JP2021142873A - Learning model generation method, probability determination method and determination device - Google Patents

Abstract

To provide a technique for predicting error event occurrence possibility at individual stations by use of operation information recorded and accumulated.SOLUTION: When station-specific operation data of a given station is inputted with data including station identification information 311A, which represents either a first class station at which error operation exceeding a fixed position to stop has not occurred during a prescribed period or a second class station at which the error operation has occurred, and station-specific operation data 312, which is statistic data of a plurality of information concerning operation performed when stopping at the station during the prescribed period, as teacher data 310A, a machine learning model 10A, which outputs a prediction value representing the possibility that the given station relevant to the first class station or the second class station, is created when the station-specific data of the given station has been inputted.SELECTED DRAWING: Figure 1

Description

The present invention relates to a probability determination method for determining the probability that an error operation will occur at a station to be stopped.

In railways, the installation of a driving information recording device (also called an operating status recording device) is obligatory by a ministerial ordinance (a ministerial ordinance that sets technical standards for railways). The operation information recording device is a device mounted on a railway vehicle and records various information related to the operation status of the train such as the position and speed of the train, the operation status of the driver, and the operation status of the main equipment (for example, patent documents). 1).

Japanese Unexamined Patent Publication No. 2008-221902

The driving information recorded by the driving information recording device is mainly used for analysis of the cause of an accident or the like. In addition, by systematically analyzing the causes of various error events including human error based on the recorded and accumulated driving information, it is also used to help prevent the recurrence of similar error events. .. Currently, the main focus is on analyzing the causes of past error events, but as a future task, new ways of utilizing recorded and accumulated driving information, such as predicting the possibility of future error events, are required. ing. For example, prediction of the possibility of error events caused by the driver's driving operation at each station.

The present invention has been made in view of the above circumstances, and an object of the present invention is to be able to generate an error event caused by a driver's driving operation at an individual station by using recorded and accumulated driving information. To provide a technique for predicting sex.

The first invention for solving the above problems is
It is performed when the first-class station and the second-class station where error operation exceeding the fixed position to be stopped did not occur during the predetermined period and stopped at the station during the predetermined period. Acquiring station-specific driving data, which is statistical data of a plurality of driving information (for example, teacher data acquisition unit 202 in FIGS. 5 and 9),
The station-specific operation data of a given station is obtained by using the station-specific operation data and the station identification information indicating whether the station-specific operation data is the information of the first category station or the information of the second category station as teacher data. To generate a machine learning model that outputs a predicted value indicating the possibility that the given station corresponds to the first classification station or the second classification station when input (for example, FIGS. 5 and 5). 9 Prediction model generation unit 204),
It is a learning model generation method including.

According to the first invention, it is possible to predict the possibility of error events occurring at individual stations by using the recorded and accumulated driving information. That is, given as teacher data, the operation data for each station of the first-class station where the error operation exceeding the fixed position of the station to be stopped as an error event does not occur and the second-class station where the error operation has occurred is used as the teacher data. By inputting the station-specific operation data of a station, it is possible to generate a machine learning model that outputs a predicted value indicating the possibility that the station corresponds to a first-class station or a second-class station. The predicted value output by the machine learning model is that the driving method performed when stopping at a given station is the driving performed at either the station where the error driving occurred or the station where the error driving did not occur. It shows how similar it is to the way it is done. Therefore, it is possible to predict the possibility of error operation occurring at a given station from the predicted value output by the machine learning model.

The second invention is the first invention.
The station-specific operation data includes the number of types of speed patterns until the station stops at a fixed position.
This is a learning model generation method.

According to the second invention, it is possible to generate a machine learning model that calculates a predicted value by including the number of types of speed patterns until the vehicle stops at a fixed position of a station as a factor of whether or not error driving occurs. ..

The third invention is the first or second invention.
The operation data for each station includes the usage rate of the maximum regular brake.
The machine learning model is a model for outputting the predicted value related to the error driving mainly due to the brake delay (for example, the predicted model 10A in the first embodiment).
This is a learning model generation method.

According to the third invention, it is possible to generate a machine learning model for calculating a predicted value by including the usage ratio of the maximum normal brake as a factor of whether or not error driving occurs.

The fourth invention is the third invention.
The station-specific driving data includes information on variations in traveling speed at a position a predetermined distance before the fixed position.
This is a learning model generation method.

According to the fourth invention, a machine learning model for calculating a predicted value is generated by including information on a variation in traveling speed at a position a predetermined distance before a fixed position as a factor of whether or not error driving occurs. Can be done. The "velocity variation" is, for example, the standard deviation of velocity.

The fifth invention is the first or second invention.
The acquisition further acquires the fixed position variable information regarding whether or not the fixed position of the station differs depending on the number of trains.
The generation generates the machine learning model by further including the fixed position variable information in the teacher data (for example, the prediction model 10B in the second embodiment).
This is a learning model generation method.

According to the fifth invention, a machine learning model that calculates a predicted value by including fixed position variable information regarding whether or not the fixed position of a station differs according to the number of trains in the factor of whether or not error driving occurs. Can be generated. The "fixed position variable information" includes, for example, the presence / absence of an n-car mark indicating a stop position target of a train having an "n" number of trains (a train of an n-car train), the number of stop position targets, and the like.

The sixth invention is
Based on the predicted value obtained by inputting station-specific operation data for a given station to the machine learning model generated by the learning model generation method of any one of the first to fifth inventions. Determining the probability that the error operation will occur at the station (for example, the probability determination unit 210 in FIGS. 5 and 9),
This is a probability determination method.

As another invention
It is performed when the first category station where the error operation exceeding the fixed position to be stopped does not occur during the predetermined period and the second category station where it occurs stops at the station during the predetermined period. Station-specific driving data, which is statistical data of a plurality of driving information, and station identification information indicating whether the station-specific driving data is the information of the first classification station or the information of the second classification station are used as teacher data. When the station-specific operation data for a given station is input, the trained trainer is to output a predicted value indicating the possibility that the given station corresponds to the first classification station or the second classification station. Prediction processing execution means (for example, prediction of FIGS. 5 and 9) for inputting station-specific operation data for a given station into a machine learning model and executing a prediction process for obtaining the predicted value of the given station. Part 208) and
With the probability determination means (for example, the probability determination unit 210 in FIGS. 5 and 9) for determining the probability that the error driving will occur at the input given station based on the prediction value obtained in the prediction process. ,
A determination device including the above may be configured.

According to the sixth invention and the like, based on the predicted value obtained by inputting the station-specific operation data of a given station into the machine learning model, the probability that an error operation will occur at the given station is determined. Can be determined. In other words, the predicted value output by the machine learning model is that the driving method performed when stopping at a given station is the driving performed at either the station where the error driving occurred or the station where the error driving did not occur. It shows how similar it is to. Therefore, for example, when an error operation has not yet occurred at a given station, but the predicted value output by the machine learning model indicates that it is highly likely that the station corresponds to the second category station where the error operation has occurred. Can determine that there is a high possibility that error operation will occur in the near future.

The explanatory view of the generation of the prediction model in 1st Embodiment. Explanatory drawing of generation of operation data for each station. Explanatory drawing for determination of probability that error driving occurs using a prediction model. An example of the judgment result. Functional configuration diagram of the judgment device. The explanatory view of the generation of the prediction model in 2nd Embodiment. Explanatory drawing for determination of probability that error driving occurs using a prediction model. An example of the judgment result. Functional configuration diagram of the judgment device.

Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings. It should be noted that the embodiments described below do not limit the present invention, and the embodiments to which the present invention can be applied are not limited to the following embodiments. Further, in the description of the drawings, the same elements are designated by the same reference numerals.

[Overview]
In the present embodiment, the driving information recorded in the driving information recording device mounted on the railway vehicle is used to determine the probability that an error driving exceeding a fixed position to be stopped at the station will occur. Error driving that exceeds a fixed position includes overrunning at a fixed position and passing through a station without stopping. The operation information is data in which the position and speed of the train, the operation status of the brake including the number of notches, the operation status, and the like are recorded in chronological order in association with the time. The operation information is the data recorded in the operation information recording device, but if it is the data of the same item, data other than the data recorded in the operation information recording device can be used as the operation information.

Specifically, it is a prediction model that is a machine learning model that shows the relationship between the driver's driving operation and error driving when stopping at a station based on the driving information when stopping at each of a plurality of stations in the past predetermined period. To generate. By using this prediction model, it is possible to predict the possibility of error driving at a given station in the near future from the driving information when the train stops at a given station. Hereinafter, two embodiments in which the factors of error operation are different will be described.

[First Embodiment]
The first embodiment is an embodiment that predicts the possibility of a braking time error in which the stop position is erroneous, mainly due to the driver's braking delay, as a driving error.

FIG. 1 is a diagram illustrating the generation of a prediction model. As shown in FIG. 1, the prediction model 10A of the first embodiment is generated by performing machine learning using a plurality of teacher data 310A. One teacher data 310A is data related to one station, and includes station identification information 311A, station-specific operation data 312, and station equipment data 313. The station identification information 311A is information indicating whether the station is a first-class station in which error driving has not occurred or a second-class station in which error driving has occurred during a predetermined period. The error driving in the first embodiment is a braking time error mainly caused by the driver's braking delay. The station-specific operation data 312 is data related to the operation by the driver when the train stops at the station, and is statistical data of a plurality of operation information when the train stops at the station during a predetermined period.

FIG. 2 is a diagram illustrating the generation of operation data 312 for each station. As shown in FIG. 2, the station-specific operation data 312 is generated by statistically calculating a plurality of operation information when the station stops at the station during a predetermined period. One driving information corresponds to one driving of the driver, and associates the speed, position, and brake notch of the train with the driving date and the driver's name at each time until the train stops at the station. This is the data. That is, the driving information is time-series data such as the position and speed of the train and the brake notch until the train stops at the target station by the driver. As a statistical calculation for a plurality of driving information, for example, the usage status of the brake notch, the average value of the speeds at a predetermined time point or a predetermined point, and the standard deviation are obtained to generate the operation data 312 for each station.

The operation data 312 by station shows the usage rate of the maximum normal brake for 5 seconds before the stop [%], the additional rate of 3 notches or more of the brake for 5 seconds before the stop [%], and the usage rate of the brake fully loosened for 5 seconds before the stop [ %], Average value of brake notch movement number for 5 seconds before stopping [times], standard deviation of brake notch movement number for 5 seconds before stopping [times], average value of additional brake amount for 5 seconds before stopping [notch], Standard deviation of the amount of additional braking for 5 seconds before stopping [notch], average value of speed 20m before the stop position [km / h], standard deviation of speed 20m before the stop position [km / h] , The average value of the speed at the point 200 m before the stop position [km / h], and the standard deviation of the speed at the point 200 m before the stop position [km / h].

Returning to FIG. 1, the station equipment data 313 included in the teacher data 310A is data on the equipment and structure of the station that may affect the driving operation of the driver when the train stops, and is a four-car marker. Presence / absence of 6-car mark, presence / absence of 8-car mark, presence / absence of 10-car mark, number of stop position targets at the platform [pieces], gradient of the installation location of the on-site signal [‰], rear end of the platform (entrance side) Includes 10 items: gradient [‰] in the center of the platform, gradient [‰] in the center of the platform, gradient [‰] at the tip of the platform (advanced side), and number of speed patterns [pieces].

The "n-car mark" is a sign indicating a stop position target of an n-car train. These n-car markers are provided at stations with different stop position targets depending on the number of trains. That is, the presence / absence of the 4-car mark, the presence / absence of the 6-car mark, the presence / absence of the 8-car mark, the presence / absence of the 10-car mark, and the number [pieces] of the stop position targets at the platform are the fixed positions of the stop position targets. It corresponds to the fixed position variable information regarding whether or not it differs depending on the number of trains.

The "number of speed patterns" is the number of types of rough speed patterns (graph of train speed with respect to position) that can be taken when stopping at the station. For example, a speed pattern from entering the station yard to stopping is created from each of a plurality of driving information when stopping at the station. Then, the created plurality of speed patterns are classified into groups having similar pattern shapes, and the number of classified groups is defined as the number of speed patterns. Since the shape of the speed pattern is mainly determined according to the signal display, the number of speed patterns is determined according to the number of traffic lights including the on-site signal and the signal display. For example, if the traffic light is only the on-site traffic light and the indications that the on-site traffic light can take are four types of blue only, blue + yellow, yellow only, and yellow + yellow, the number of speed patterns is "4".

The prediction model 10A is generated by performing machine learning with station-specific operation data 312 and station equipment data 313 as inputs and station identification information 311A as output among the data included in the teacher data 310A.

In the present embodiment, the type of machine learning is logistic regression analysis, but other machine learning such as a neural network may be used. That is, the prediction model 10A is generated by performing logistic regression analysis using the station-specific operation data 312 and the station equipment data 313 of the teacher data 310A as explanatory variables and the station identification information 311A as the objective variable. The explanatory variables are a total of 21 items, including 11 items of station-specific operation data 312 and 10 items of station equipment data 313. The objective variable is a binary variable in which the first classification station is "0" and the second classification station is "1". The prediction model formula that is the result of the logistic regression analysis is the model formula shown in formula (1).

In the model formula shown in the formula (1), out of the 21 explanatory variables, the usage rate [%] (explanatory variable X ₁ ) of the maximum regular brake for 5 seconds before the stop, and the speed at a point 200 m before the stop position. [Km / h] (explanatory variable X ₂ ), standard deviation of velocity [km / h] (explanatory variable X ₃ ) at a point 200 m before the stop position, number of velocity patterns [pieces] ( the explanatory variable X _4), four explanatory variables X ₁ to X ₄ of the shows that are affecting the objective variable. As a result of conducting the machine learning experiment of this embodiment based on the one-year operation information conducted at 92 stations, as an example, the coefficients a _{1 to} a _{4 of the explanatory variables X 1 to} X _{4 and} the coefficients a 1 to a 4 of the explanatory variables X 1 to X 4 and The constant term a ₀ was, for example, a ₁ = 3.69, a ₂ = 0.37, a ₃ = −0.72, a ₄ = 0.90, a ₀ = -22.79. The predicted value P calculated by the model formula of the formula (1) is a numerical value in the range of "0.0 to 1.0", and there is a possibility that the station is a second classification station, that is, an error operation occurs. Represents a probability.

FIG. 3 is a diagram for explaining the determination of the probability that error driving will occur at a given station using the prediction model 10A. As shown in FIG. 3, station data 330, which is data related to the station to be determined (“Z station” in the example of FIG. 3), is input to the prediction model 10A. The station data 330 includes station-specific operation data and station equipment data. _{Explanatory variable X 1} included in the operation data for each station, the usage rate [%] of the maximum regular brake for 5 seconds before the stop, and explanatory variable X ₂ the average value of the speed 200 m before the stop position [km / h], the explanatory variables X ₃ is a stop position velocity and standard deviation [km / h] of at 200m before the point, contained in the station facility data, the speed pattern number and an explanatory variable X _4, 4 pieces of the The values of the explanatory variables X _{1 to} X ₄ are input to the model formula represented by the formula (1) to calculate the predicted value P.

Since the predicted value P indicates the possibility of being a second-class station, the closer the predicted value P is to "1", the more likely it is to correspond to the second-class station, and conversely, the closer it is to "0", the more likely it is. It can be said that there is a high possibility that it corresponds to the first category station. Therefore, for example, if the threshold value of the predicted value P is "0.5" and the predicted value P is less than "0.5", the station to be determined corresponds to the first classification station, and the predicted value P is "0.5". If it is the above, it can be determined that the station to be determined corresponds to the second category station. Of course, instead of judging in two stages, it corresponds to the second classification station, which is likely / low to correspond to the first classification station depending on whether the predicted value P is close to "0" or "1". It may be judged in a total of 5 stages of high / low possibility and neither.

Then, if it is determined that the station corresponds to the second classification station, but no error driving has occurred during the predetermined target period, it is determined that the station to be determined is likely to have an error driving in the future. .. This is a station that corresponds to the second category station from the past driving information, that is, a station that is likely to cause error driving, but no error driving occurs during a predetermined period. In this case, it is highly likely that error operation will occur in the future. It should be noted that the closer the predicted value P is to "1", the higher the possibility that error driving will occur in the future, and the degree of the possibility may be predicted.

FIG. 4 is an example of a determination result (experimental result) by the determination method of the first embodiment. Using the driving information when the trains stopped at each of the 92 stations as teacher data 310A, a prediction model 10A represented by the model formula shown in the formula (1) was generated by logistic regression analysis. Of the 92 stations, there are 46 first-class stations and 46 second-class stations. Next, each of the 92 stations used as the teacher data 310A was set as the station to be determined, and the station data 330 was input to the generated prediction model 10A to calculate the predicted value P.

FIG. 4 is a diagram in which the predicted values P of each of the 92 stations calculated in this way are plotted. The plot of the predicted value P of the first classification station is shown on the left side, and the plot of the predicted value P of the second classification station is shown on the right side. The vertical axis is the predicted value P. Further, among the first-class stations, only one station in which the error operation first occurs after a predetermined period is shown by the plot of "white triangle".

According to FIG. 4, the predicted value P of less than 80% of the first-class stations is less than "0.5", and it is predicted that the stations correspond to the first-class stations. Further, among the first-class stations, the predicted value P of the station where the error driving first occurs after the predetermined period is "0.59", and it is predicted that the station corresponds to the second-class station. In addition, the predicted value P of less than 80% of the second classification station is "0.5" or more, and it is predicted that it corresponds to the second classification station.

As in this determination example, the station used to generate the prediction model 10A as the teacher data 310 is set as the station to be determined using the prediction model 10A. It is possible to extract stations where error driving is likely to occur mainly due to brake delay. That is, although it is the first classification station, there is a possibility that an error operation will occur in the future at a station where the predicted value predicted from the prediction model 10A is close to "1" (for example, the predicted value P is "0.5" or more). Can be extracted as a high station.

FIG. 5 is a functional configuration diagram of the determination device 1A. As described above, the determination device 1A uses the driving information when the train stops at the target station to generate a prediction model 10A showing the relationship between the driver's driving operation and the error driving when the train stops at the target station. Using this prediction model 10A, it is a device that determines the probability that an error operation will occur at a given station to be determined. As shown in FIG. 5, the determination device 1A includes an input unit 102, a display unit 104, a sound output unit 106, a communication unit 108, a processing unit 200, and a storage unit 300, and is a kind of determination device 1A. Realized as a computer system. The determination device 1A may be realized by one computer, or may be configured by connecting a plurality of computers.

The input unit 102 is realized by an input device such as a keyboard, a mouse, a touch panel, and various switches, and outputs an operation signal corresponding to the performed operation to the processing unit 200. The display unit 104 is realized by a display device such as a liquid crystal display or a touch panel, and performs various displays based on the display signal from the processing unit 200. The sound output unit 106 is realized by a sound output device such as a speaker, and outputs various sounds based on the sound signal from the processing unit 200. The communication unit 108 is a communication device realized by, for example, a wireless communication module, a router, a modem, a jack or a control circuit of a wired communication cable, and is connected to a given communication network to perform data communication with an external device. conduct.

The processing unit 200 is a processor realized by an arithmetic unit such as a CPU (Central Processing Unit) or an FPGA (Field Programmable Gate Array) or an arithmetic circuit, and is a program or data stored in the storage unit 300, an input unit 102, or a communication unit. Overall control of the determination device 1A is performed based on the input data from the unit 108 and the like. Further, the processing unit 200 has a teacher data acquisition unit 202, a prediction model generation unit 204, a station data acquisition unit 206, a prediction unit 208, and a probability determination unit 210 as functional processing blocks. Each of these functional units included in the processing unit 200 can be realized by software by the processing unit 200 executing a program, or can be realized by a dedicated arithmetic circuit. In the present embodiment, the former will be described as being realized by software.

The teacher data acquisition unit 202 describes the first-class station in which the error operation exceeding the fixed position of the station to be stopped has not occurred during the predetermined period and the second-class station in which the error operation has occurred during the predetermined period. The teacher data 310A including the station-specific driving data 312, which is statistical data of a plurality of driving information performed when the train stops at the station, is acquired. The teacher data 310A includes station identification information 311A and station equipment data 313, in addition to station-specific operation data 312. The error driving indicated by the station identification information 311A is a braking time error in which the stop position is erroneous mainly due to the driver's braking delay. The operation data 312 by station includes the usage ratio of the maximum regular brake for 5 seconds before the stop (explanatory variable X ₁ ), the average value of the speed 200 m before the stop position (explanatory variable X ₂ ), and the predetermined distance from the fixed position. _{The standard deviation of the speed (explanatory variable X 3} ) at the point 200 m before the stop position, which is information on the variation of the speed in front, is included, and the station equipment data 313 stops at the stop position target which is the fixed position of the station. The number of types of speed patterns (explanatory variable X ₄ ) is included (see FIG. 1).

The prediction model generation unit 204 generates station-specific operation data, station equipment data, and station identification information indicating whether the station-specific operation data and station equipment data are information on a first-class station or information on a second-class station. A machine learning model that outputs predicted values indicating the possibility that a given station corresponds to a first-class station or a second-class station when the station-specific operation data of a given station is input as teacher data. To generate.

That is, the prediction model generation unit 204 performs machine learning using the teacher data 310A acquired by the teacher data acquisition unit 202, and station data including station-specific operation data and station equipment data of a given determination target station. When 330 is input, a prediction model 10A, which is a machine learning model that outputs a prediction value P indicating the possibility that the station to be determined corresponds to the first classification station or the second classification station, is generated (FIG. 1). reference). When logistic regression analysis is performed as machine learning, a model formula showing an example in formula (1) is generated as the prediction model 10A. The data related to the prediction model 10A generated by the prediction model generation unit 204 is stored as the prediction model data 320. The prediction model data 320 is data that defines the prediction model 10A, and specifically, is data of coefficients a _{0 to} a ₄ of the model formula showing an example in the formula (1) (see FIG. 1).

The station data acquisition unit 206 acquires station data 330 including station-specific operation data and station equipment data related to a given station (see FIG. 3).

The prediction unit 208 inputs station-specific operation data and station equipment data related to a given station into the machine learning model, and executes a prediction process for obtaining a predicted value of the given station. That is, the prediction unit 208 inputs the station data 330 acquired by the station data acquisition unit 206 into the model formula (1) generated as the prediction model 10A by the prediction model generation unit 204, so that the station to be determined can be determined. The predicted value P indicating the possibility of corresponding to the first classification station or the second classification station is calculated.

The probability determination unit 210 determines the probability that an error driving will occur at a given station based on the predicted value obtained in the prediction process. That is, in the probability determination unit 210, the predicted value P calculated by the prediction unit 208 is equal to or higher than the threshold value (for example, “0.5”), and an error operation occurs at the station to be determined during a predetermined period. If not, it is determined that there is a high possibility that an error operation will occur at the station to be determined in the near future.

The storage unit 300 is realized by a storage device such as an IC (Integrated Circuit) memory such as a ROM (Read Only Memory) or a RAM (Random Access Memory) or a hard disk, and the processing unit 200 controls the determination device 1A in an integrated manner. Is stored as a work area of the processing unit 200, and the calculation result executed by the processing unit 200 and the input data from the input unit 102 and the communication unit 108 are temporarily stored. NS. In the present embodiment, the storage unit 300 contains the determination program 302, the teacher data 310A acquired by the teacher data acquisition unit 202, the prediction model data 320A generated by the prediction model generation unit 204, and the station data acquisition unit 206. The station data 330 acquired by the above and the determination result data 340A determined by the probability determination unit 210 are stored.

[Second Embodiment]
In the second embodiment, as a driving error, it is predicted that there is a possibility that a stop position error may occur in which the stop position is erroneous mainly due to the driver's misunderstanding of the stop position. In the second embodiment below, the points different from those of the first embodiment will be mainly described. Further, the same components as those in the above-described first embodiment are designated by the same reference numerals, and detailed description thereof will be omitted.

FIG. 6 is a diagram illustrating the generation of the prediction model. As shown in FIG. 6, the prediction model 10B of the second embodiment is generated by performing machine learning using a plurality of teacher data 310B as in the first embodiment. The error driving indicated by the station identification information 311B included in the teacher data 310B is a stop position error in which the stop position is erroneous mainly due to the driver's misunderstanding of the stop position, and is different from the brake time error in the first embodiment.

The prediction model 10B is generated by performing machine learning with station-specific operation data 312 and station equipment data 313 as inputs and station identification information 311B as output among the data included in the teacher data 310B. That is, the prediction model 10B is generated by performing logistic regression analysis using the station-specific operation data 312 and the station equipment data 313 of the teacher data 310 as explanatory variables and the station identification information 311B as the objective variable. The prediction model formula that is the result of the logistic regression analysis is the model formula shown in formula (2).

Model formula shown in Equation (2), among the 21 pieces of explanatory variables, the average value of the velocity at 20m before the point of the stop position [km / h] (as explanatory variables X _5), whether the four-car mark ( the explanatory variable _{X 6),} the presence or absence of 6-Car mark (the explanatory variable _{X 7),} the speed pattern number [pieces] (explanatory variable _{X 8),} 4 pieces of explanatory variables _X 5 to X ₈ of, It shows that it affects the objective variable. As a result of conducting the machine learning experiment of this embodiment based on the one-year operation information conducted at 72 stations, as an example, the coefficients a _{5 to} a _{8 of the explanatory variables X 5 to} X _{8 and} the coefficients a 5 to a 8 of the explanatory variables X 5 to X 8 and The constant term a ₉ was, for example, a ₅ = 0.80, a ₆ = 1.42, a ₇ = 1.15, a ₈ = 1.07, a ₉ = -18.30. The predicted value P calculated by the model formula of the formula (2) is a numerical value in the range of "0.0 to 1.0", and there is a possibility that the station is a second classification station, that is, an error operation occurs. Represents a probability. The second embodiment predicts the possibility of error driving having different main causes, and since the station identification information 311B included in the teacher data 310B is different from the first embodiment, it is different from the prediction model 10A of the first embodiment. Generates a different prediction model 10B.

FIG. 7 is a diagram for explaining the determination of the probability that error driving will occur at a given station using the prediction model 10B. As shown in FIG. 7, station data 330, which is data related to the station to be determined (“Z station” in the example of FIG. 7), is input to the prediction model 10B. That is, the average value [km / h] of the speed at the point 20 m before the stop position, which is the _{explanatory variable X 5} included in the station data 330 for each station _{, and the explanatory variable X 6 included in the station equipment data.} and the presence of four-car mark is, the presence of six-car mark an explanatory variable X _7, speed pattern number and an explanatory variable X _8, the value of four explanatory variables X ₅ to X ₈ in the formula ( The predicted value P is calculated by inputting into the model formula represented by 2).

Then, as in the first embodiment, for example, if the threshold value of the predicted value P is "0.5" and the predicted value P is less than "0.5", the station to be determined corresponds to the first classification station. If the predicted value P is "0.5" or more, it is determined that the station to be determined corresponds to the second category station.

FIG. 8 is an example of a determination result (experimental result) by the determination method of the second embodiment. Using the driving information when the trains stopped at each of the 72 stations as the teacher data 310B, the prediction model 10B represented by the model formula shown in the formula (2) was generated by logistic regression analysis. Of the 72 stations, there are 36 first-class stations and 36 second-class stations. Next, each of the 72 stations used as the teacher data 310A was set as the station to be determined, and the station data 330 was input to the generated prediction model 10B to calculate the predicted value P.

FIG. 8 is a diagram in which the predicted values P of each of the 72 stations calculated in this way are plotted. The plot of the predicted value P of the first classification station is shown on the left side, and the plot of the predicted value P of the second classification station is shown on the right side. The vertical axis is the predicted value P. Further, among the first-class stations, only one station in which the error operation first occurs after a predetermined period is shown by the plot of "white triangle".

According to FIG. 8, the predicted value P of less than 80% of the first-class stations is less than “0.5”, and it is predicted that the stations correspond to the first-class stations. Further, among the first-class stations, the predicted value P of the station where the error driving first occurs after the predetermined period is "0.71", and it is predicted that the station corresponds to the second-class station. In addition, the predicted value P of less than 80% of the second classification station is "0.5" or more, and it is predicted that it corresponds to the second classification station.

As in this determination example, the station used to generate the prediction model 10B as the teacher data 310 is set as the station to be determined using the prediction model 10B. It is possible to extract stations that are likely to cause error driving mainly due to misunderstanding of the stop position target. That is, although it is the first classification station, there is a possibility that an error operation will occur in the future at a station where the predicted value predicted from the prediction model 10B is close to "1" (for example, the predicted value P is "0.5" or more). Can be extracted as a high station.

FIG. 9 is a functional configuration diagram of the determination device 1B. As described above, the determination device 1B uses the driving information when the train stops at the target station to generate a prediction model 10B showing the relationship between the driver's driving operation and the error driving when the train stops at the target station. Using this prediction model 10B, it is a device that determines the probability that an error operation will occur at a given station to be determined.

In the determination device 1B, the teacher data acquisition unit 202 acquires the teacher data 310B. The error driving indicated by the station identification information 311B included in the teacher data 310B is a stop position error in which the stop position is erroneous mainly due to the driver's misunderstanding of the stop position.

The prediction model generation unit 204 performs machine learning using the teacher data 310B acquired by the teacher data acquisition unit 202 to generate the prediction model 10B (see FIG. 6). When logistic regression analysis is performed as machine learning, a model formula showing an example in formula (2) is generated as the prediction model 10B. The data related to the prediction model 10B generated by the prediction model generation unit 204 is stored as the prediction model data 320B. Predictive model data 320B is data defining the predictive model 10B, specifically, the data of the coefficient _a 5 ~a ₉ model type, one example of which is shown in Equation (2).

The prediction unit 208 inputs the station data 330 acquired by the station data acquisition unit 206 into the model formula (2) generated as the prediction model 10B by the prediction model generation unit 204, so that the station to be determined is the first station. The predicted value P indicating the possibility of corresponding to the classification station or the second classification station is calculated.

[Action effect]
As described above, according to the first embodiment and the second embodiment, it is possible to predict the possibility of future error events occurring at individual stations by using the recorded and accumulated operation information. That is, given as teacher data, the operation data for each station of the first category station where the error operation exceeding the fixed position of the station to be stopped as an error event has not occurred and the second category station where the error operation has occurred is given. By inputting the station-specific operation data of a station, it is possible to generate a machine learning model that outputs a predicted value indicating the possibility that the station corresponds to a first-class station or a second-class station. The predicted value output by the machine learning model is that the driving method performed when stopping at a given station is the driving performed at either the station where the error driving occurred or the station where the error driving did not occur. It shows how similar it is to the way it is done. Therefore, it is possible to predict the possibility of error operation occurring at a given station from the predicted value output by the machine learning model.

It should be noted that the applicable embodiments of the present invention are not limited to the above-described embodiments, and of course, they can be appropriately changed without departing from the spirit of the present invention.

1A, 1B ... Judgment device 200 ... Processing unit 202 ... Teacher data acquisition unit 204 ... Prediction model generation unit 206 ... Station data acquisition unit 208 ... Prediction unit 210 ... Probability judgment unit 300A, 300B ... Storage unit 302 ... Judgment program 310A, 310B ... Teacher data 320A, 320B ... Prediction model data 330 ... Station data 340A, 340B ... Judgment result data 10A, 10B ... Prediction model

Claims (7)

It is performed when the first-class station and the second-class station where error operation exceeding the fixed position to be stopped did not occur during the predetermined period and stopped at the station during the predetermined period. Acquiring station-specific driving data, which is statistical data of multiple driving information,
The station-specific operation data of a given station is obtained by using the station-specific operation data and the station identification information indicating whether the station-specific operation data is the information of the first classification station or the information of the second classification station as teacher data. To generate a machine learning model that outputs a predicted value indicating the possibility that the given station corresponds to the first classification station or the second classification station when input.
Learning model generation method including. The station-specific operation data includes the number of types of speed patterns until the station stops at a fixed position.
The learning model generation method according to claim 1. The operation data for each station includes the usage rate of the maximum regular brake.
The machine learning model is a model for outputting the predicted value related to the error driving mainly due to the brake delay.
The learning model generation method according to claim 1 or 2. The station-specific driving data includes information on variations in traveling speed at a position a predetermined distance before the fixed position.
The learning model generation method according to claim 3. The acquisition further acquires the fixed position variable information regarding whether or not the fixed position of the station differs depending on the number of trains.
The generation generates the machine learning model by further including the fixed position variable information in the teacher data.
The learning model generation method according to claim 1 or 2. Based on the predicted value obtained by inputting station-specific operation data for a given station to the machine learning model generated by the learning model generation method according to any one of claims 1 to 5. To determine the probability that the error driving will occur at the station.
Probability determination method. It is performed when the first category station where the error operation exceeding the fixed position to be stopped does not occur during the predetermined period and the second category station where it occurs stops at the station during the predetermined period. Station-specific driving data, which is statistical data of a plurality of driving information, and station identification information indicating whether the station-specific driving data is the information of the first classification station or the information of the second classification station are used as teacher data. When the station-specific operation data for a given station is input, the trained trainer is to output a predicted value indicating the possibility that the given station corresponds to the first classification station or the second classification station. A prediction processing execution means for inputting station-specific operation data for a given station into a machine learning model and executing a prediction process for obtaining the predicted value of the given station.
Based on the predicted value obtained in the prediction process, a probability determining means for determining the probability that the error driving will occur at the input given station, and a probability determining means.
Judgment device including.

Images