WO2021181687A1 - Prediction model creation device, prediction model creation method, and program - Google Patents

Prediction model creation device, prediction model creation method, and program Download PDF

Info

Publication number
WO2021181687A1
WO2021181687A1 PCT/JP2020/011219 JP2020011219W WO2021181687A1 WO 2021181687 A1 WO2021181687 A1 WO 2021181687A1 JP 2020011219 W JP2020011219 W JP 2020011219W WO 2021181687 A1 WO2021181687 A1 WO 2021181687A1
Authority
WO
WIPO (PCT)
Prior art keywords
prediction model
positioning error
failure rate
distribution
observations
Prior art date
Application number
PCT/JP2020/011219
Other languages
French (fr)
Japanese (ja)
Inventor
山田　昌弘
剛志 是永
健司 ▲高▼尾
陽平 知識
Original Assignee
三菱重工機械システム株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 三菱重工機械システム株式会社 filed Critical 三菱重工機械システム株式会社
Priority to PCT/JP2020/011219 priority Critical patent/WO2021181687A1/en
Priority to JP2022505712A priority patent/JP7235931B2/en
Publication of WO2021181687A1 publication Critical patent/WO2021181687A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
    • G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
    • G01C21/26—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 specially adapted for navigation in a road network
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S19/00—Satellite radio beacon positioning systems; Determining position, velocity or attitude using signals transmitted by such systems
    • G01S19/01—Satellite radio beacon positioning systems transmitting time-stamped messages, e.g. GPS [Global Positioning System], GLONASS [Global Orbiting Navigation Satellite System] or GALILEO
    • G01S19/13—Receivers
    • G01S19/23—Testing, monitoring, correcting or calibrating of receiver elements

Definitions

  • the present invention relates to a predictive model creation device for predicting a positioning error distribution, a predictive model creation method, and a program.
  • Patent Document 1 describes a technique for calculating a positioning error based on an error between a pseudo distance from a plurality of satellites and a distance from a positioned position to the satellite.
  • a vehicle traveling on the road A shown in FIG. 8 positions the traveling position of its own vehicle by a satellite positioning system, it may be traveling on a nearby road B depending on the degree of error in the positioned position information. Such results may be obtained.
  • misrecognition of the road on which the vehicle travels becomes a problem.
  • a technique for calculating the probability of making a mistake on the road on which the vehicle is traveling hereinafter referred to as a failure rate
  • Both L1 and L2 in FIG. 8 show the distribution of positioning error.
  • the positioning error is normally distributed, and the standard deviation of the normal distribution is defined as the "positioning error".
  • the distribution L1 shows a probability distribution when the positioning error is smaller than the distribution L2. See FIG.
  • the vertical axis of the graph of FIG. 9 shows the failure rate, and the horizontal axis shows the magnitude of the positioning error.
  • the difference D1 between the positioning error R1 and the positioning error R2 and the difference D2 between the positioning error R2 and the positioning error R3 are equal.
  • the difference D3 between the failure rate when the positioning error R1 and the failure rate when the positioning error R2 is R2 is significantly different from the difference D4 between the failure rate when the positioning error R2 and the failure rate when the positioning error R3 is set. That is, the magnitude of the positioning error is not proportional to the failure rate. Therefore, when dealing with the failure rate, the difference D1 between the positioning error R1 and the positioning error R2 and the difference D2 between the positioning error R2 and the positioning error R3 should not be treated as equivalent.
  • a prediction model when creating a prediction model, learning is performed so as to minimize or maximize a predetermined evaluation function. For example, when a prediction model is created by using the least squares error as an evaluation function and reducing the error between the predicted value and the measured value of the positioning error, the positioning error is predicted to be R3 when the true magnitude of the positioning error is R2.
  • R1 when the true magnitude of the positioning error is R2, and both of them have the value of the evaluation function. It is possible that they will be equivalent.
  • this prediction model may become a model in which the failure rate of erroneously recognizing road A and road B is significantly different from the actual model. With such a prediction model, it is not possible to grasp the accurate failure rate.
  • the present disclosure provides a predictive model creation device, a predictive model creation method, and a program capable of solving the above-mentioned problems.
  • the prediction model creation device is a prediction model creation device that creates a prediction model that predicts a positioning error indicating an error in the position positioned by the satellite positioning system, and is a feature of the terrain at a certain point.
  • a unit a prediction model creation unit that creates a prediction model that outputs the probability distribution of the positioning error when the topographical feature data is input, a probability distribution of the positioning error output by the prediction model, a predetermined function, and the like.
  • the evaluation value of the failure rate is calculated based on the number of observations for each positioning error and the function. It includes an evaluation unit that calculates based on the failure rate, and an update unit that updates the prediction model so that the evaluation value becomes an optimum value.
  • the prediction model creation method is a prediction model creation method for creating a prediction model that predicts a positioning error indicating an error in the position positioned by the satellite positioning system, and is a feature of the terrain at a certain point.
  • the topographical feature data is input, there is a step of creating a prediction model that outputs the probability distribution of the positioning error, and in the step of creating the prediction model, the probability of the positioning error output by the prediction model.
  • the evaluation value of the failure rate is the number of observations for each positioning error and the above. It is calculated based on the observation failure rate calculated based on the function, and the prediction model is updated so that the evaluation value becomes the optimum value.
  • the program is a prediction model creation method for creating a prediction model for predicting a positioning error indicating an error in the position positioned by a satellite positioning system on a computer, and is a feature of the terrain at a certain point.
  • the topographical feature data is input, there is a step of creating a prediction model that outputs the probability distribution of the positioning error, and in the step of creating the prediction model, the probability of the positioning error output by the prediction model.
  • the evaluation value of the failure rate is the number of observations for each positioning error and the above. The process of updating the prediction model so that the evaluation value becomes the optimum value, which is calculated based on the observation failure rate calculated based on the function, is executed.
  • prediction model creation device prediction model creation method and program, it is possible to create a prediction model that estimates the distribution of the error of the positioned position in the positioning using satellites.
  • FIG. 1 is a diagram illustrating a positioning error and a failure rate.
  • FIG. 2 is a diagram illustrating a positioning error and a failure rate. It is a figure which shows an example of the hardware configuration of the prediction model creation apparatus in each embodiment of this invention.
  • FIG. 1 is a diagram showing an example of a prediction model creation device according to the first embodiment of the present invention.
  • the prediction model creation device 10 includes a data acquisition unit 11, a prediction model creation unit 12, and a storage unit 13.
  • the data acquisition unit 11 acquires the training data necessary for creating the prediction model M that predicts the probability distribution of the positioning error of the satellite positioning system at an arbitrary point.
  • the training data includes, for example, topographical feature data indicating the topographical features of each of a plurality of points, speed data when the vehicle travels at that point, and positioning error when the point is traveled multiple times. It contains information such as the number of observations and the total number of runs (total number of observations).
  • the prediction model creation unit 12 creates a prediction model M that predicts the probability distribution of the positioning error of the satellite positioning system so that an accurate failure rate can be calculated.
  • the prediction model creation unit 12 includes a prediction value calculation unit 121, an evaluation unit 122, and an update unit 123.
  • the predicted value calculation unit 121 estimates the true distribution of the positioning error.
  • the predicted value calculation unit 121 calculates the predicted value of the failure rate based on the estimated positioning error distribution. As will be described later, the predicted value of the failure rate is calculated as a distribution.
  • the evaluation unit 122 evaluates by comparing the failure rate calculated based on the positioning error predicted by the prediction model M with the failure rate calculated from the distribution of the observed positioning error. In this embodiment, the failure rate is used as the evaluation function (loss function) instead of the positioning error.
  • the update unit 123 updates the parameters of the prediction model M based on the evaluation of the evaluation unit 122 with respect to the failure rate based on the prediction model M.
  • the storage unit 13 stores the prediction model M and various data.
  • the positioning error indicates the magnitude of the error of the position information positioned by the satellite positioning system, and indicates that an error can occur in the range of a circle centered on the position information and having the positioning error as a radius, for example.
  • the variation in positioning error is represented by a probability distribution. For example, while traveling a certain point Q 100 times by a vehicle, the position information and the positioning error at the point Q are received from the satellite positioning system.
  • the prediction model M of the present embodiment outputs the probability distribution of the positioning error observed at the point Q when the topographical feature data of the point Q, the speed of the vehicle, and the like are input.
  • the failure rate f is calculated by the following equation (1).
  • f f 2 ⁇ r 2 + f 3 ⁇ r 3 + ... + f 11 ⁇ r 11 ...
  • f k failure rate when the positioning error is k in [m] r k is the true probability that the positioning error becomes k [m].
  • the true probability, number of positioning error of n k is k [m] is observed, the total number of observations is N, the value of n k ⁇ N which is assumed when the N ⁇ ⁇ .
  • the failure rate Since the posterior probability p follows the Dirichlet distribution, the failure rate also has a distribution, and the distribution can be calculated by the following equation (3).
  • f ⁇ (f 2 ⁇ r 2 + f 3 ⁇ r 3 + ⁇ + f 11 ⁇ r 11) ⁇ p (r 2, r 3, ⁇ , r 11
  • the predicted value calculation unit 121 calculates the estimated value of the true positioning error distribution using the formula (2), and calculates the observed value of the failure rate by the formula (3).
  • the probability distribution of the positioning error predicted by the prediction model M and the failure rate f pred calculated from the equation (1) are obtained.
  • the distribution of positioning error was estimated by the Dirichlet distribution, but it may be calculated by the product of the binomial distribution instead of the Dirichlet distribution.
  • evaluation function Assuming that the probability distribution of the positioning error predicted by the prediction model M and the failure rate calculated from the equation (1) are f- pred , the evaluation value L of the f- pred is the equation (4) from the observation result as shown in the following equation (4). It is the cumulative probability of the posterior probability of the failure rate estimated by 3) near the predicted value.
  • b in the equation (4) is a margin of error of the failure rate.
  • the update unit 123 adjusts the parameters of the prediction model M so that the evaluation value L of the equation (4) is maximized.
  • a recurrent neural network RNN
  • Topographical feature data, velocity data, etc. are used as input variables for the prediction model M.
  • the update unit 123 adjusts the parameters based on the distribution of the failure rate.
  • FIGS. 2 and 3 are a first diagram and a second diagram for explaining a method of creating the prediction model of the present disclosure, respectively.
  • FIG. 2 is a diagram when the total number of observations is small
  • FIG. 3 is a diagram when the total number of observations is large.
  • FIG. 2A is a graph showing the positioning error observation frequency space.
  • n 2 and n 3 indicate the number of times the positioning error 2 [m] was observed and the number of times the positioning error 3 [m] was observed, respectively.
  • n 2 and n 3 of n 2 to n 11 are taken out and shown in the figure. This also applies to the following.
  • FIG. 2B is a graph showing the positioning error distribution space.
  • R 2 and r 3 indicate the probability that the positioning error 2 [m] is observed and the probability that the positioning error 3 [m] is observed, respectively.
  • Point X 12 in FIG. 2 (a) shows that the number of times that the positioning error 2 [m] has been observed in n 2, and the number of times that the positioning error 3 [m] is observed is n 3 ..
  • the distribution Y 12 in FIG. 2 (b) shows the distribution of the true posterior probabilities of the positioning error because the number of observations is a combination of the values of the points X 12. That is, the distribution Y 12 shows the Dirichlet distribution according to the equation (2) in the case of the point X 12. Since the total number of observations in FIG. 2 is small, the spread of the distribution Y 12 is large as compared with the case of FIG. 3 described below.
  • FIG. 2C shows a graph showing the distribution W of the failure rate (observation failure rate) calculated from the posterior distribution of the positioning error, and the failure rate f pred based on the predicted value Z 12 by the prediction model M.
  • the vertical axis of FIG. 2C shows the probability, and the horizontal axis shows the failure rate.
  • the predicted value calculation unit 121 Based on the true posterior distribution Y 12 , the predicted value calculation unit 121 has a large number of random numbers (r 1 , r 2 , ..., R 11 ) that follow the Dirichlet distribution by a method such as MCMC, that is, positioning according to the Dirichlet distribution. Generate an error distribution.
  • the predicted value calculation unit 121 calculates the graph W showing the distribution of the observation failure rate from the generated positioning error distribution and the equation (3).
  • the evaluation unit 122 calculates the failure rate f pred from the predicted value Z 12 in FIG. 2 (b) and the equation (1).
  • the width b in FIG. 2C corresponds to the number of observations or a predetermined allowable width, and corresponds to b in the formula (4).
  • the size of the shaded area Lw indicates an evaluation value for the correctness of the prediction, and corresponds to L in the equation (4).
  • Evaluation unit 122 in the vicinity of the failure rate f pred calculated from the predicted value Z 12, to evaluate the correctness of the predicted value Z 12 depending distribution W of the observed failure rate how much of gather.
  • the prediction model M is such that the region Lw included in the width of b centered on the failure rate f pred based on the prediction value Z 12 becomes large.
  • Parameter coefficient of RNN network
  • FIG. 3A shows a positioning error observation frequency space when the number of observations is large.
  • n 2 and n 3 are the number of times the positioning errors 2 [m] and 3 [m] were observed, respectively, but the total number of observations is larger than in the case of FIG.
  • the spread of the positioning error distribution Y 12 becomes small. This indicates that the positioning error distribution Y 12 has converged closer to the true distribution due to the increase in the number of observations.
  • Point Z 12 in FIG. 3B is a predicted value by the prediction model M.
  • the predicted value calculation unit 121 uses the random numbers generated based on the positioning error distribution Y 12 of FIG. 3 (b) and the equation (3) to be used in FIG.
  • the distribution W of the observation failure rate shown in c) is calculated.
  • the evaluation unit 122 calculates the failure rate f pred shown in FIG. 3 (c) by using the predicted value Z 12 in FIG. 3 (b) and the equation (1).
  • the width Wh of the observation failure rate distribution W becomes narrower as the spread of the positioning error distribution Y 12 decreases.
  • the expected value Wc of the distribution W and the failure rate f pred based on the predicted value are separated from each other, and the failure rate f pred is outside the width Wh of the distribution W.
  • the size is smaller than that in FIG. 2 (c).
  • the prediction model creating unit 12 creates a prediction model M for calculating such a predicted value. For example, if a prediction model M is created using the positioning errors observed at points Q1 and Q2 with different topographical features as training data, and the number of observations at point Q1 is small and the number of observations at point Q2 is large, prediction is made.
  • the model creation unit 12 makes sure that the expected value Wc of the distribution W having a spread as illustrated in FIG. 2C and the failure rate fpred are separated by a certain degree. Is allowed, and when the topographical feature data of the point Q2 is input, it is predicted to output a predicted value such that the expected value Wc of the narrow distribution W as illustrated in FIG. 3C and the failure rate f pred approach each other. Create a model.
  • FIG. 4 is a flowchart showing an example of the prediction model creation process of the present disclosure.
  • the data acquisition unit 11 acquires the learning data (step S11).
  • the data acquisition unit 11 has, for each of the plurality of points, the topographical feature data of that point, the speed of the vehicle, the magnitude of the positioning error observed when passing the target point a plurality of times, the number of observations thereof, and the total number of observations.
  • the topographical feature data is, for example, the number and height of buildings existing in each direction around the target point, the route (road shape) on which the vehicle traveled, and the like.
  • Both the terrain feature data and the vehicle speed are factors that affect the reception of signals from the satellite positioning system at the target point.
  • Topographic feature data can be obtained by analyzing three-dimensional map data.
  • the vehicle speed and positioning error the vehicle is actually driven multiple times in the same direction at the same speed at the same target point, and the speed at that time, the position information acquired from the satellite positioning system, and the positioning error information are recorded. Can be obtained by doing.
  • the terrain feature data, the vehicle speed, the number of observations for each positioning error, and the total number of observations are set as one set of learning data.
  • the data acquisition unit 11 acquires learning data for a plurality of points having various topographical features and records them in the storage unit 13.
  • the prediction model creation unit 12 selects a plurality of sets from the learning data recorded in the storage unit 13 (step S12).
  • the prediction model creation unit 12 acquires the terrain feature data and the vehicle speed for each of the selected learning data and inputs them to the prediction model M being trained.
  • the prediction model M predicts the positioning error distribution corresponding to the input terrain feature data and the vehicle speed (step S13).
  • the prediction model M is "a building with a height of 20 m in the north direction, which includes topographical feature data and vehicle speed, and runs on a road passing from west to east on the south side of the building, from west to east at a speed of west to east.
  • the prediction model creation unit 12 evaluates the predicted value (probability distribution of positioning error) (step S14).
  • the prediction value is evaluated using the failure rate, instead of evaluating the difference between the probability distribution of the positioning error output by the prediction model M and the probability distribution of the actually observed positioning error.
  • the predicted value calculation unit 121 uses the number of observations and the total number of observations of each positioning error of the training data, and the equation (2) to distribute the Dirichlet distribution (FIGS. 2 (b) and 3 (b)). Y 12 ) is calculated.
  • the predicted value calculation unit 121 generates a random number based on the Dirichlet distribution, and calculates the distribution of the observation failure rate (W in FIGS.
  • the evaluation unit 122 uses the predicted values (Z 12 in FIGS. 2 (b) and 3 (b)) predicted by the prediction model M during learning in step S13 and the equation (1) to determine the failure rate based on the predicted values.
  • f pred (Fig 2 (b), f pred in FIG. 3 (b)) is calculated.
  • the evaluation unit 122 calculates the evaluation value L (Lw in FIGS. 2 (c) and 3 (c)) by the formula (4).
  • the update unit 123 updates the parameters of the prediction model M so as to maximize the evaluation value L (step S15).
  • the prediction model creation unit 12 determines whether or not the end condition is satisfied (step S16). For example, the prediction model creation unit 12 determines that the end condition is satisfied when the value of the evaluation value L reaches a predetermined target value. Alternatively, the prediction model creation unit 12 determines that the end condition is satisfied when the number of times of learning (the number of times of execution of steps S12 to S15) reaches a predetermined threshold value. If the end condition is not satisfied (step S16; No), the prediction model creation unit 12 repeats the process from step S12. When the end condition is satisfied (step S16; Yes), the prediction model creating unit 12 ends the process of creating the prediction model M. The prediction model creation unit 12 records the learned prediction model M in the storage unit 13.
  • the prediction model M that predicts the probability distribution of the positioning error when the terrain feature data and the vehicle speed are input, but the prediction model creation unit 12 creates the positioning error when the terrain feature data is input. It may be configured to create a prediction model M that predicts the probability distribution of. This also applies to the second embodiment.
  • FIG. 5 is a diagram showing an example of a prediction model creation device according to the second embodiment of the present disclosure.
  • the prediction model creation device 10A includes a data acquisition unit 11, a prediction model creation unit 12, a storage unit 13, and a reliability evaluation unit 14. The description of the same configuration as that of the first embodiment will be omitted.
  • the reliability evaluation unit 14 evaluates the reliability of the probability distribution of the positioning error predicted by the prediction model M.
  • the reliability evaluation unit 14 executes either or both of (1) reliability evaluation based on the uncertainty of the model and (2) reliability evaluation based on the total number of observations during learning.
  • the reliability evaluation unit 14 applies a method called dropout in which some of the elements of the network are intentionally omitted. ..
  • the reliability evaluation unit 14 drops out different networks each time a predetermined number of times, and records the predicted value output by the predicted model M to which the dropout has been applied for the same input.
  • the reliability evaluation unit 14 evaluates the variation of the predicted value output by the prediction model M. When the variation is large, the reliability of the probability distribution of the positioning error predicted by the prediction model M is low, and when the variation is small, the reliability is evaluated as high.
  • the reliability evaluation unit 14 when the spread of the Diricle distribution is large, the reliability of the probability distribution of the positioning error predicted by the prediction model M is low (the total number of observations is small), and when the spread is small (the total number of observations is large). ), The reliability is evaluated as high.
  • FIG. 6 is a diagram illustrating an evaluation method of a prediction model according to the second embodiment of the present disclosure.
  • 6 (a) to 6 (c) are diagrams for explaining the evaluation process when the total number of observations is large.
  • FIG. 6A shows a feature space.
  • the feature quantity is a parameter to be input to the prediction model.
  • Point H 1 is feature value g 1 (e.g., feature values of the terrain) indicated by the value of G 1, characteristic quantity g 2 (e.g., vehicle speed) of the input parameter value of is in G2.
  • FIG. 6B shows a positioning error distribution space.
  • the point I 1 indicates the output when the input parameter of the point H 1 is input to the trained prediction model M, that is, the probability distribution of the positioning error.
  • Distribution J 1 indicates the variability that occurred when the dropout was applied, the variability according to the total number of observations, or both.
  • Reliability evaluation unit 14 applies a drop out trained predictive model M, a predictive value when inputting the input parameters as at point H 1, is recorded in the storage unit 13.
  • the reliability evaluation unit 14 records the calculated posterior probability distribution in the storage unit 13.
  • the distribution J 1 is either or both of a plot of the predicted values recorded when the dropout is applied and a calculated posterior probability distribution.
  • the distribution W 1 of the observation failure rate is calculated based on the distribution J 1 , the distribution as shown in FIG. 6 (c) is obtained.
  • FIG. 6 (d) to 6 (e) are diagrams for explaining the evaluation process when the total number of observations is small.
  • FIG. 6 (d) shows the feature space
  • FIG. 6 (e) shows the positioning error distribution space.
  • Point I 2 is a predicted value when the input parameter of point H 2 is input to the trained prediction model M.
  • Distribution J 2 for the predicted value I 2 is the variation occurred in dropout application, or variations corresponding to all the observation times, or show both.
  • Reliability evaluation unit 14 performs the same process as described in FIG. 6 (b), calculates a distribution J 2 in FIG. 6 (e). When the distribution W 2 of the observation failure rate is calculated based on the distribution J 2 , the distribution as shown in FIG. 6 (f) is obtained.
  • the reliability evaluation unit 14 evaluates the reliability of the predicted value based on the magnitude of the spread of the distribution J 1 and the distribution J 2. For example, the reliability evaluation unit 14 compares the area of the distribution J 1 and the distribution J 2 with a predetermined threshold, and the predicted value predicted by the prediction model M based on the fact that the area of the distribution J 1 is less than the threshold. I 1 is evaluated as highly reliable. For example, the reliability evaluation unit 14 evaluates that the predicted value I 2 predicted by the prediction model M is highly reliable based on the area of the distribution J 2 being equal to or larger than the threshold value.
  • FIG. 7 is a flowchart showing an example of the evaluation process of the prediction model according to the second embodiment of the present disclosure.
  • the reliability evaluation unit 14 evaluates the reliability based on the uncertainty of the model (step S21). With respect to a certain predicted value, the reliability evaluation unit 14 inputs the input parameter n value when the predicted value is obtained into the prediction model M to which the dropout is applied, and acquires the predicted value. The reliability evaluation unit 14 performs the same process a plurality of times by changing the network to drop out, and calculates the variation of the predicted value. Next, the reliability evaluation unit 14 evaluates the reliability based on the total number of observations during learning (step S22).
  • the reliability of the predicted value predicted by the trained prediction model M can be evaluated. For example, when the reliability of the predicted value is low, as shown in FIG. 6 (f), the distribution W 2 of the failure rate calculated based on the predicted value has a certain spread or more. In such a case, a look at the distribution W 2 spread, "this point, the failure rate is because there is a possibility of erroneous recognition of the road uncertain, it is better not to perform processing related to road pricing" a decision such as It can be carried out.
  • FIG. 10 is a diagram showing an example of the hardware configuration of the prediction model creation device according to each embodiment of the present disclosure.
  • the computer 900 is, for example, a PC (Personal Computer) including a CPU 901, a main storage device 902, an auxiliary storage device 903, an input / output interface 904, and a communication interface 905, a server terminal device, and the like.
  • the above-mentioned prediction model creation devices 10 and 10A are mounted on the computer 900.
  • the operation of each processing unit described above is stored in the auxiliary storage device 903 in the form of a program.
  • the CPU 901 reads the program from the auxiliary storage device 903, expands it to the main storage device 902, and executes the above processing according to the program.
  • the CPU 901 secures a storage area corresponding to the storage unit 13 in the main storage device 902 according to the program.
  • the CPU 901 secures a storage area for storing the data being processed in the auxiliary storage device 903 according to the program.
  • the auxiliary storage device 903 is an example of a non-temporary tangible medium.
  • Other examples of non-temporary tangible media include magnetic disks, magneto-optical disks, CD-ROMs, DVD-ROMs, semiconductor memories, etc. connected via the input / output interface 904.
  • the distributed computer 900 may expand the program to the main storage device 902 and execute the above processing.
  • the program may be for realizing a part of the above-mentioned functions.
  • the program may be a so-called difference file (difference program) that realizes the above-mentioned function in combination with another program already stored in the auxiliary storage device 903.
  • the prediction model creation devices 10 and 10A are prediction model creation devices for creating a prediction model for predicting a positioning error indicating an error in the position positioned by the satellite positioning system, and are at a certain point. Acquires topographical feature data indicating topographical features, the positioning error with respect to the position of the moving body observed by the satellite positioning system when the moving body passes through the point a plurality of times, and the number of observations for each positioning error. Data acquisition unit 11 to create a prediction model M that creates a prediction model M that outputs the probability distribution of the positioning error when the topographical feature data is input, and a prediction model creation unit 12 that outputs the probability distribution of the positioning error that the prediction model M outputs.
  • the evaluation value of the failure rate is set for each positioning error.
  • An evaluation unit 122 that calculates based on the number of observations and the observation failure rate calculated based on the function (Equation (3)), and an update unit that updates the prediction model M so that the evaluation value becomes an optimum value. 123 and. This makes it possible to create a prediction model M that predicts a positioning error for which an accurate failure rate can be calculated.
  • the predetermined function indicates a function for calculating the failure rate, and the equations (1) and (3) correspond to the predetermined functions.
  • the prediction model creation devices 10, 10A are the prediction model creation devices 10, 10A of (1), and the number of observations for each positioning error is acquired by the data acquisition unit 11. Based on the distribution of posterior probabilities of the positioning error to be the number of observations and the function (Equation (3)), the distribution W of the observation failure rate is calculated, and the distribution W is used as the failure rate.
  • the predicted value calculation unit 121 which outputs the predicted value of the above, is further provided. This makes it possible to calculate the true distribution of failure rates.
  • the prediction model creation devices 10, 10A are the prediction model creation devices 10, 10A of (1) to (2), and the evaluation unit 122 is based on the prediction model M.
  • the area Lw where the calculated region of the predetermined width b centered on the failure rate and the region occupied by the distribution W indicated by the predicted value of the failure rate overlap is calculated as the evaluation value L, and the update unit 123 ,
  • the parameters of the prediction model M are updated so that the evaluation value L becomes the maximum. This makes it possible to create a prediction model M capable of calculating an accurate failure rate.
  • the prediction model creation device 10A is the prediction model creation device 10A of (1) to (3), and the reliability of the prediction value of the probability distribution of the positioning error output by the prediction model.
  • a reliability evaluation unit 14 for evaluating sex is further provided.
  • the prediction model creation device 10A is the prediction model creation device 10A of (4), and when the prediction model M is constructed by a neural network, the reliability evaluation unit 14 , The reliability is evaluated based on the variation at the time of applying the dropout to the prediction model when the predetermined topographical feature data and the speed of the moving object are input to the prediction model.
  • the prediction model creation device 10A according to the sixth aspect is the prediction model creation device 10A of (4) to (5), and the reliability evaluation unit 14 is the magnitude of the total value of the number of observations.
  • the spread of the posterior probability distribution of the predicted value based on the above is calculated, and the reliability is evaluated based on the spread of the distribution.
  • the prediction model creation method is a prediction model creation method for creating a prediction model for predicting a positioning error indicating an error in the position positioned by the satellite positioning system, and is a feature of the terrain at a certain point.
  • the topographical feature data is input, there is a step of creating a prediction model that outputs the probability distribution of the positioning error, and in the step of creating the prediction model, the probability of the positioning error output by the prediction model.
  • the evaluation value of the failure rate is the number of observations for each positioning error and the above. It is calculated based on the observation failure rate calculated based on the function, and the prediction model is updated so that the evaluation value becomes the optimum value.
  • the program according to the eighth aspect is a prediction model creation method for creating a prediction model for predicting a positioning error indicating an error of a position positioned by a satellite positioning system on a computer, and is a feature of the terrain at a certain point.
  • the topographical feature data is input, there is a step of creating a prediction model that outputs the probability distribution of the positioning error, and in the step of creating the prediction model, the probability of the positioning error output by the prediction model.
  • the evaluation value of the failure rate is the number of observations for each positioning error and the above. The process of updating the prediction model so that the evaluation value becomes the optimum value, which is calculated based on the observation failure rate calculated based on the function, is executed.
  • prediction model creation device prediction model creation method and program, it is possible to create a prediction model that estimates the distribution of the error of the positioned position in the positioning using satellites.

Abstract

This prediction model creation device is for creating a prediction model for predicting positioning errors indicating errors in positions measured by a satellite positioning system. The prediction model creation device performs calculation on the basis of a probability distribution of positioning errors, calculates an evaluated value on the basis of a failure rate that indicates a possibility of occurrence of misrecognition with respect to each position measured by the satellite positioning system and that is based on a predicted value predicted by the prediction model and on the basis of an observation failure rate calculated on the basis of a function and the number of times of observation for each positioning error, and performs learning to obtain a prediction model such that the evaluated value becomes an optimal value.

Description

予測モデル作成装置、予測モデル作成方法及びプログラムPredictive model creation device, predictive model creation method and program
　本発明は、測位誤差分布を予測する予測モデル作成装置、予測モデル作成方法及びプログラムに関する。 The present invention relates to a predictive model creation device for predicting a positioning error distribution, a predictive model creation method, and a program.
　ＧＮＳＳなどの衛星測位システムを利用すると、位置情報と、その位置情報の誤差の大きさを示す測位誤差情報が得られる。特許文献１には、複数の衛星からの擬似距離と、測位された位置から衛星までの距離との誤差により、測位誤差を算出する技術が記載されている。 By using a satellite positioning system such as GNSS, position information and positioning error information indicating the magnitude of the error in the position information can be obtained. Patent Document 1 describes a technique for calculating a positioning error based on an error between a pseudo distance from a plurality of satellites and a distance from a positioned position to the satellite.
　図８を参照する。図８に示す道路Ａを走行している車両が、自車両の走行位置を衛星測位システムによって測位した場合、測位された位置情報の誤差の程度によっては、近くの道路Ｂを走行しているかのような結果が得られる可能性がある。走行する道路に応じて課金を行うような場合、車両が走行する道路の誤認識は問題となる。衛星測位システムが提供する測位誤差情報を利用して、走行中の道路を間違える確率（以下、失敗率と記載する。）を算出し、道路の誤認識に備える技術が提供されている。 Refer to FIG. When a vehicle traveling on the road A shown in FIG. 8 positions the traveling position of its own vehicle by a satellite positioning system, it may be traveling on a nearby road B depending on the degree of error in the positioned position information. Such results may be obtained. In the case of charging according to the road on which the vehicle travels, misrecognition of the road on which the vehicle travels becomes a problem. Utilizing the positioning error information provided by the satellite positioning system, a technique for calculating the probability of making a mistake on the road on which the vehicle is traveling (hereinafter referred to as a failure rate) and preparing for misrecognition of the road is provided.
　図８のＬ１，Ｌ２は共に測位誤差の分布を示している。測位誤差は正規分布し、正規分布の標準偏差を「測位誤差」とする。分布Ｌ１は、分布Ｌ２よりも測位誤差が小さい場合の確率分布を示している。図９を参照する。図９のグラフの縦軸は失敗率、横軸は測位誤差の大きさを示す。図９において、測位誤差Ｒ１と測位誤差Ｒ２の差Ｄ１、測位誤差Ｒ２と測位誤差Ｒ３の差Ｄ２は等しい。しかし、測位誤差Ｒ１のときの失敗率と測位誤差Ｒ２のときの失敗率の差Ｄ３は、測位誤差Ｒ２のときの失敗率と測位誤差Ｒ３のときの失敗率の差Ｄ４と大きく異なる。つまり、測位誤差の大きさと失敗率は比例しない。従って、失敗率を取り扱う場合には測位誤差Ｒ１と測位誤差Ｒ２の差Ｄ１、測位誤差Ｒ２と測位誤差Ｒ３の差Ｄ２を同等として扱うべきではない。 Both L1 and L2 in FIG. 8 show the distribution of positioning error. The positioning error is normally distributed, and the standard deviation of the normal distribution is defined as the "positioning error". The distribution L1 shows a probability distribution when the positioning error is smaller than the distribution L2. See FIG. The vertical axis of the graph of FIG. 9 shows the failure rate, and the horizontal axis shows the magnitude of the positioning error. In FIG. 9, the difference D1 between the positioning error R1 and the positioning error R2 and the difference D2 between the positioning error R2 and the positioning error R3 are equal. However, the difference D3 between the failure rate when the positioning error R1 and the failure rate when the positioning error R2 is R2 is significantly different from the difference D4 between the failure rate when the positioning error R2 and the failure rate when the positioning error R3 is set. That is, the magnitude of the positioning error is not proportional to the failure rate. Therefore, when dealing with the failure rate, the difference D1 between the positioning error R1 and the positioning error R2 and the difference D2 between the positioning error R2 and the positioning error R3 should not be treated as equivalent.
特開２０１９－１５６３５号公報Japanese Unexamined Patent Publication No. 2019-15635
　一般に、予測モデルを作成する場合には、所定の評価関数を最小化、または、最大化するように学習を行う。例えば、最小二乗誤差を評価関数とし、測位誤差の予測値と実測値との誤差を小さくする手法で予測モデルを作成した場合、真の測位誤差の大きさがＲ２のときに測位誤差Ｒ３と予測する予測モデルを作成する可能性もあれば、真の測位誤差の大きさがＲ２のときに測位誤差Ｒ１と予測するような予測モデルを作成する可能性もあり、そのどちらも評価関数の値が同等になってしまう事が考えられる。すると、この予測モデルは、道路Ａと道路Ｂを誤認識する失敗率が実際とは大きく異なるモデルになってしまう事がある。このような予測モデルでは、正確な失敗率を把握することができない。 Generally, when creating a prediction model, learning is performed so as to minimize or maximize a predetermined evaluation function. For example, when a prediction model is created by using the least squares error as an evaluation function and reducing the error between the predicted value and the measured value of the positioning error, the positioning error is predicted to be R3 when the true magnitude of the positioning error is R2. There is a possibility of creating a prediction model that predicts the positioning error R1 when the true magnitude of the positioning error is R2, and both of them have the value of the evaluation function. It is possible that they will be equivalent. Then, this prediction model may become a model in which the failure rate of erroneously recognizing road A and road B is significantly different from the actual model. With such a prediction model, it is not possible to grasp the accurate failure rate.
　本開示は、上述の課題を解決することのできる予測モデル作成装置、予測モデル作成方法及びプログラムを提供する。 The present disclosure provides a predictive model creation device, a predictive model creation method, and a program capable of solving the above-mentioned problems.
　本発明の一態様によれば、予測モデル作成装置は、衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成装置であって、ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するデータ取得部と、前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルを作成する予測モデル作成部と、前記予測モデルが出力する前記測位誤差の確率分布と、所定の関数と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数に基づいて算出される観測失敗率に基づいて算出する評価部と、前記評価値が最適値となるように前記予測モデルを更新する更新部と、を備える。 According to one aspect of the present invention, the prediction model creation device is a prediction model creation device that creates a prediction model that predicts a positioning error indicating an error in the position positioned by the satellite positioning system, and is a feature of the terrain at a certain point. Data acquisition to acquire the topographical feature data indicating the above, the positioning error with respect to the position of the moving object observed by the satellite positioning system when the moving object passes through the point a plurality of times, and the number of observations for each positioning error. A unit, a prediction model creation unit that creates a prediction model that outputs the probability distribution of the positioning error when the topographical feature data is input, a probability distribution of the positioning error output by the prediction model, a predetermined function, and the like. With respect to the failure rate that indicates the magnitude of the possibility of misrecognition of the position, which is calculated based on the above, the evaluation value of the failure rate is calculated based on the number of observations for each positioning error and the function. It includes an evaluation unit that calculates based on the failure rate, and an update unit that updates the prediction model so that the evaluation value becomes an optimum value.
　本発明の一態様によれば、予測モデル作成方法は、衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成方法であって、ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するステップと、前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルを作成するステップと、を有し、前記予測モデルを作成するステップでは、前記予測モデルが出力する前記測位誤差の確率分布と、所定の関数と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数に基づいて算出される観測失敗率に基づいて算出し、前記評価値が最適値となるように前記予測モデルを更新する。 According to one aspect of the present invention, the prediction model creation method is a prediction model creation method for creating a prediction model that predicts a positioning error indicating an error in the position positioned by the satellite positioning system, and is a feature of the terrain at a certain point. The step of acquiring the topographical feature data indicating the above, the positioning error with respect to the position of the moving object observed by the satellite positioning system when the moving object passes through the point a plurality of times, and the number of observations for each positioning error. When the topographical feature data is input, there is a step of creating a prediction model that outputs the probability distribution of the positioning error, and in the step of creating the prediction model, the probability of the positioning error output by the prediction model. Regarding the failure rate, which is calculated based on the distribution and a predetermined function and indicates the magnitude of the possibility of misrecognition of the position, the evaluation value of the failure rate is the number of observations for each positioning error and the above. It is calculated based on the observation failure rate calculated based on the function, and the prediction model is updated so that the evaluation value becomes the optimum value.
　本発明の一態様によれば、プログラムは、コンピュータに、衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成方法であって、ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するステップと、前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルを作成するステップと、を有し、前記予測モデルを作成するステップでは、前記予測モデルが出力する前記測位誤差の確率分布と、所定の関数と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数に基づいて算出される観測失敗率に基づいて算出し、前記評価値が最適値となるように前記予測モデルを更新する処理、を実行させる。 According to one aspect of the present invention, the program is a prediction model creation method for creating a prediction model for predicting a positioning error indicating an error in the position positioned by a satellite positioning system on a computer, and is a feature of the terrain at a certain point. The step of acquiring the topographical feature data indicating the above, the positioning error with respect to the position of the moving body observed by the satellite positioning system when the moving body passes through the point a plurality of times, and the number of observations for each positioning error. When the topographical feature data is input, there is a step of creating a prediction model that outputs the probability distribution of the positioning error, and in the step of creating the prediction model, the probability of the positioning error output by the prediction model. Regarding the failure rate, which is calculated based on the distribution and a predetermined function and indicates the magnitude of the possibility of misrecognition of the position, the evaluation value of the failure rate is the number of observations for each positioning error and the above. The process of updating the prediction model so that the evaluation value becomes the optimum value, which is calculated based on the observation failure rate calculated based on the function, is executed.
　上記した予測モデル作成装置、予測モデル作成方法及びプログラムによれば、衛星を利用した測位における測位された位置の誤差の分布を推定する予測モデルを作成することができる。 According to the above-mentioned prediction model creation device, prediction model creation method and program, it is possible to create a prediction model that estimates the distribution of the error of the positioned position in the positioning using satellites.
本開示の第一実施形態における予測モデル作成装置の一例を示す図である。It is a figure which shows an example of the prediction model making apparatus in 1st Embodiment of this disclosure. 本開示の予測モデルの作成方法を説明する第１の図である。It is the first figure explaining the method of making the prediction model of this disclosure. 本開示の予測モデルの作成方法を説明する第２の図である。It is a 2nd figure explaining the method of making the prediction model of this disclosure. 本開示の予測モデル作成処理の一例を示すフローチャートである。It is a flowchart which shows an example of the prediction model creation process of this disclosure. 本開示の第二実施形態における予測モデル作成装置の一例を示す図である。It is a figure which shows an example of the prediction model making apparatus in the 2nd Embodiment of this disclosure. 本開示の第二実施形態における予測モデルの評価方法について説明する図である。It is a figure explaining the evaluation method of the prediction model in the 2nd Embodiment of this disclosure. 本開示の第二実施形態における予測モデルの評価処理の一例を示すフローチャートである。It is a flowchart which shows an example of the evaluation process of the prediction model in the 2nd Embodiment of this disclosure. 測位誤差と失敗率について説明する第１図である。FIG. 1 is a diagram illustrating a positioning error and a failure rate. 測位誤差と失敗率について説明する第２図である。FIG. 2 is a diagram illustrating a positioning error and a failure rate. 本発明の各実施形態における予測モデル作成装置のハードウェア構成の一例を示す図である。It is a figure which shows an example of the hardware configuration of the prediction model creation apparatus in each embodiment of this invention.
＜第一実施形態＞
　以下、本発明の第一実施形態による測位誤差の予測方法について図１～図４を参照して説明する。
（構成）
　図１は、本発明の第一実施形態における予測モデル作成装置の一例を示す図である。図１に示すように予測モデル作成装置１０は、データ取得部１１と、予測モデル作成部１２と、記憶部１３と、を備える。 <First Embodiment>
Hereinafter, a method for predicting a positioning error according to the first embodiment of the present invention will be described with reference to FIGS. 1 to 4.
(composition)
FIG. 1 is a diagram showing an example of a prediction model creation device according to the first embodiment of the present invention. As shown in FIG. 1, the prediction model creation device 10 includes a data acquisition unit 11, a prediction model creation unit 12, and a storage unit 13.
　データ取得部１１は、任意の地点における衛星測位システムの測位誤差の確率分布を予測する予測モデルＭの作成に必要な学習データを取得する。学習データには、例えば、複数の地点のそれぞれについて、その地点の地形の特徴を示す地形特徴データ、その地点を車両で走行したときの速度データ、その地点を複数回走行した際の測位誤差ごとの観測回数、全走行回数（全観測回数）などの情報が含まれている。 The data acquisition unit 11 acquires the training data necessary for creating the prediction model M that predicts the probability distribution of the positioning error of the satellite positioning system at an arbitrary point. The training data includes, for example, topographical feature data indicating the topographical features of each of a plurality of points, speed data when the vehicle travels at that point, and positioning error when the point is traveled multiple times. It contains information such as the number of observations and the total number of runs (total number of observations).
　予測モデル作成部１２は、正確な失敗率が算出できるような衛星測位システムの測位誤差の確率分布を予測する予測モデルＭを作成する。予測モデルＭは、例えば、測位誤差＝２［ｍ］が観測される確率“ｒ２”、測位誤差＝３［ｍ］が観測される確率“ｒ３”、・・・、測位誤差＝１１［ｍ］が観測される確率“ｒ１１”といった情報を出力する。予測モデル作成部１２は、予測値算出部１２１と、評価部１２２と、更新部１２３と、を備える。 The prediction model creation unit 12 creates a prediction model M that predicts the probability distribution of the positioning error of the satellite positioning system so that an accurate failure rate can be calculated. In the prediction model M, for example, the probability that a positioning error = 2 [m] is observed "r 2 ", the probability that a positioning error = 3 [m] is observed "r 3 ", ..., The positioning error = 11 [ It outputs information such as the probability "r 11 " that [m] is observed. The prediction model creation unit 12 includes a prediction value calculation unit 121, an evaluation unit 122, and an update unit 123.
　予測値算出部１２１は、測位誤差の真の分布を推定する。予測値算出部１２１は、推定した測位誤差分布に基づいて、失敗率の予測値を算出する。後述するように失敗率の予測値は、分布として算出される。
　評価部１２２は、予測モデルＭが予測する測位誤差に基づいて算出された失敗率を、観測された測位誤差の分布から算出された失敗率と比較して評価を行う。本実施形態では、測位誤差ではなく、失敗率を評価関数(損失関数)として用いる。
　更新部１２３は、予測モデルＭに基づく失敗率に対する評価部１２２の評価に基づいて、予測モデルＭのパラメータを更新する。 The predicted value calculation unit 121 estimates the true distribution of the positioning error. The predicted value calculation unit 121 calculates the predicted value of the failure rate based on the estimated positioning error distribution. As will be described later, the predicted value of the failure rate is calculated as a distribution.
The evaluation unit 122 evaluates by comparing the failure rate calculated based on the positioning error predicted by the prediction model M with the failure rate calculated from the distribution of the observed positioning error. In this embodiment, the failure rate is used as the evaluation function (loss function) instead of the positioning error.
The update unit 123 updates the parameters of the prediction model M based on the evaluation of the evaluation unit 122 with respect to the failure rate based on the prediction model M.
　記憶部１３は、予測モデルＭや各種データを記憶する。 The storage unit 13 stores the prediction model M and various data.
（測位誤差について）
　衛星測位システムを用いた測位では、同一地点で複数回の測位を行っても、測位誤差は一定ではなくばらつきが生じる。測位誤差とは、衛星測位システムが測位した位置情報の誤差の大きさを示し、例えば、位置情報を中心、測位誤差を半径とする円の範囲で誤差が生じ得ることを示す。本実施形態では、測位誤差のばらつきを確率分布で表す。例えば、ある地点Ｑを車両で１００回走行しながら、地点Ｑにおける位置情報と測位誤差を衛星測位システムから受信する。例えば、１００回のうち測位誤差２［ｍ］が３０回、測位誤差３［ｍ］が２０回、・・・、測位誤差１１［ｍ］が1回観測された場合、地点Ｐの測位誤差が２［ｍ］である確率は３０％、３［ｍ］である確率は２０％、・・・、１１［ｍ］である確率は１％となる。本実施形態の予測モデルＭは、地点Ｑの地形特徴データ、車両の速度などを入力すると、地点Ｑで観測される測位誤差の確率分布を出力する。 (About positioning error)
In positioning using a satellite positioning system, even if positioning is performed a plurality of times at the same point, the positioning error is not constant and varies. The positioning error indicates the magnitude of the error of the position information positioned by the satellite positioning system, and indicates that an error can occur in the range of a circle centered on the position information and having the positioning error as a radius, for example. In this embodiment, the variation in positioning error is represented by a probability distribution. For example, while traveling a certain point Q 100 times by a vehicle, the position information and the positioning error at the point Q are received from the satellite positioning system. For example, if the positioning error 2 [m] is observed 30 times, the positioning error 3 [m] is 20 times, ..., And the positioning error 11 [m] is observed once out of 100 times, the positioning error at the point P is The probability of being 2 [m] is 30%, the probability of being 3 [m] is 20%, ..., The probability of being 11 [m] is 1%. The prediction model M of the present embodiment outputs the probability distribution of the positioning error observed at the point Q when the topographical feature data of the point Q, the speed of the vehicle, and the like are input.
（失敗率の算出方法）
　失敗率ｆは、次の式（１）で算出する。
　ｆ＝ｆ２・ｒ２＋ｆ３・ｒ３＋・・・＋ｆ１１・ｒ１１・・・（１）
　ここで、ｆｋは測位誤差がｋ［ｍ］のときの失敗率、ｒｋは測位誤差がｋ［ｍ］となる真の確率である。真の確率とは、ｎｋを測位誤差がｋ［ｍ］が観測された回数、観測回数の合計をＮとし、Ｎ→∞とした際に想定されるｎｋ÷Ｎの値である。 (Calculation method of failure rate)
The failure rate f is calculated by the following equation (1).
f = f 2 · r 2 + f 3 · r 3 + ... + f 11 · r 11 ... (1)
Here, f k failure rate when the positioning error is k in [m], r k is the true probability that the positioning error becomes k [m]. The true probability, number of positioning error of n k is k [m] is observed, the total number of observations is N, the value of n k ÷ N which is assumed when the N → ∞.
（測位誤差分布の推定）
　ｒｋは一般には観測できないため、ｎｋ、Ｎの値から推定する。ｎｋ(ｋ＝２，３、・・・１１)、Ｎが既知である場合に、真の確率がｒｋである事後確率ｐ（ｒ２，ｒ３，・・・，ｒ１１｜ｎ２，ｎ３，・・・，ｎ１１）は、ディリクレ分布に従う。事後確率ｐを以下の式（２）に示す。 (Estimation of positioning error distribution)
Since r k is generally unobservable, it is estimated from the values of n k and N. n k (k = 2,3, ··· 11), if N is known, the posterior probability p (r2, r3 true probability is r k, ···, r11 | n2 , n3, · ..., N 11 ) follows the Dirichlet distribution. The posterior probability p is shown in the following equation (2).
Figure JPOXMLDOC01-appb-M000001
Figure JPOXMLDOC01-appb-M000001
（失敗率の真値）
　事後確率ｐがディリクレ分布に従うので、失敗率も分布を持ち、その分布は、以下の式（３）で算出することができる。
　ｆ＝∫（ｆ２・ｒ２＋ｆ３・ｒ３＋・・・＋ｆ１１・ｒ１１）×ｐ（ｒ２，ｒ３，・・・，ｒ１１｜ｎ２，ｎ３，・・・，ｎ１１）ｄｒ２ｄｒ３・・・ｄｒ１１・・・（３）
　予測値算出部１２１は、式（２）を用いて真の測位誤差分布の推定値を算出し、式（３）によって失敗率の観測値を算出する。さらに本実施形態では、失敗率の真値を式（３）で算出されるｆと仮定し、予測モデルＭが予測する測位誤差の確率分布と式（１）から算出される失敗率ｆｐｒｅｄを、式（３）で算出されるｆに基づいて評価する。
　なお、式（２）ではディリクレ分布によって測位誤差の分布を推定したが、ディリクレ分布ではなく2項分布の積によって算出してもよい。 (True value of failure rate)
Since the posterior probability p follows the Dirichlet distribution, the failure rate also has a distribution, and the distribution can be calculated by the following equation (3).
f = ∫ (f 2 · r 2 + f 3 · r 3 + ··· + f 11 · r 11) × p (r 2, r 3, ···, r 11 | n 2, n 3, ···, n 11 ) dr 2 dr 3 ... dr 11 ... (3)
The predicted value calculation unit 121 calculates the estimated value of the true positioning error distribution using the formula (2), and calculates the observed value of the failure rate by the formula (3). Further, in the present embodiment, assuming that the true value of the failure rate is f calculated by the equation (3), the probability distribution of the positioning error predicted by the prediction model M and the failure rate f pred calculated from the equation (1) are obtained. , Evaluate based on f calculated by equation (3).
In Eq. (2), the distribution of positioning error was estimated by the Dirichlet distribution, but it may be calculated by the product of the binomial distribution instead of the Dirichlet distribution.
（評価関数）
　予測モデルＭが予測する測位誤差の確率分布と式（１）から算出される失敗率をｆｐｒｅｄとすると、ｆｐｒｅｄの評価値Ｌを以下の式（４）のように、観測結果から式（３）によって推定される失敗率の事後確率の、予測値近傍での累積確率とする。ここで、式（４）のｂは、失敗率の許容誤差である。 (Evaluation function)
Assuming that the probability distribution of the positioning error predicted by the prediction model M and the failure rate calculated from the equation (1) are f- pred , the evaluation value L of the f- pred is the equation (4) from the observation result as shown in the following equation (4). It is the cumulative probability of the posterior probability of the failure rate estimated by 3) near the predicted value. Here, b in the equation (4) is a margin of error of the failure rate.
Figure JPOXMLDOC01-appb-M000002
Figure JPOXMLDOC01-appb-M000002
　更新部１２３は、式（４）の評価値Ｌが最大化するように、予測モデルＭのパラメータを調整する。予測モデルＭには、再起型ニューラルネットワーク(Recurrent Neural Network: RNN)を用いることができる。予測モデルＭへの入力変数として、地形特徴データ、速度データ等を用いる。評価値Ｌを最大化する際には、予測値算出部１２１は、確率をそのまま扱うのではなく、マルコフ連鎖モンテカルロ(Markov Chain Monte Carlo: MCMC)法などにより、事後確率ｐの分布に従って乱数を生成して複数のｒｋ（ｋ＝２～１１）の組みを生成し、生成されたｒｋの組みを用いて失敗率の分布を算出する。更新部１２３は、この失敗率の分布に基づいてパラメータの調整を行う。 The update unit 123 adjusts the parameters of the prediction model M so that the evaluation value L of the equation (4) is maximized. A recurrent neural network (RNN) can be used as the prediction model M. Topographical feature data, velocity data, etc. are used as input variables for the prediction model M. When maximizing the evaluation value L, the predicted value calculation unit 121 does not handle the probability as it is, but generates a random number according to the distribution of the posterior probability p by the Markov Chain Monte Carlo (MCMC) method or the like. and generating a set of a plurality of r k (k = 2 ~ 11 ), calculates a distribution of failure rate by using the set of generated r k. The update unit 123 adjusts the parameters based on the distribution of the failure rate.
　次に図２、図３を用いて予測モデルＭの作成方法について説明する。
　図２、図３は、それぞれ、本開示の予測モデルの作成方法を説明する第１の図、第２の図である。図２は全観測回数が少ない場合、図３は全観測回数が多い場合の図である。
　図２（ａ）は、測位誤差観測回数空間を示すグラフである。ｎ２、ｎ３はそれぞれ、測位誤差２［ｍ］が観測された回数、測位誤差３［ｍ］が観測された回数を示す。説明の便宜のため、ｎ２～ｎ１１のうちのｎ２、ｎ３を取り出して図示している。このことは、以下においても同様である。図２（ｂ）は、測位誤差分布空間を示すグラフである。ｒ２、ｒ３はそれぞれ、測位誤差２［ｍ］が観測される確率、測位誤差３［ｍ］が観測される確率を示す。 Next, a method of creating the prediction model M will be described with reference to FIGS. 2 and 3.
2 and 3 are a first diagram and a second diagram for explaining a method of creating the prediction model of the present disclosure, respectively. FIG. 2 is a diagram when the total number of observations is small, and FIG. 3 is a diagram when the total number of observations is large.
FIG. 2A is a graph showing the positioning error observation frequency space. n 2 and n 3 indicate the number of times the positioning error 2 [m] was observed and the number of times the positioning error 3 [m] was observed, respectively. For convenience of explanation, n 2 and n 3 of n 2 to n 11 are taken out and shown in the figure. This also applies to the following. FIG. 2B is a graph showing the positioning error distribution space. R 2 and r 3 indicate the probability that the positioning error 2 [m] is observed and the probability that the positioning error 3 [m] is observed, respectively.
　図２（ａ）の点Ｘ１２は、測位誤差２［ｍ］が観測された回数がｎ２で、且つ、測位誤差３［ｍ］が観測された回数がｎ３であることを示している。図２（ｂ）の分布Ｙ１２は、観測回数が点Ｘ１２の値の組合せとなるための測位誤差の真の事後確率の分布を示す。つまり、分布Ｙ１２は、点Ｘ１２の場合の式（２）によるディリクレ分布を示す。図２は全観測回数が少ないため、次に説明する図３の場合と比較して、分布Ｙ１２の広がりが大きい。図２（ｂ）の点Ｚ１２は、観測回数がｎ２且つｎ３の場合の、学習中の予測モデルＭによる確率分布（確率ｒ２、ｒ３）の予測値を示す。ここで、予測値Ｚ１２は分布Ｙ１２の中心から離れているが、観測回数が少なく事後分布の広がりが大きいことから一定程度確からしい予測であると評価される。 Point X 12 in FIG. 2 (a) shows that the number of times that the positioning error 2 [m] has been observed in n 2, and the number of times that the positioning error 3 [m] is observed is n 3 .. The distribution Y 12 in FIG. 2 (b) shows the distribution of the true posterior probabilities of the positioning error because the number of observations is a combination of the values of the points X 12. That is, the distribution Y 12 shows the Dirichlet distribution according to the equation (2) in the case of the point X 12. Since the total number of observations in FIG. 2 is small, the spread of the distribution Y 12 is large as compared with the case of FIG. 3 described below. Point Z 12 in FIG. 2 (b), when the number of observations is n 2 and n 3, shows the predicted value of the probability distribution with predictive model M in the training (probability r 2, r 3). Here, the predicted value Z 12 is far from the center of the distribution Y 12 , but it is evaluated to be a certain degree of certainty because the number of observations is small and the posterior distribution spreads widely.
　図２（ｃ）に、測位誤差の事後分布から算出される失敗率（観測失敗率）の分布Ｗを示すグラフと、予測モデルＭによる予測値Ｚ１２に基づく失敗率ｆｐｒｅｄを示す。図２（ｃ）の縦軸は確率、横軸は失敗率を示す。予測値算出部１２１は、真の事後分布Ｙ１２に基づいて、ＭＣＭＣ等の方法によってディリクレ分布に従う乱数として多数の（ｒ１，ｒ２、・・・、ｒ１１）、つまり、ディリクレ分布に従う測位誤差分布を生成する。予測値算出部１２１は、生成した測位誤差分布と、式（３）から観測失敗率の分布を示すグラフＷを算出する。評価部１２２は、図２（ｂ）の予測値Ｚ１２と式（１）から失敗率ｆｐｒｅｄを算出する。図２（ｃ）の幅ｂは、観測回数に応じた、または、所定の許容幅を示し、式（４）のｂに相当する。斜線を記した領域Ｌｗの大きさは、予測の正しさに対する評価値を示し、式（４）のＬに相当する。評価部１２２は、予測値Ｚ１２から算出された失敗率ｆｐｒｅｄの近傍に、観測失敗率の分布Ｗがどの程度の集まるかによって予測値Ｚ１２の正しさの評価を行う。更新部１２３は、グラフＷと失敗率の座標軸で形成される領域のうち、予測値Ｚ１２に基づく失敗率ｆｐｒｅｄを中心とするｂの幅に含まれる領域Ｌｗが大きくなるように予測モデルＭのパラメータ（ＲＮＮのネットワークの係数）を学習する。 FIG. 2C shows a graph showing the distribution W of the failure rate (observation failure rate) calculated from the posterior distribution of the positioning error, and the failure rate f pred based on the predicted value Z 12 by the prediction model M. The vertical axis of FIG. 2C shows the probability, and the horizontal axis shows the failure rate. Based on the true posterior distribution Y 12 , the predicted value calculation unit 121 has a large number of random numbers (r 1 , r 2 , ..., R 11 ) that follow the Dirichlet distribution by a method such as MCMC, that is, positioning according to the Dirichlet distribution. Generate an error distribution. The predicted value calculation unit 121 calculates the graph W showing the distribution of the observation failure rate from the generated positioning error distribution and the equation (3). The evaluation unit 122 calculates the failure rate f pred from the predicted value Z 12 in FIG. 2 (b) and the equation (1). The width b in FIG. 2C corresponds to the number of observations or a predetermined allowable width, and corresponds to b in the formula (4). The size of the shaded area Lw indicates an evaluation value for the correctness of the prediction, and corresponds to L in the equation (4). Evaluation unit 122, in the vicinity of the failure rate f pred calculated from the predicted value Z 12, to evaluate the correctness of the predicted value Z 12 depending distribution W of the observed failure rate how much of gather. In the update unit 123, among the regions formed by the graph W and the coordinate axes of the failure rate, the prediction model M is such that the region Lw included in the width of b centered on the failure rate f pred based on the prediction value Z 12 becomes large. Parameter (coefficient of RNN network) is learned.
　図３（ａ）に観測回数が多い場合の測位誤差観測回数空間を示す。ｎ２、ｎ３はそれぞれ、測位誤差２［ｍ］、３［ｍ］が観測された回数であるが、図（ａ）の場合よりも全観測回数が多い。この場合、図３（ｂ）の測位誤差分布空間では、測位誤差分布Ｙ１２の広がりが小さくなる。これは、観測回数の増加により、測位誤差分布Ｙ１２がより真の分布に近づき収束したことを示している。図３（ｂ）の点Ｚ１２は、予測モデルＭによる予測値である。図２（ｃ）を用いて説明したように、予測値算出部１２１は、図３（ｂ）の測位誤差分布Ｙ１２に基づいて生成された乱数と式（３）を用いて、図３（ｃ）に示す観測失敗率の分布Ｗを算出する。評価部１２２は、図３（ｂ）の予測値Ｚ１２と式（１）を用いて、図３（ｃ）に示す失敗率ｆｐｒｅｄを算出する。図３（ｃ）に示すように、全観測回数が多い場合、測位誤差分布Ｙ１２の広がりの低下に合わせて、観測失敗率の分布Ｗの幅Ｗｈも狭くなる。図３（ｃ）の例では、分布Ｗの期待値Ｗｃと予測値に基づく失敗率ｆｐｒｅｄの距離が離れていて、失敗率ｆｐｒｅｄが分布Ｗの幅Ｗｈの外にあるため、領域Ｌｗの大きさが図２（ｃ）の場合と比較して小さい。仮に、失敗率ｆｐｒｅｄが分布Ｗの期待値Ｗｃ付近となれば、分布Ｗの幅Ｗｈは、許容幅ｂ内に含まれることとなり、領域Ｌｗの面積（式（４）の評価値Ｌ）は最大となる。このような状態となるときの予測値Ｚ１２が目標とする予測値であり、予測モデル作成部１２は、このような予測値を算出する予測モデルＭを作成する。例えば、地形の特徴が異なる地点Ｑ１と地点Ｑ２で観測された測位誤差などを学習データとして予測モデルＭを作成し、地点Ｑ１での観測回数が少なく、地点Ｑ２での観測回数が多い場合、予測モデル作成部１２は、地点Ｑ１の地形特徴データが入力されたときには、図２（ｃ）に例示するような広がりを持った分布Ｗの期待値Ｗｃと失敗率ｆｐｒｅｄが一定程度離れていることが許容され、且つ、地点Ｑ２の地形特徴データが入力されたときには、図３（ｃ）に例示するような狭い分布Ｗの期待値Ｗｃと失敗率ｆｐｒｅｄが近づくような予測値を出力する予測モデルを作成する。 FIG. 3A shows a positioning error observation frequency space when the number of observations is large. n 2 and n 3 are the number of times the positioning errors 2 [m] and 3 [m] were observed, respectively, but the total number of observations is larger than in the case of FIG. In this case, in the positioning error distribution space of FIG. 3B, the spread of the positioning error distribution Y 12 becomes small. This indicates that the positioning error distribution Y 12 has converged closer to the true distribution due to the increase in the number of observations. Point Z 12 in FIG. 3B is a predicted value by the prediction model M. As described with reference to FIG. 2 (c), the predicted value calculation unit 121 uses the random numbers generated based on the positioning error distribution Y 12 of FIG. 3 (b) and the equation (3) to be used in FIG. 3 (3). The distribution W of the observation failure rate shown in c) is calculated. The evaluation unit 122 calculates the failure rate f pred shown in FIG. 3 (c) by using the predicted value Z 12 in FIG. 3 (b) and the equation (1). As shown in FIG. 3C, when the total number of observations is large, the width Wh of the observation failure rate distribution W becomes narrower as the spread of the positioning error distribution Y 12 decreases. In the example of FIG. 3C, the expected value Wc of the distribution W and the failure rate f pred based on the predicted value are separated from each other, and the failure rate f pred is outside the width Wh of the distribution W. The size is smaller than that in FIG. 2 (c). If the failure rate f pred is close to the expected value Wc of the distribution W, the width Wh of the distribution W is included in the permissible width b, and the area of the region Lw (evaluation value L of the equation (4)) is It becomes the maximum. The predicted value Z 12 in such a state is the target predicted value, and the prediction model creating unit 12 creates a prediction model M for calculating such a predicted value. For example, if a prediction model M is created using the positioning errors observed at points Q1 and Q2 with different topographical features as training data, and the number of observations at point Q1 is small and the number of observations at point Q2 is large, prediction is made. When the topographical feature data of the point Q1 is input, the model creation unit 12 makes sure that the expected value Wc of the distribution W having a spread as illustrated in FIG. 2C and the failure rate fpred are separated by a certain degree. Is allowed, and when the topographical feature data of the point Q2 is input, it is predicted to output a predicted value such that the expected value Wc of the narrow distribution W as illustrated in FIG. 3C and the failure rate f pred approach each other. Create a model.
　次に予測モデル作成処理の流れについて説明する。
　図４は、本開示の予測モデル作成処理の一例を示すフローチャートである。
　まず、データ取得部１１が、学習データを取得する（ステップＳ１１）。例えば、データ取得部１１は、複数の地点のそれぞれについて、その地点の地形特徴データ、車両の速度、対象地点を複数回通過した際に観測された測位誤差の大きさとその観測回数、全観測回数を取得する。地形特徴データとは、例えば、対象地点を中心として各方向に存在する建物の数やその高さ、車両が走行した経路（道路の形状）などである。地形特徴データ、車両の速度は何れも対象地点での衛星測位システムからの信号の受信に影響を及ぼす要素である。地形特徴データは、３次元の地図データを解析して得ることができる。車両の速度と測位誤差については、同一の対象地点について、実際に車両を同一方向に同一の速度で複数回走行させ、そのときの速度、衛星測位システムから取得した位置情報、測位誤差情報を記録することによって得ることができる。本実施形態では、地形特徴データ、車両の速度と、測位誤差ごとの観測回数、全観測回数を１組みの学習データとする。例えば、ある地点を１０回走行して、測位誤差２［ｍ］が５回、３［ｍ］が３回、４［ｍ］が２回観測された場合、測位誤差ごとの観測回数とは、「測位誤差２［ｍ］が５回、３［ｍ］が３回、４［ｍ］が２回」という情報であり、全観測回数は１０回である。データ取得部１１は、様々な地形の特徴を有する複数の地点について、学習データを取得し、記憶部１３に記録する。 Next, the flow of the prediction model creation process will be described.
FIG. 4 is a flowchart showing an example of the prediction model creation process of the present disclosure.
First, the data acquisition unit 11 acquires the learning data (step S11). For example, the data acquisition unit 11 has, for each of the plurality of points, the topographical feature data of that point, the speed of the vehicle, the magnitude of the positioning error observed when passing the target point a plurality of times, the number of observations thereof, and the total number of observations. To get. The topographical feature data is, for example, the number and height of buildings existing in each direction around the target point, the route (road shape) on which the vehicle traveled, and the like. Both the terrain feature data and the vehicle speed are factors that affect the reception of signals from the satellite positioning system at the target point. Topographic feature data can be obtained by analyzing three-dimensional map data. Regarding the vehicle speed and positioning error, the vehicle is actually driven multiple times in the same direction at the same speed at the same target point, and the speed at that time, the position information acquired from the satellite positioning system, and the positioning error information are recorded. Can be obtained by doing. In the present embodiment, the terrain feature data, the vehicle speed, the number of observations for each positioning error, and the total number of observations are set as one set of learning data. For example, when a certain point is traveled 10 times and a positioning error of 2 [m] is observed 5 times, 3 [m] is observed 3 times, and 4 [m] is observed 2 times, the number of observations for each positioning error is The information is that "positioning error 2 [m] is 5 times, 3 [m] is 3 times, and 4 [m] is 2 times", and the total number of observations is 10. The data acquisition unit 11 acquires learning data for a plurality of points having various topographical features and records them in the storage unit 13.
　次に予測モデル作成部１２が、記憶部１３に記録された学習データの中から複数組みを選択する（ステップＳ１２）。次に予測モデル作成部１２は、選択した学習データのそれぞれについて、地形特徴データと車両の速度を取得して、学習中の予測モデルＭに入力する。予測モデルＭは、入力された地形特徴データと車両の速度に対応する測位誤差分布を予測する（ステップＳ１３）。例えば、予測モデルＭは、地形特徴データと車両の速度が含まれる「北の方向に高さ２０ｍの建物等が存在し、その建物の南側を西から東へ通る道路を、西から東へ時速４０ｋｍで走行する」という入力パラメータを取得すると、同様の地形の特徴を有する地点を同じ条件で走行した際に、実際に観測される測位誤差の確率分布に近い値を出力する。 Next, the prediction model creation unit 12 selects a plurality of sets from the learning data recorded in the storage unit 13 (step S12). Next, the prediction model creation unit 12 acquires the terrain feature data and the vehicle speed for each of the selected learning data and inputs them to the prediction model M being trained. The prediction model M predicts the positioning error distribution corresponding to the input terrain feature data and the vehicle speed (step S13). For example, the prediction model M is "a building with a height of 20 m in the north direction, which includes topographical feature data and vehicle speed, and runs on a road passing from west to east on the south side of the building, from west to east at a speed of west to east. When the input parameter "traveling at 40 km" is acquired, a value close to the probability distribution of the actually observed positioning error is output when traveling at a point having the same terrain characteristics under the same conditions.
　次に予測モデル作成部１２は、予測値（測位誤差の確率分布）の評価を行う（ステップＳ１４）。上述のとおり、本開示では、予測モデルＭが出力する測位誤差の確率分布と、実際に観測された測位誤差の確率分布との差を評価するのではなく、失敗率を用いて予測値の評価を行う。具体的には、まず、予測値算出部１２１が、学習データの各測位誤差の観測回数と全観測回数、式（２）を用いてディリクレ分布（図２（ｂ）、図３（ｂ）のＹ１２）を算出する。次に予測値算出部１２１は、ディリクレ分布に基づく乱数を発生させ、乱数と式（３）を用いて、観測失敗率の分布（図２（ｃ）、図３（ｃ）のＷ）を算出する。次に評価部１２２が、学習中の予測モデルＭがステップＳ１３で予測した予測値（図２（ｂ）、図３（ｂ）のＺ１２）と式（１）により、予測値に基づく失敗率ｆｐｒｅｄ（図２（ｂ）、図３（ｂ）のｆｐｒｅｄ）を算出する。そして、評価部１２２は、式（４）により、評価値Ｌ（図２（ｃ）、図３（ｃ）のＬｗ）を算出する。 Next, the prediction model creation unit 12 evaluates the predicted value (probability distribution of positioning error) (step S14). As described above, in the present disclosure, the prediction value is evaluated using the failure rate, instead of evaluating the difference between the probability distribution of the positioning error output by the prediction model M and the probability distribution of the actually observed positioning error. I do. Specifically, first, the predicted value calculation unit 121 uses the number of observations and the total number of observations of each positioning error of the training data, and the equation (2) to distribute the Dirichlet distribution (FIGS. 2 (b) and 3 (b)). Y 12 ) is calculated. Next, the predicted value calculation unit 121 generates a random number based on the Dirichlet distribution, and calculates the distribution of the observation failure rate (W in FIGS. 2 (c) and 3 (c)) using the random number and the equation (3). do. Next, the evaluation unit 122 uses the predicted values (Z 12 in FIGS. 2 (b) and 3 (b)) predicted by the prediction model M during learning in step S13 and the equation (1) to determine the failure rate based on the predicted values. f pred (Fig 2 (b), f pred in FIG. 3 (b)) is calculated. Then, the evaluation unit 122 calculates the evaluation value L (Lw in FIGS. 2 (c) and 3 (c)) by the formula (4).
　次に更新部１２３が、評価値Ｌを最大化するように予測モデルＭのパラメータを更新する（ステップＳ１５）。次に予測モデル作成部１２は、終了条件が成立するかどうかを判定する（ステップＳ１６）。例えば、予測モデル作成部１２は、評価値Ｌの値が所定の目標値に達したら終了条件が成立すると判定する。あるいは、予測モデル作成部１２は、学習の回数（ステップＳ１２～Ｓ１５の実行回数）が所定の閾値に達したら終了条件が成立すると判定する。終了条件が成立しない場合（ステップＳ１６；Ｎｏ）、予測モデル作成部１２は、ステップＳ１２からの処理を繰り返す。終了条件が成立する場合（ステップＳ１６；Ｙｅｓ）、予測モデル作成部１２は、予測モデルＭの作成処理を終了する。予測モデル作成部１２は、学習済みの予測モデルＭを記憶部１３に記録する。 Next, the update unit 123 updates the parameters of the prediction model M so as to maximize the evaluation value L (step S15). Next, the prediction model creation unit 12 determines whether or not the end condition is satisfied (step S16). For example, the prediction model creation unit 12 determines that the end condition is satisfied when the value of the evaluation value L reaches a predetermined target value. Alternatively, the prediction model creation unit 12 determines that the end condition is satisfied when the number of times of learning (the number of times of execution of steps S12 to S15) reaches a predetermined threshold value. If the end condition is not satisfied (step S16; No), the prediction model creation unit 12 repeats the process from step S12. When the end condition is satisfied (step S16; Yes), the prediction model creating unit 12 ends the process of creating the prediction model M. The prediction model creation unit 12 records the learned prediction model M in the storage unit 13.
　本実施形態によれば、失敗率を評価関数の指標とすることにより、走行する道路を正しく認識することができるか否かを適切に評価することができる測位誤差の予測モデルを作成することができる。
　なお、上記例では、地形特徴データと車両の速度を入力すると測位誤差の確率分布を予測する予測モデルＭを作成することとしたが、予測モデル作成部１２は、地形特徴データを入力すると測位誤差の確率分布を予測する予測モデルＭを作成するように構成されていてもよい。このことは、第二実施形態についても同様である。 According to the present embodiment, by using the failure rate as an index of the evaluation function, it is possible to create a positioning error prediction model that can appropriately evaluate whether or not the road on which the vehicle is traveling can be correctly recognized. can.
In the above example, it is decided to create the prediction model M that predicts the probability distribution of the positioning error when the terrain feature data and the vehicle speed are input, but the prediction model creation unit 12 creates the positioning error when the terrain feature data is input. It may be configured to create a prediction model M that predicts the probability distribution of. This also applies to the second embodiment.
＜第二実施形態＞
　次に図５～図７を参照して、第二実施形態における予測モデル作成装置について説明する。
　図５は、本開示の第二実施形態における予測モデル作成装置の一例を示す図である。予測モデル作成装置１０Ａは、データ取得部１１と、予測モデル作成部１２と、記憶部１３と、信頼性評価部１４と、を備える。第一実施形態と同様の構成については説明を省略する。 <Second embodiment>
Next, the prediction model creation device according to the second embodiment will be described with reference to FIGS. 5 to 7.
FIG. 5 is a diagram showing an example of a prediction model creation device according to the second embodiment of the present disclosure. The prediction model creation device 10A includes a data acquisition unit 11, a prediction model creation unit 12, a storage unit 13, and a reliability evaluation unit 14. The description of the same configuration as that of the first embodiment will be omitted.
　信頼性評価部１４は、予測モデルＭによって予測された測位誤差の確率分布の信頼性について評価する。信頼性評価部１４は、（１）モデルの不確かさに基づく信頼性の評価と、（２）学習時の全観測回数に基づく信頼性の評価の何れか又は両方を実行する。 The reliability evaluation unit 14 evaluates the reliability of the probability distribution of the positioning error predicted by the prediction model M. The reliability evaluation unit 14 executes either or both of (1) reliability evaluation based on the uncertainty of the model and (2) reliability evaluation based on the total number of observations during learning.
（１）モデルの不確かさに基づく信頼性の評価
　予測モデルＭがＲＮＮ等のネットワークの場合、信頼性評価部１４は、ネットワークの要素の一部を故意に欠落させるドロップアウトと呼ばれる手法を適用する。例えば、信頼性評価部１４は、所定回数だけ、その都度、異なるネットワークのドロップアウトを行い、同じ入力に対して、ドロップアウト適用済みの予測モデルＭが出力する予測値を記録する。信頼性評価部１４は、予測モデルＭが出力した予測値のばらつきを評価する。ばらつきが大きい場合、予測モデルＭによって予測された測位誤差の確率分布の信頼性は低く、ばらつきが小さければ、信頼性は高いと評価する。 (1) Evaluation of reliability based on model uncertainty When the prediction model M is a network such as RNN, the reliability evaluation unit 14 applies a method called dropout in which some of the elements of the network are intentionally omitted. .. For example, the reliability evaluation unit 14 drops out different networks each time a predetermined number of times, and records the predicted value output by the predicted model M to which the dropout has been applied for the same input. The reliability evaluation unit 14 evaluates the variation of the predicted value output by the prediction model M. When the variation is large, the reliability of the probability distribution of the positioning error predicted by the prediction model M is low, and when the variation is small, the reliability is evaluated as high.
（２）学習時の全観測回数に基づく信頼性の評価
　図３（ｃ）を用いて説明したように、学習時に、全観測回数が多い場合は予測値ｆｐｒｅｄが観測失敗率の分布Ｗと離れていると評価値Ｌが大きく低下する。反対に、図２（ｃ）を用いて説明したように、全観測回数が少ない場合には予測値ｆｐｒｅｄが観測失敗率の分布Ｗと離れていても、分布Ｗの広がりが大きいため、評価値Ｌの低下が小さい。この性質を利用して、学習時の観測回数に基づく信頼性の評価として式（２）のディリクレ分布を用いる。ただし、予測モデルＭの作成後は、予測値に対応したｎｋ（ｋ＝２，３，・・・，１１）が存在しないため、ｎｋ＝Ｎ×ｒｋとおいてディリクレ分布を算出する。信頼性評価部１４は、ディリクレ分布の広がりが大きい場合、予測モデルＭによって予測された測位誤差の確率分布の信頼性は低く（全観測回数が少ない）、広がりが小さければ（全観測回数が多い）、信頼性は高いと評価する。 (2) Evaluation of reliability based on the total number of observations during learning As explained using Fig. 3 (c), when the total number of observations is large during learning, the predicted value f pred is the distribution W of the observation failure rate. If they are far apart, the evaluation value L drops significantly. On the contrary, as explained with reference to FIG. 2C, when the total number of observations is small, even if the predicted value f pred is different from the distribution W of the observation failure rate, the distribution W spreads widely, so that the evaluation is made. The decrease in value L is small. Utilizing this property, the Dirichlet distribution of Eq. (2) is used as an evaluation of reliability based on the number of observations during learning. However, since the n k (k = 2, 3, ..., 11) corresponding to the predicted value does not exist after the prediction model M is created, the Dirichlet distribution is calculated by setting n k = N × r k. In the reliability evaluation unit 14, when the spread of the Diricle distribution is large, the reliability of the probability distribution of the positioning error predicted by the prediction model M is low (the total number of observations is small), and when the spread is small (the total number of observations is large). ), The reliability is evaluated as high.
　図６は、本開示の第二実施形態における予測モデルの評価方法について説明する図である。
　図６（ａ）～図６（ｃ）は、全観測回数が多い場合の評価処理を説明する図である。図６（ａ）は、特徴量空間を示す。特徴量とは予測モデルに入力するパラメータである。点Ｈ１は、特徴量ｇ１（例えば、地形の特徴量）の値がＧ１で、特徴量ｇ２（例えば、車両の速度）の値がＧ２である入力パラメータを示す。図６（ｂ）は、測位誤差分布空間を示す。点Ｉ１は、点Ｈ１の入力パラメータを、学習済みの予測モデルＭに入力した時の出力、つまり測位誤差の確率分布を示す。分布Ｊ１は、ドロップアウト適用時に発生したばらつき、または、全観測回数に応じたばらつき、または、その両方を示す。信頼性評価部１４は、学習済みの予測モデルＭにドロップアウトを適用して、点Ｈ１の入力パラメータを入力したときの予測値を、記憶部１３に記録する。信頼性評価部１４は、Ｉ１が示すｒ２，ｒ３，・・・，ｒ１１（それぞれ、測位誤差＝２、３、・・・、１１［ｍ］が観測される確率）と、予測モデル作成時の観測回数の合計値Ｎ（Ｎは既知である。）を用いて、ｎｋ＝Ｎ×ｒｋ（ｋ＝２~１１）を算出する。そして、信頼性評価部１４は、ｎｋ、ｒｋ、Ｎ、と式（２）により２～１１［ｍ］の各測位誤差がそれぞれｎｋ回（ｋ＝２~１１）ずつ観測されるときの測位誤差の事後確率の分布（ディリクレ分布）を算出する。信頼性評価部１４は、算出した事後確率の分布を、記憶部１３に記録する。分布Ｊ１は、ドロップアウト適用時に記録した予測値をプロットしたもの、算出した事後確率の分布、の何れか又は両方である。分布Ｊ１に基づいて、観測失敗率の分布Ｗ１を算出すると図６（ｃ）のような分布が得られる。 FIG. 6 is a diagram illustrating an evaluation method of a prediction model according to the second embodiment of the present disclosure.
6 (a) to 6 (c) are diagrams for explaining the evaluation process when the total number of observations is large. FIG. 6A shows a feature space. The feature quantity is a parameter to be input to the prediction model. Point H 1 is feature value g 1 (e.g., feature values of the terrain) indicated by the value of G 1, characteristic quantity g 2 (e.g., vehicle speed) of the input parameter value of is in G2. FIG. 6B shows a positioning error distribution space. The point I 1 indicates the output when the input parameter of the point H 1 is input to the trained prediction model M, that is, the probability distribution of the positioning error. Distribution J 1 indicates the variability that occurred when the dropout was applied, the variability according to the total number of observations, or both. Reliability evaluation unit 14 applies a drop out trained predictive model M, a predictive value when inputting the input parameters as at point H 1, is recorded in the storage unit 13. Reliability evaluation unit 14, I 1 is shown r2, r3, ..., r 11 and (respectively, the positioning error = 2,3, ..., 11 the probability of [m] is observed), create predictive models Using the total value N (N is known) of the number of observations at the time, n k = N × r k (k = 2 to 11) is calculated. The reliability evaluation unit 14, when n k, r k, N, and each of the positioning error of 2 ~ 11 [m] according to equation (2) is observed by each n k times (k = 2 ~ 11) Calculate the posterior probability distribution (Dirichlet distribution) of the positioning error of. The reliability evaluation unit 14 records the calculated posterior probability distribution in the storage unit 13. The distribution J 1 is either or both of a plot of the predicted values recorded when the dropout is applied and a calculated posterior probability distribution. When the distribution W 1 of the observation failure rate is calculated based on the distribution J 1 , the distribution as shown in FIG. 6 (c) is obtained.
　図６（ｄ）～図６（ｅ）は、全観測回数が少ない場合の評価処理を説明する図である。図６（ｄ）は特徴量空間を示し、図６（ｅ）は測位誤差分布空間を示す。点Ｉ２は、点Ｈ２の入力パラメータを、学習済みの予測モデルＭに入力した時の予測値である。予測値Ｉ２についての分布Ｊ２は、ドロップアウト適用時に発生したばらつき、または、全観測回数に応じたばらつき、または、その両方を示す。信頼性評価部１４は、図６（ｂ）で説明した処理と同様の処理を行って、図６（ｅ）の分布Ｊ２を算出する。分布Ｊ２に基づいて、観測失敗率の分布Ｗ２を算出すると図６（ｆ）のような分布が得られる。 6 (d) to 6 (e) are diagrams for explaining the evaluation process when the total number of observations is small. FIG. 6 (d) shows the feature space, and FIG. 6 (e) shows the positioning error distribution space. Point I 2 is a predicted value when the input parameter of point H 2 is input to the trained prediction model M. Distribution J 2 for the predicted value I 2 is the variation occurred in dropout application, or variations corresponding to all the observation times, or show both. Reliability evaluation unit 14 performs the same process as described in FIG. 6 (b), calculates a distribution J 2 in FIG. 6 (e). When the distribution W 2 of the observation failure rate is calculated based on the distribution J 2 , the distribution as shown in FIG. 6 (f) is obtained.
　信頼性評価部１４は、分布Ｊ１、分布Ｊ２の広がりの大きさに基づいて予測値の信頼性を評価する。例えば、信頼性評価部１４は、分布Ｊ１、分布Ｊ２の面積と所定の閾値を比較して、分布Ｊ１の面積が閾値未満であることに基づいて、予測モデルＭが予測した予測値Ｉ１は信頼性が高いと評価する。例えば、信頼性評価部１４は、分布Ｊ２の面積が閾値以上であることに基づいて、予測モデルＭが予測した予測値Ｉ２は信頼性が高いと評価する。 The reliability evaluation unit 14 evaluates the reliability of the predicted value based on the magnitude of the spread of the distribution J 1 and the distribution J 2. For example, the reliability evaluation unit 14 compares the area of the distribution J 1 and the distribution J 2 with a predetermined threshold, and the predicted value predicted by the prediction model M based on the fact that the area of the distribution J 1 is less than the threshold. I 1 is evaluated as highly reliable. For example, the reliability evaluation unit 14 evaluates that the predicted value I 2 predicted by the prediction model M is highly reliable based on the area of the distribution J 2 being equal to or larger than the threshold value.
　次に予測値の評価処理の流れについて説明する。
　図７は、本開示の第二実施形態における予測モデルの評価処理の一例を示すフローチャートである。
　まず、信頼性評価部１４は、モデルの不確かさに基づく信頼性の評価を行う（ステップＳ２１）。信頼性評価部１４は、ある予測値に関して、その予測値が得られたときの入力パラメータｎ値を、ドロップアウトを適用した予測モデルＭに入力し、予測値を取得する。信頼性評価部１４は、ドロップアウトするネットワークを変えて同様の処理を複数回行い、予測値のばらつきを算出する。
　次に信頼性評価部１４は、学習時の全観測回数に基づく信頼性の評価を行う（ステップＳ２２）。信頼性評価部１４は、評価対象の予測値ｆｐｒｅｄに関して、ｎｋ＝Ｎ×ｒｋを算出し、式（２）から事後確率の分布を算出する。
　次に信頼性評価部１４は、予測値を評価する（ステップＳ２３）。信頼性評価部１４は、ステップＳ２１、Ｓ２２で得られた分布の大きさに基づいて、予測値の信頼性を評価する。 Next, the flow of the evaluation process of the predicted value will be described.
FIG. 7 is a flowchart showing an example of the evaluation process of the prediction model according to the second embodiment of the present disclosure.
First, the reliability evaluation unit 14 evaluates the reliability based on the uncertainty of the model (step S21). With respect to a certain predicted value, the reliability evaluation unit 14 inputs the input parameter n value when the predicted value is obtained into the prediction model M to which the dropout is applied, and acquires the predicted value. The reliability evaluation unit 14 performs the same process a plurality of times by changing the network to drop out, and calculates the variation of the predicted value.
Next, the reliability evaluation unit 14 evaluates the reliability based on the total number of observations during learning (step S22). The reliability evaluation unit 14 calculates n k = N × r k with respect to the predicted value f pred to be evaluated, and calculates the distribution of posterior probabilities from the equation (2).
Next, the reliability evaluation unit 14 evaluates the predicted value (step S23). The reliability evaluation unit 14 evaluates the reliability of the predicted value based on the size of the distribution obtained in steps S21 and S22.
　本実施形態によれば、第一実施形態の効果に加え、学習済みの予測モデルＭによって予測された予測値の信頼性を評価することができる。例えば、予測値の信頼性が低い場合、図６（ｆ）に示すように、その予測値に基づいて算出される失敗率の分布Ｗ２は一定以上の広がりを持つ。このような場合、分布Ｗ２の広がりをみて、「この地点は、失敗率が不確定で道路の誤認識の可能性があるため、道路課金に関する処理を行わない方がよい」等の判断を行うことができる。 According to the present embodiment, in addition to the effect of the first embodiment, the reliability of the predicted value predicted by the trained prediction model M can be evaluated. For example, when the reliability of the predicted value is low, as shown in FIG. 6 (f), the distribution W 2 of the failure rate calculated based on the predicted value has a certain spread or more. In such a case, a look at the distribution W 2 spread, "this point, the failure rate is because there is a possibility of erroneous recognition of the road uncertain, it is better not to perform processing related to road pricing" a decision such as It can be carried out.
　図１０は、本開示の各実施形態における予測モデル作成装置のハードウェア構成の一例を示す図である。
　コンピュータ９００は、ＣＰＵ９０１、主記憶装置９０２、補助記憶装置９０３、入出力インタフェース９０４、通信インタフェース９０５を備える例えばＰＣ（Personal Computer）、サーバ端末装置などである。上述の予測モデル作成装置１０，１０Ａは、コンピュータ９００に実装される。そして、上述した各処理部の動作は、プログラムの形式で補助記憶装置９０３に記憶されている。ＣＰＵ９０１は、プログラムを補助記憶装置９０３から読み出して主記憶装置９０２に展開し、当該プログラムに従って上記処理を実行する。ＣＰＵ９０１は、プログラムに従って、記憶部１３に対応する記憶領域を主記憶装置９０２に確保する。ＣＰＵ９０１は、プログラムに従って、処理中のデータを記憶する記憶領域を補助記憶装置９０３に確保する。 FIG. 10 is a diagram showing an example of the hardware configuration of the prediction model creation device according to each embodiment of the present disclosure.
The computer 900 is, for example, a PC (Personal Computer) including a CPU 901, a main storage device 902, an auxiliary storage device 903, an input / output interface 904, and a communication interface 905, a server terminal device, and the like. The above-mentioned prediction model creation devices 10 and 10A are mounted on the computer 900. The operation of each processing unit described above is stored in the auxiliary storage device 903 in the form of a program. The CPU 901 reads the program from the auxiliary storage device 903, expands it to the main storage device 902, and executes the above processing according to the program. The CPU 901 secures a storage area corresponding to the storage unit 13 in the main storage device 902 according to the program. The CPU 901 secures a storage area for storing the data being processed in the auxiliary storage device 903 according to the program.
　少なくとも１つの実施形態において、補助記憶装置９０３は、一時的でない有形の媒体の一例である。一時的でない有形の媒体の他の例としては、入出力インタフェース９０４を介して接続される磁気ディスク、光磁気ディスク、ＣＤ－ＲＯＭ、ＤＶＤ－ＲＯＭ、半導体メモリ等が挙げられる。このプログラムが通信回線によってコンピュータ９００に配信される場合、配信を受けたコンピュータ９００が当該プログラムを主記憶装置９０２に展開し、上記処理を実行しても良い。当該プログラムは、前述した機能の一部を実現するためのものであっても良い。さらに、当該プログラムは、前述した機能を補助記憶装置９０３に既に記憶されている他のプログラムとの組み合わせで実現するもの、いわゆる差分ファイル（差分プログラム）であっても良い。 In at least one embodiment, the auxiliary storage device 903 is an example of a non-temporary tangible medium. Other examples of non-temporary tangible media include magnetic disks, magneto-optical disks, CD-ROMs, DVD-ROMs, semiconductor memories, etc. connected via the input / output interface 904. When this program is distributed to the computer 900 by a communication line, the distributed computer 900 may expand the program to the main storage device 902 and execute the above processing. The program may be for realizing a part of the above-mentioned functions. Further, the program may be a so-called difference file (difference program) that realizes the above-mentioned function in combination with another program already stored in the auxiliary storage device 903.
　その他、本発明の趣旨を逸脱しない範囲で、上記した実施の形態における構成要素を周知の構成要素に置き換えることは適宜可能である。この発明の技術範囲は上記の実施形態に限られるものではなく、本発明の趣旨を逸脱しない範囲において種々の変更を加えることが可能である。 In addition, it is possible to replace the components in the above-described embodiment with well-known components as appropriate without departing from the spirit of the present invention. The technical scope of the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present invention.
＜付記＞
　実施形態に記載の予測モデル作成装置１０，１０Ａ、予測モデル作成方法およびプログラムは、例えば以下のように把握される。 <Additional notes>
The prediction model creation devices 10, 10A, the prediction model creation method, and the program described in the embodiment are grasped as follows, for example.
（１）第１の態様に係る予測モデル作成装置１０，１０Ａは、衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成装置であって、ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するデータ取得部１１と、前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルＭを作成する予測モデル作成部１２と、前記予測モデルＭが出力する前記測位誤差の確率分布と、所定の関数（式（１））と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数（式（３））に基づいて算出される観測失敗率に基づいて算出する評価部１２２と、前記評価値が最適値となるように前記予測モデルＭを更新する更新部１２３と、を備える。
　これにより、正確な失敗率が算出可能な測位誤差を予測する予測モデルＭを作成することができる。所定の関数とは、失敗率を算出する関数を示し、式（１）や式（３）が所定の関数に該当する。 (1) The prediction model creation devices 10 and 10A according to the first aspect are prediction model creation devices for creating a prediction model for predicting a positioning error indicating an error in the position positioned by the satellite positioning system, and are at a certain point. Acquires topographical feature data indicating topographical features, the positioning error with respect to the position of the moving body observed by the satellite positioning system when the moving body passes through the point a plurality of times, and the number of observations for each positioning error. Data acquisition unit 11 to create a prediction model M that creates a prediction model M that outputs the probability distribution of the positioning error when the topographical feature data is input, and a prediction model creation unit 12 that outputs the probability distribution of the positioning error that the prediction model M outputs. With respect to the failure rate, which is calculated based on the predetermined function (Equation (1)) and indicates the magnitude of the possibility of erroneous recognition of the position, the evaluation value of the failure rate is set for each positioning error. An evaluation unit 122 that calculates based on the number of observations and the observation failure rate calculated based on the function (Equation (3)), and an update unit that updates the prediction model M so that the evaluation value becomes an optimum value. 123 and.
This makes it possible to create a prediction model M that predicts a positioning error for which an accurate failure rate can be calculated. The predetermined function indicates a function for calculating the failure rate, and the equations (1) and (3) correspond to the predetermined functions.
（２）第２の態様に係る予測モデル作成装置１０，１０Ａは、（１）の予測モデル作成装置１０，１０Ａであって、前記測位誤差ごとの観測回数が、前記データ取得部１１によって取得された前記観測回数となるための前記測位誤差の事後確率の分布と、前記関数（式（３））と、に基づいて、前記観測失敗率の分布Ｗを算出し、該分布Ｗを前記失敗率の予測値として出力する予測値算出部１２１、をさらに備える。
　これにより、失敗率の真の分布を算出することができる。 (2) The prediction model creation devices 10, 10A according to the second aspect are the prediction model creation devices 10, 10A of (1), and the number of observations for each positioning error is acquired by the data acquisition unit 11. Based on the distribution of posterior probabilities of the positioning error to be the number of observations and the function (Equation (3)), the distribution W of the observation failure rate is calculated, and the distribution W is used as the failure rate. The predicted value calculation unit 121, which outputs the predicted value of the above, is further provided.
This makes it possible to calculate the true distribution of failure rates.
（３）第３の態様に係る予測モデル作成装置１０，１０Ａは、（１）～（２）の予測モデル作成装置１０，１０Ａであって、前記評価部１２２は、前記予測モデルMに基づいて算出された前記失敗率を中心とする所定幅ｂの領域と、前記失敗率の予測値が示す分布Ｗが占める領域と、が重なる面積Ｌｗを前記評価値Ｌとして算出し、前記更新部１２３は、前記評価値Ｌが最大となるように前記予測モデルＭのパラメータを更新する。
　これにより、正確な失敗率を算出できる予測モデルＭを作成することができる。 (3) The prediction model creation devices 10, 10A according to the third aspect are the prediction model creation devices 10, 10A of (1) to (2), and the evaluation unit 122 is based on the prediction model M. The area Lw where the calculated region of the predetermined width b centered on the failure rate and the region occupied by the distribution W indicated by the predicted value of the failure rate overlap is calculated as the evaluation value L, and the update unit 123 , The parameters of the prediction model M are updated so that the evaluation value L becomes the maximum.
This makes it possible to create a prediction model M capable of calculating an accurate failure rate.
（４）第４の態様に係る予測モデル作成装置１０Ａは、（１）～（３）の予測モデル作成装置１０Ａであって、前記予測モデルが出力する前記測位誤差の確率分布の予測値に対する信頼性を評価する信頼性評価部１４、をさらに備える。 (4) The prediction model creation device 10A according to the fourth aspect is the prediction model creation device 10A of (1) to (3), and the reliability of the prediction value of the probability distribution of the positioning error output by the prediction model. A reliability evaluation unit 14 for evaluating sex is further provided.
（５）第５の態様に係る予測モデル作成装置１０Ａは、（４）の予測モデル作成装置１０Ａであって、前記予測モデルＭがニューラルネットワークで構築されている場合、前記信頼性評価部１４は、所定の前記地形特徴データおよび前記移動体の速度を前記予測モデルに入力した場合の、前記予測モデルへのドロップアウトの適用時のばらつきに基づいて、前記信頼性の評価を行う。 (5) The prediction model creation device 10A according to the fifth aspect is the prediction model creation device 10A of (4), and when the prediction model M is constructed by a neural network, the reliability evaluation unit 14 , The reliability is evaluated based on the variation at the time of applying the dropout to the prediction model when the predetermined topographical feature data and the speed of the moving object are input to the prediction model.
（６）第６の態様に係る予測モデル作成装置１０Ａは、（４）～（５）の予測モデル作成装置１０Ａであって、前記信頼性評価部１４は、前記観測回数の合計値の大きさに基づく前記予測値の事後確率の分布の広がりを算出し、前記分布の広がりに基づいて、前記信頼性の評価を行う。
　第４～第６の態様に係る予測モデル作成装置１０Ａによれば、予測モデルＭの作成後であっても、予測モデルＭが出力する予測値について、評価を行うことができる。 (6) The prediction model creation device 10A according to the sixth aspect is the prediction model creation device 10A of (4) to (5), and the reliability evaluation unit 14 is the magnitude of the total value of the number of observations. The spread of the posterior probability distribution of the predicted value based on the above is calculated, and the reliability is evaluated based on the spread of the distribution.
According to the prediction model creation device 10A according to the fourth to sixth aspects, it is possible to evaluate the prediction value output by the prediction model M even after the prediction model M is created.
（７）第７の態様に係る予測モデル作成方法は、衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成方法であって、ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するステップと、前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルを作成するステップと、を有し、前記予測モデルを作成するステップでは、前記予測モデルが出力する前記測位誤差の確率分布と、所定の関数と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数に基づいて算出される観測失敗率に基づいて算出し、前記評価値が最適値となるように前記予測モデルを更新する。 (7) The prediction model creation method according to the seventh aspect is a prediction model creation method for creating a prediction model for predicting a positioning error indicating an error in the position positioned by the satellite positioning system, and is a feature of the terrain at a certain point. The step of acquiring the topographical feature data indicating the above, the positioning error with respect to the position of the moving object observed by the satellite positioning system when the moving object passes through the point a plurality of times, and the number of observations for each positioning error. When the topographical feature data is input, there is a step of creating a prediction model that outputs the probability distribution of the positioning error, and in the step of creating the prediction model, the probability of the positioning error output by the prediction model. Regarding the failure rate, which is calculated based on the distribution and a predetermined function and indicates the magnitude of the possibility of misrecognition of the position, the evaluation value of the failure rate is the number of observations for each positioning error and the above. It is calculated based on the observation failure rate calculated based on the function, and the prediction model is updated so that the evaluation value becomes the optimum value.
（８）第８の態様に係るプログラムは、コンピュータに、衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成方法であって、ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するステップと、前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルを作成するステップと、を有し、前記予測モデルを作成するステップでは、前記予測モデルが出力する前記測位誤差の確率分布と、所定の関数と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数に基づいて算出される観測失敗率に基づいて算出し、前記評価値が最適値となるように前記予測モデルを更新する処理、を実行させる。 (8) The program according to the eighth aspect is a prediction model creation method for creating a prediction model for predicting a positioning error indicating an error of a position positioned by a satellite positioning system on a computer, and is a feature of the terrain at a certain point. The step of acquiring the topographical feature data indicating the above, the positioning error with respect to the position of the moving body observed by the satellite positioning system when the moving body passes through the point a plurality of times, and the number of observations for each positioning error. When the topographical feature data is input, there is a step of creating a prediction model that outputs the probability distribution of the positioning error, and in the step of creating the prediction model, the probability of the positioning error output by the prediction model. Regarding the failure rate, which is calculated based on the distribution and a predetermined function and indicates the magnitude of the possibility of misrecognition of the position, the evaluation value of the failure rate is the number of observations for each positioning error and the above. The process of updating the prediction model so that the evaluation value becomes the optimum value, which is calculated based on the observation failure rate calculated based on the function, is executed.
　上記した予測モデル作成装置、予測モデル作成方法及びプログラムによれば、衛星を利用した測位における測位された位置の誤差の分布を推定する予測モデルを作成することができる。 According to the above-mentioned prediction model creation device, prediction model creation method and program, it is possible to create a prediction model that estimates the distribution of the error of the positioned position in the positioning using satellites.
１０、１０Ａ　　　予測モデル作成装置
１１　　　データ取得部
１２　　　予測モデル作成部
１２１　　予測値算出部
１２２　　評価部
１２３　　更新部
１３　　　記憶部
１４　　　信頼性評価部
９００　　　コンピュータ
９０１　　　ＣＰＵ
９０２　　　主記憶装置
９０３　　　補助記憶装置
９０４　　　入出力インタフェース
９０５　　　通信インタフェース 10, 10A Prediction model creation device 11 Data acquisition unit 12 Prediction model creation unit 121 Prediction value calculation unit 122 Evaluation unit 123 Update unit 13 Storage unit 14 Reliability evaluation unit 900 Computer 901 CPU
902 Main storage device 903 Auxiliary storage device 904 Input / output interface 905 Communication interface

Claims (8)

  1. 　衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成装置であって、
    　ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するデータ取得部と、
    　前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルを作成する予測モデル作成部と、
    　前記予測モデルが出力する前記測位誤差の確率分布と、所定の関数と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数に基づいて算出される観測失敗率に基づいて算出する評価部と、
    　前記評価値が最適値となるように前記予測モデルを更新する更新部と、
    　を備える予測モデル作成装置。     It is a prediction model creation device that creates a prediction model that predicts the positioning error, which indicates the position error measured by the satellite positioning system.
    Topographical feature data showing the topographical features of a certain point, the positioning error with respect to the position of the moving object observed by the satellite positioning system when the moving object passes through the point a plurality of times, and the number of observations for each positioning error. , And the data acquisition unit to acquire
    A prediction model creation unit that creates a prediction model that outputs the probability distribution of the positioning error when the terrain feature data is input, and a prediction model creation unit.
    The evaluation value of the failure rate is calculated based on the probability distribution of the positioning error output by the prediction model and a predetermined function, which indicates the magnitude of the possibility of erroneous recognition of the position. With an evaluation unit that calculates based on the number of observations for each positioning error and the observation failure rate calculated based on the function.
    An update unit that updates the prediction model so that the evaluation value becomes the optimum value,
    Predictive model creation device.
  2. 　前記測位誤差ごとの観測回数が、前記データ取得部によって取得された前記観測回数となるための前記測位誤差の事後確率の分布と、前記関数と、に基づいて、前記観測失敗率の分布を算出し、該分布を前記失敗率の予測値として出力する予測値算出部、
    　をさらに備える請求項１に記載の予測モデル作成装置。     The distribution of the observation failure rate is calculated based on the posterior probability distribution of the positioning error and the function so that the number of observations for each positioning error becomes the number of observations acquired by the data acquisition unit. Then, the predicted value calculation unit that outputs the distribution as the predicted value of the failure rate,
    The predictive model creating apparatus according to claim 1.
  3. 　前記評価部は、前記予測モデルに基づいて算出された前記失敗率を中心とする所定幅の領域と、前記失敗率の予測値が示す分布が占める領域と、が重なる面積を前記評価値として算出し、
    　前記更新部は、前記評価値が最大となるように前記予測モデルのパラメータを更新する、
    　請求項２に記載の予測モデル作成装置。     The evaluation unit calculates as the evaluation value the area where the area having a predetermined width centered on the failure rate calculated based on the prediction model and the area occupied by the distribution indicated by the prediction value of the failure rate overlap. death,
    The update unit updates the parameters of the prediction model so that the evaluation value is maximized.
    The prediction model creation device according to claim 2.
  4. 　前記予測モデルが出力する前記測位誤差の確率分布の予測値に対する信頼性を評価する信頼性評価部、をさらに備える請求項１から請求項３の何れか１項に記載の予測モデル作成装置。 The prediction model creating device according to any one of claims 1 to 3, further comprising a reliability evaluation unit for evaluating the reliability of the probability distribution of the positioning error output by the prediction model with respect to the predicted value.
  5. 　前記予測モデルがニューラルネットワークで構築されている場合、前記信頼性評価部は、所定の前記地形特徴データおよび前記移動体の速度を前記予測モデルに入力した場合の、前記予測モデルへのドロップアウトの適用時のばらつきに基づいて、前記信頼性の評価を行う、
    　請求項４に記載の予測モデル作成装置。     When the prediction model is constructed by a neural network, the reliability evaluation unit drops out to the prediction model when the predetermined terrain feature data and the speed of the moving object are input to the prediction model. The reliability is evaluated based on the variation at the time of application.
    The prediction model creation device according to claim 4.
  6. 　前記信頼性評価部は、前記観測回数の合計値の大きさに基づく前記予測値の事後確率の分布の広がりを算出し、前記分布の広がりに基づいて、前記信頼性の評価を行う、
    　請求項４又は請求項５に記載の予測モデル作成装置。     The reliability evaluation unit calculates the spread of the posterior probability distribution of the predicted value based on the magnitude of the total number of observations, and evaluates the reliability based on the spread of the distribution.
    The prediction model creating apparatus according to claim 4 or 5.
  7. 　衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成方法であって、
    　ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するステップと、
    　前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルを作成するステップと、を有し、
    　前記予測モデルを作成するステップでは、
    　前記予測モデルが出力する前記測位誤差の確率分布と、所定の関数と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数に基づいて算出される観測失敗率に基づいて算出し、
    　前記評価値が最適値となるように前記予測モデルを更新する、
    　予測モデル作成方法。     It is a prediction model creation method that creates a prediction model that predicts the positioning error that indicates the position error that the satellite positioning system has positioned.
    Topographical feature data showing the topographical features of a certain point, the positioning error with respect to the position of the moving object observed by the satellite positioning system when the moving object passes through the point a plurality of times, and the number of observations for each positioning error. , And the steps to get
    When the terrain feature data is input, it has a step of creating a prediction model that outputs the probability distribution of the positioning error.
    In the step of creating the prediction model,
    The evaluation value of the failure rate is calculated based on the probability distribution of the positioning error output by the prediction model and a predetermined function, which indicates the magnitude of the possibility of erroneous recognition of the position. Is calculated based on the number of observations for each positioning error and the observation failure rate calculated based on the function.
    The prediction model is updated so that the evaluation value becomes the optimum value.
    How to create a predictive model.
  8. 　コンピュータに、
    　衛星測位システムが測位した位置の誤差を示す測位誤差を予測する予測モデルを作成する予測モデル作成方法であって、
    　ある地点の地形の特徴を示す地形特徴データと、前記地点を移動体で複数回通過したときに前記衛星測位システムが観測した前記移動体の位置に対する前記測位誤差および前記測位誤差ごとの観測回数と、を取得するステップと、
    　前記地形特徴データを入力すると、前記測位誤差の確率分布を出力する予測モデルを作成するステップと、を有し、
    　前記予測モデルを作成するステップでは、
    　前記予測モデルが出力する前記測位誤差の確率分布と、所定の関数と、に基づいて算出される、前記位置に対する誤認識が生じる可能性の大きさを示す失敗率について、前記失敗率の評価値を、前記測位誤差ごとの観測回数と前記関数に基づいて算出される観測失敗率に基づいて算出し、
    　前記評価値が最適値となるように前記予測モデルを更新する処理、
    　を実行させるプログラム。     On the computer
    It is a prediction model creation method that creates a prediction model that predicts the positioning error that indicates the position error that the satellite positioning system has positioned.
    Topographical feature data showing the topographical features of a certain point, the positioning error with respect to the position of the moving object observed by the satellite positioning system when the moving object passes through the point a plurality of times, and the number of observations for each positioning error. , And the steps to get
    When the terrain feature data is input, it has a step of creating a prediction model that outputs the probability distribution of the positioning error.
    In the step of creating the prediction model,
    The evaluation value of the failure rate is calculated based on the probability distribution of the positioning error output by the prediction model and a predetermined function, which indicates the magnitude of the possibility of erroneous recognition of the position. Is calculated based on the number of observations for each positioning error and the observation failure rate calculated based on the function.
    A process of updating the prediction model so that the evaluation value becomes an optimum value.
    A program that executes.
PCT/JP2020/011219 2020-03-13 2020-03-13 Prediction model creation device, prediction model creation method, and program WO2021181687A1 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/JP2020/011219 WO2021181687A1 (en) 2020-03-13 2020-03-13 Prediction model creation device, prediction model creation method, and program
JP2022505712A JP7235931B2 (en) 2020-03-13 2020-03-13 Prediction model creation device, prediction model creation method and program

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2020/011219 WO2021181687A1 (en) 2020-03-13 2020-03-13 Prediction model creation device, prediction model creation method, and program

Publications (1)

Publication Number Publication Date
WO2021181687A1 true WO2021181687A1 (en) 2021-09-16

Family

ID=77672172

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2020/011219 WO2021181687A1 (en) 2020-03-13 2020-03-13 Prediction model creation device, prediction model creation method, and program

Country Status (2)

Country Link
JP (1) JP7235931B2 (en)
WO (1) WO2021181687A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113834792A (en) * 2021-09-22 2021-12-24 中国科学院合肥物质科学研究院 MAX-DOAS-based 50-meter-resolution trace gas profile inversion method

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH0792388B2 (en) * 1989-04-17 1995-10-09 住友電気工業株式会社 Position detector
JPH0944244A (en) * 1995-07-27 1997-02-14 Shinko Electric Co Ltd Travel controller for unmanned vehicle
JP2003215228A (en) * 2002-01-23 2003-07-30 Hitachi Ltd Mobile terminal with position indication function and position indication method
JP2007203860A (en) * 2006-02-01 2007-08-16 Xanavi Informatics Corp Vehicle speed control device, and method and program for setting target speed in same device
US20120139775A1 (en) * 2010-12-02 2012-06-07 Honda Motor Co., Ltd. Method for Testing GNSS-Based Positioning Systems in Obstructed Environments

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH0792388B2 (en) * 1989-04-17 1995-10-09 住友電気工業株式会社 Position detector
JPH0944244A (en) * 1995-07-27 1997-02-14 Shinko Electric Co Ltd Travel controller for unmanned vehicle
JP2003215228A (en) * 2002-01-23 2003-07-30 Hitachi Ltd Mobile terminal with position indication function and position indication method
JP2007203860A (en) * 2006-02-01 2007-08-16 Xanavi Informatics Corp Vehicle speed control device, and method and program for setting target speed in same device
US20120139775A1 (en) * 2010-12-02 2012-06-07 Honda Motor Co., Ltd. Method for Testing GNSS-Based Positioning Systems in Obstructed Environments

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113834792A (en) * 2021-09-22 2021-12-24 中国科学院合肥物质科学研究院 MAX-DOAS-based 50-meter-resolution trace gas profile inversion method

Also Published As

Publication number Publication date
JPWO2021181687A1 (en) 2021-09-16
JP7235931B2 (en) 2023-03-08

Similar Documents

Publication Publication Date Title
US20170139423A1 (en) Control system and method for multi-vehicle systems
Scholte et al. A nonlinear set‐membership filter for on‐line applications
JP6525325B2 (en) Method and device for determining device location
JP2020535562A (en) Devices and methods to control the system
US20110302116A1 (en) Data processing device, data processing method, and program
JP5089281B2 (en) State estimation device and state estimation method
Ren et al. A fuzzy logic map matching for wheelchair navigation
CN112800616B (en) Equipment residual life self-adaptive prediction method based on proportional acceleration degradation modeling
EP2981136A1 (en) Terminal device, position estimation method, position estimation system, and program
US11615293B2 (en) Reinforcement learning with a stochastic action set
Mohamed et al. Population MCMC methods for history matching and uncertainty quantification
WO2021181687A1 (en) Prediction model creation device, prediction model creation method, and program
CN111256710A (en) Map matching method and system
Hakobyan et al. Distributionally robust risk map for learning-based motion planning and control: A semidefinite programming approach
Ghosh et al. A new simulation metric to determine safe environments and controllers for systems with unknown dynamics
Ngufor et al. Unsupervised labeling of data for supervised learning and its application to medical claims prediction
Louati Cloud-assisted collaborative estimation for next-generation automobile sensing
WO2021090413A1 (en) Control device, control system, control method, and program
CN110749325A (en) Flight path planning method and device
EP4266004A1 (en) User trajectory positioning method, electronic device and computer storage medium
US11475255B2 (en) Method for adaptive context length control for on-line edge learning
Král et al. Policy search for active fault diagnosis with partially observable state
KR102105032B1 (en) An efficient pareto set selection method for optimization of multi-objective systems
CN113075713A (en) Vehicle relative pose measuring method, system, equipment and storage medium
Mayer et al. Stochastic Optimal Control Using Gaussian Process Regression over Probability Distributions

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20923897

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2022505712

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20923897

Country of ref document: EP

Kind code of ref document: A1