired by the learning unit; and the communication quality data at a future time that is a predetermined time after the occurrence of the time-series pattern. a storage unit that stores trained data including a label data set including the label data set;
a predicted value acquisition unit that acquires a predicted value of the communication quality data after a predetermined time from a time-series pattern of the communication quality data input at the current time based on the learned data;
The predicted value of the communication quality data acquired by the predicted value acquisition unit and the true value of the predicted value , which is the actual measured value of the communication quality data after the predetermined time. a prediction error distribution acquisition unit that acquires a prediction error distribution that is a distribution of prediction error values that are differences;
a risk value acquisition unit that acquires a risk value indicating the possibility that a communication failure will occur at a future time that is a predetermined time after the current time, based on the prediction error distribution and the prediction value;
A communication status prediction device comprising: The learning data generation processing unit includes:
When an average vector acquisition period is set, a unit section including a predetermined number of the communication quality data is set, and the communication quality data included in the unit section is time series data a _i (i: natural number), a plurality of A vector v_a _i having the time series data a _i as an element and an average vector v_a _ave of the time series data included in the average vector acquisition period are obtained, and the vector v_a _i and the average vector v_a _ave are obtain the norm of the difference vector of as the abnormality degree,
The communication status prediction device according to claim 1. The learning data generation processing unit includes:
comparing the degree of abnormality with a predetermined threshold, and if the degree of abnormality is greater than the threshold, identifying a time-series pattern of communication quality data corresponding to the degree of abnormality as the excluded data;
The communication status prediction device according to claim 1 or 2. The learning data generation processing unit includes:
An average vector acquisition period is set, and the abnormality degrees acquired in the average vector acquisition period are sorted, and the abnormality degree value is included in the top x% (x: real number) from the highest abnormality degree. identifying a time-series pattern of communication quality data corresponding to the abnormality level as the excluded data;
The communication status prediction device according to claim 1 or 2. Communication status prediction performed using a communication status prediction device including a learning data generation processing unit, a learning unit, a storage unit, a predicted value acquisition unit, a prediction error distribution acquisition unit, and a risk value acquisition unit A method,
The learning data generation processing unit acquires the degree of abnormality of the time-series pattern of the communication quality data, identifies excluded data that is data to be excluded from the learning process according to the acquired degree of abnormality, and a learning data generation processing step of acquiring data from which the excluded data is excluded from the time-series pattern of the quality data as learning processing target data;
A learning step in which the learning unit performs a learning process using the learning processing target data acquired in the learning data generation processing step, the learning step comprising: an occurrence pattern of a time-series pattern of the communication quality data; the learning step of performing the learning process using the communication quality data at a future time that is after a predetermined time after the pattern has occurred;
The storage unit stores a rarity calculation data set including the occurrence pattern of the time series pattern of the communication quality data acquired in the learning step, and a future time that is a predetermined time after the occurrence of the time series pattern. a label data set including the communication quality data;
a predicted value acquisition step in which the predicted value acquisition unit acquires a predicted value of the communication quality data after a predetermined time from a time-series pattern of communication quality data input at the current time based on the learned data; and,
The prediction error distribution acquisition unit calculates the predicted value of the communication quality data acquired in the predicted value acquisition step and the true value of the predicted value, and the actual measurement of the communication quality data after the predetermined time. a prediction error distribution acquisition step of acquiring a prediction error distribution that is a distribution of prediction error values that are the difference from the true value that is the value ;
a risk value acquisition step in which the risk value acquisition unit acquires a risk value indicating the possibility that a communication failure will occur at a future time that is a predetermined time after the current time , based on the prediction error distribution and the predicted value; ,
A communication status prediction method comprising: A program for causing a computer to execute the communication status prediction method according to claim 5.

Images