JP7371040B2 - Operation status estimation system and operation status estimation method - Google Patents

Description

The present invention relates to a technique for estimating delays of moving objects such as buses, trains, streetcars, ships, and taxis.

Mobile vehicles such as buses, trains, streetcars, ships, and some taxis operate according to timetables. Although it is desirable that these vehicles operate according to timetables, delays may occur due to various reasons such as accidents, construction, weather, and traffic jams. In the case of constant delays, it may be possible to respond by changing the timetable, but it is not possible to depart earlier than the normal timetable (as an example, this is regulated within Japan). , it is not possible to appropriately determine a timetable based on the average travel time. Therefore, for these reasons, even if the timetable is updated, there are situations where permanent delays still occur. Furthermore, in order to make it easier for users to understand, timetables are often set such that flights on the same route run at the same time (for example, with an hour difference) even if their departure times are different.

Nowadays, with the development of information and communication technology, it has become possible to use location information of mobile objects that changes from moment to moment. In addition, studies are progressing both domestically and internationally on data formats that can be commonly used by carriers, such as GTFS (General Transit Feed Specification) and GTFS-RT. Against this background, there is a need for a common delay prediction technique to improve user convenience.

Here, Patent Document 1 describes a calculation unit that calculates a plurality of explanatory variables using at least one of the arrival time and departure time of each train at each station in a predetermined section of the railway; The delay time at a predetermined station and a predetermined train in a section is set as an objective variable, and an explanatory variable having a predetermined contribution to the objective variable is selected from among the plurality of explanatory variables, and a regression equation for the objective variable is created. and a display control unit that causes a display unit to display information regarding the delay time based on the regression equation.

JP 2019-123479 Publication

Nowadays, GPS devices are often attached to moving objects to identify the location, but the technology described in Patent Document 1 is based on the arrival time and departure time of each railway station. Therefore, there is a problem in that the delay cannot be estimated by reflecting real-time information on the moving object obtained by GPS (Global Positioning System) or the like.

Furthermore, for buses, streetcars, etc., there may be no means to obtain accurate arrival and departure times at stops (destinations, such as bus stops). Furthermore, if there are no users, they may simply pass through the station, making it difficult to obtain arrival and departure times. Therefore, information corresponding to the arrival time and departure time for each station, which is the basic data of the technology described in Patent Document 1, cannot be obtained, and the technology described in Patent Document 1 cannot be utilized.

Therefore, an object of the present invention is to provide an operation state estimation system and an operation state estimation method that can estimate the current delay time with finer granularity.

According to a first aspect of the present invention, the following driving state estimation system is provided. That is, the operating state estimation system includes a processor. The processor (1) executes the passage time estimation model construction unit and uses at least the route information, the timetable, and the position information of the mobile body based on the travel history, or calculates the position of the stop of the mobile body. Input information on the current location of the mobile object in units that have commonality in the operation pattern, using at least information on the timetable at the stop and the information that can be obtained by linearly interpolating the above information. A transit time estimation model is constructed to estimate the elapsed time from the departure of the mobile object according to the following. The processor (2) executes the inference unit to determine the current delay time of the mobile object from the elapsed time output from the passing time estimation model. The processor (3) executes the disclosure section to present to the user the operating state of the mobile body determined by the execution of the inference section.

According to the second aspect of the present invention, the following driving state estimation method is provided. That is, the operation state estimation method uses at least (1) route information, a timetable, and position information of a mobile body based on a travel history, or the position of a stop of a mobile body and the time at the stop. Using at least the information in the table and the information that can be obtained by linearly interpolating the information, the movement is performed in accordance with the input of information on the current position of the mobile object in units that have commonality in the operation pattern. Build a transit time estimation model that estimates the elapsed time from the departure of the moving object, (2) calculate the current delay time of the moving object from the elapsed time output from the transit time estimation model, and (3) calculate the determined movement. The operating state of the body is presented to the user.

According to the present invention, by using a passing time estimation model that estimates the elapsed time from departure for information specifying the position of an arbitrary moving object, it becomes possible to estimate the current delay time with fine granularity.

Furthermore, if information to identify the location of a moving object in operation is obtained, it is possible to predict the delay time of the moving object even if accurate arrival and departure times cannot be obtained. . For example, buses and streetcars may pass through without passengers, and even if it is difficult to obtain accurate arrival and departure times at the stop, the delay time of the moving object is Can be predicted.

FIG. 2 is a diagram for explaining the system configuration and functional configuration in this embodiment. FIG. 1 is a diagram for explaining the hardware configuration of a system in this embodiment. The figure which shows an example of stool data. The figure which shows an example of timetable data. The figure which shows an example of route data. The figure which shows an example of real-time position data. The figure which shows an example of passage time estimation model data. The figure which shows an example of estimated passage time and delay time data. The figure which shows an example of delay time data for each route number. The figure which shows an example of delay time prediction model training data. The figure which shows an example of delay time prediction model data. The figure which shows an example of predicted delay time data by route order. The figure which shows an example of reliability data. Flowchart for explaining model construction processing. 5 is a flowchart for explaining details of a learning process of a passing time estimation model. Flowchart for explaining standardization of estimated delay time sequences. A diagram for explaining a delay time prediction model. Flowchart for explaining inference processing. The figure which shows an example of an operation status display screen.

Hereinafter, typical embodiments for carrying out the present invention will be described with reference to the drawings as appropriate. This embodiment will be described as an example in which a bus is used as a moving object. However, the same applies to other moving bodies such as trains, streetcars, ships, and taxis. First, I will give an overview.

The processing of the driving state estimation system in this embodiment is divided into a construction phase and an inference phase. The construction phase consists of static data such as flights, timetables, routes, etc., dynamic data obtained from vehicles and environmental equipment during actual operation such as location, departure/arrival information, congestion status in vehicles, and weather information. and surrounding information such as traffic information, and performs processing to construct a passing time estimation model and delay time prediction model used for estimating operating conditions. The inference phase uses the above-mentioned static data, dynamic data of the vehicle in motion, and surrounding information to calculate the estimated value of the current or future delay time and information regarding the reliability of the estimation result. and performs the process of presenting the results to the user.

In the construction phase, the static data and dynamic data described above are first collected. Dynamic data is collected for one month, for example, at the time of initial construction. However, this period can be freely selected. Next, using static data and dynamic data, a transit time estimation model is constructed that estimates the elapsed time (sometimes referred to as transit time) from the first train to a certain location on the route as 0. Ru. Here, if the same location is passed twice or more on the route, information that can identify the number of visits is taken into consideration. In order to simplify the explanation, it is assumed that the above-mentioned information is taken into consideration even when the term "position" is simply written hereinafter.

Intuitively, the passing time estimation model is like a timetable for an arbitrary location, and by using this model and real-time location information, more fine-grained state estimation becomes possible.

Although it is possible to calculate the estimated delay time at each time using dynamic data and a transit time estimation model (in other words, it is possible to calculate the estimated delay time at the time when position information is acquired), If the estimated delay time at each time for each flight is called an estimated delay time sequence, the length of the estimated delay time sequence will differ depending on the day because the delay of each operated flight differs from day to day. Therefore, standardization (fixed length) using positions (predetermined positions) on a typical route is performed. Then, using information such as the standardized estimated delay time sequence, date and time, and the aforementioned peripheral information, the delay time (referred to as predicted delay time) is calculated from the partially given estimated delay time sequence and other information. A delay time prediction model is constructed.

In the inference phase, dynamic data about the moving vehicle is first collected. Then, the estimated delay time for the running vehicle is calculated using the collected dynamic data, static data, and passing time estimation model. Next, a predicted delay time is calculated using the estimated delay time, peripheral information, and delay time prediction model. Furthermore, a reliability level is calculated based on the reconstruction error and the adjustment (mask) of the input used for prediction. This information is then presented to the user as the operating status.

Next, the configuration of the driving state estimation system in this embodiment will be described with reference to FIG. 1. The operation status estimation system includes an operation status estimation server 10 that estimates the operation status, a vehicle 11 that transports passengers, an environmental equipment 12 that measures the status of the vehicle with a sensor such as a camera, a traffic information service 13, and weather information. It includes a service 14 and a user terminal 15 such as a smartphone or a personal computer (PC). Note that the user 16 who operates the user terminal 15 may or may not be a passenger of the vehicle 11 .

The components of the driving state estimation system are interconnected via WWW (World Wide Web). Note that in this embodiment, the connection is made via WWW, but other communication means may be used. Further, the above-mentioned constituent elements are merely examples, and the number of elements may be increased or decreased. For example, the operation state estimation server 10 may be composed of two or more servers. Moreover, it is sufficient if the above-mentioned dynamic data can be collected, and the dynamic data may be acquired from either the vehicle 11 or the environmental equipment 12, or from both. Furthermore, when acquiring dynamic data from the vehicle 11, the environmental equipment 12 may be omitted.

Next, the correspondence between functions and hardware will be explained with reference to FIGS. 1 and 2. A CPU (Central Processing Unit) 1H101 loads a program stored in a ROM (Read Only Memory) 1H102 or an external storage device 1H104 into a RAM (Random Access Memory) 1H103, and uses a communication I/F (Interfacial ce) 1H105, mouse, keyboard, etc. By controlling an external input device 1H106 represented by 1H106 and an external output device 1H107 represented by a display, etc., the collection unit 101, passage time estimation model construction unit 102, and delay The time prediction model construction unit 103, the inference unit 104, the publishing unit 105, and the data management unit 106 are realized (function).

Next, the structure of each data will be explained with reference to the figures. First, with reference to FIG. 3, the flight data 1D1 collected by the traffic status estimation server 10 from the traffic information service 13 and managed by the data management unit 106 will be described. Flight data 1D1 is data representing basic information of the operated flight, and includes flight ID (1D101), flight name (1D102), timetable ID (1D103), group ID (1D104), and route ID ( 1D105). The flight ID (1D101) is an identifier for identifying the flight. The flight name (1D102) is a name corresponding to the flight ID (1D101). The timetable ID (1D103) is an identifier for identifying a timetable. The group ID (1D104) is an identifier to ensure that the routes of two flights are the same and that the elapsed time in the timetable (the elapsed time when the first departure time is set to time 0) is also the same. be. In other words, it is an identifier for ensuring that there is a commonality in the movement patterns of mobile objects. Route ID (1D105) is an identifier for specifying a route.

A detailed explanation of the data shown in FIG. 3 will be supplemented. "XY-LINE-OUT-W600" represents an outbound flight from station X to station Y that departs at 6 o'clock on weekdays, and operates according to the route specified by R0 and the timetable specified by ST10000. . "XY-LINE-OUT-H610" represents a holiday flight corresponding to "XY-LINE-OUT-W600". In this example, the routes to and from holidays are the same, but the departure times are different, so the group IDs (1D104) are different. On the other hand, "XY-LINE-OUT-W700" represents an outbound flight from Station X to Station Y that departs one hour after "XY-LINE-OUT-W600". This flight has the same route and timetable elapsed time as "XY-LINE-OUT-W600", so G10000, which has the same group ID (1D104) as "XY-LINE-OUT-W600", is registered. There is.

Next, with reference to FIG. 4, timetable data 1D2 collected by the operation state estimation server 10 from the traffic information service 13 and managed by the data management unit 106 will be described. Timetable data 1D2 is data representing a timetable, and includes a timetable ID (1D201), timetable order (1D202), arrival time (1D203), departure time (1D204), and stop ID (1D205). , is provided. The timetable ID (1D201) corresponds to the timetable ID (1D103) of the flight data 1D1, and the timetable data is specified in combination with the timetable order (1D202). The timetable order (1D202) represents the order of stops of a certain flight, and represents that the stops are visited in order from the smallest number to the largest number. Note that the numbers do not need to be consecutive. The arrival time (1D203) and departure time (1D204) represent the arrival time and departure time of the stop. Note that if there is almost no stopping time, the same time is registered for these two. Buses often stop at the same time, but trains often stop for one to several minutes. The stop ID (1D205) represents an identifier for specifying the stop of a certain flight.

A detailed explanation of the data shown in FIG. 4 will be supplemented. This example includes the aforementioned timetable "XY-LINE-OUT-W600" and indicates that the stop IDs are "S1, S2, S3, . . . ". It also indicates that the arrival and departure times will be "6:00, 6:01, 6:04, . . .".

Next, with reference to FIG. 5, route data 1D3 that is collected by the driving state estimation server 10 from the traffic information service 13 and managed by the data management unit 106 will be described. The route data 1D3 represents the route along which the vehicle travels, and includes a route ID (1D301), a route number (1D302), a route latitude (1D303), a route longitude (1D304), a stop ID (1D305), A stop name (1D306) is provided. Note that the route data 1D3 has finer granularity than the timetable data 1D2. The route ID (1D301) corresponds to the route ID (1D105) of the flight data 1D1, and is an identifier for specifying the route data 1D3 together with the route number (1D302). The route number (1D302) represents the running order of each position on the route, and consecutive numbers such as 0, 1, 2, . . . are registered. The route latitude (1D303) and the route longitude (1D304) represent the position in the geographic coordinate system. The stop ID (1D305) and stop name (1D306) represent an identifier and a name for specifying a stop if the position is a stop. If it is not a stop position (for example, if it is not a bus stop), N/A is registered.

A detailed explanation of the data shown in FIG. 5 will be supplemented. This example shows a route where the route ID (1D105) is R0, and position information and stop information are registered for each consecutive route number (ID302). In particular, route numbers 0, 2, and 4 are stop positions, and a stop ID (1D305) and a stop name (1D306) are registered. On the other hand, route numbers 1 and 3 are not stops, and N/A is registered in the stop ID (1D305) and stop name (1D306).

Next, real-time position data 1D4 collected by the driving state estimation server 10 from the vehicle 11 or the environmental equipment 12 and managed by the data management unit 106 will be described with reference to FIG. The real-time position data 1D4 is data related to a position acquired in real time, and is highly accurate data with respect to the measurement time. The real-time position data 1D4 includes a flight ID (1D401), a date and time (1D402), a vehicle latitude (1D403), a vehicle longitude (1D404), and a passed route number (1D405). The flight ID (1D401) corresponds to the flight ID (1D101) of the flight data 1D1, and the real-time location data 1D4 is specified using the flight ID (1D401) and date and time (1D402). The date and time (1D402) represents the measurement time of the vehicle position. The vehicle latitude (1D403) and the vehicle longitude (1D404) represent the position in the geographic coordinate system. In addition, the passed route number (1D405) is registered to distinguish cases where the same location is passed multiple times as described above.

A detailed explanation of the data shown in FIG. 6 will be supplemented. This example includes location data collected for the flight "XY-LINE-OUT-W600" and records four data values from 6:00 to 6:01:31. Further, the passed route number (1D405) is 0 at the beginning and becomes 1 at the stage of the third data. This indicates that the first stop was passed before the third data.

Next, the passage time estimation model data 1D5 generated by the operation state estimation server 10 and managed by the data management unit 106 will be explained with reference to FIG. The passage time estimation model data 1D5 includes a group ID (1D501), a construction date (1D502), a model parameter (1D503), and an offset (1D504). The group ID (1D501) corresponds to the group ID (1D104) of flight data 1D1, and if two flights have the same route and the elapsed time of the timetable (when the first departure time is set to time 0), This is an identifier to ensure that the elapsed time) is also the same. In this embodiment, a passing time estimation model is constructed for each group ID (1D501). Here, the construction date (1D502) is the date on which the passing time estimation model was constructed. The model parameter (1D503) is a parameter for expressing the passing time estimation model. The offset (1D504) is a value representing the average difference between the constructed passage time estimation model and the timetable. In this embodiment, the unit is minutes. Note that the offset (1D504) is not an essential requirement and may be omitted.

A detailed explanation of the data shown in FIG. 7 will be supplemented. In this example, a passing time estimation model is constructed on February 1, 2021 regarding group ID "G10000". Further, the offset (1D504) is 0.3 minutes.

Next, the estimated passage time and delay time data 1D6 generated by the operation state estimation server 10 and managed by the data management unit 106 will be described with reference to FIG. Estimated passage time and delay time data 1D6 includes flight ID (1D601), date and time (1D602), vehicle latitude (1D603), vehicle longitude (1D604), passed route number (1D605), and estimated passage time ( 1D606) and a delay time (1D607). The flight ID (1D601) corresponds to the flight ID (1D101) of the flight data 1D1, and the estimated transit time and delay time data 1D6 are specified using the flight ID (1D601) and date and time (1D602). The date and time (1D602) represents the measurement time of the vehicle position. The vehicle longitude (1D603) and the vehicle longitude (1D604) represent the position in the geographic coordinate system. The passed route number (1D605) is registered to distinguish cases where the same location is passed multiple times. The estimated passing time (1D606) is an estimated value of the passing time calculated using the above-mentioned passing time estimation model, and is the elapsed time with the time of departure from the starting location as 0. In this embodiment, the unit is minutes. The delay time (1D607) is an estimated value of the delay time obtained from the difference between the date and time (1D602) and the estimated passing time (1D606). The delay time (1D607) can be considered to represent the estimated delay time at each time in the above-mentioned outline.

A detailed explanation of the data shown in FIG. 8 will be supplemented. In this example, estimated transit time and delay time data 1D6 for flight "XY-LINE-OUT-W600" are registered, and the first departure is at the fourth location (latitude: 35.399571, longitude: 139.539084). The estimated passing time is 1.4 minutes when the departure time from the location is 0, and the delay time is 0.13 minutes.

Next, the route number delay time data 1D7 generated by the operation state estimation server 10 and managed by the data management unit 106 will be explained with reference to FIG. The delay time data 1D7 for each route number includes a flight ID (1D701), a starting date and time (1D702), a route number (1D703), and a delay time (1D704). The flight ID (1D701) corresponds to the flight ID (1D101) of the flight data 1D1, and the delay time data 1D7 for each route number is specified using the first departure date and time (1D702) and the route number (1D703). The first departure date and time (1D702) is the first departure date and time of the flight. The route number (1D703) corresponds to the route number (1D302) of the route data 1D3, and represents the running order of each position on the route, and consecutive numbers such as 0, 1, 2, etc. are registered. be done. The delay time (1D704) is an estimated value of the delay time at each position. The delay time (1D704) can be considered to represent the standardized estimated delay time in the above-mentioned outline.

A detailed explanation of the data shown in FIG. 9 will be supplemented. In this example, delay time data 1D7 for each route number of flight "XY-LINE-OUT-W600" is registered, and the fourth position (route ID: R0, route number: 3, latitude: 35.399711, Longitude: 139.539513), the delay time is 0.12 minutes.

Next, referring to FIG. 10, the delay is generated by combining the information collected from the traffic information service 13 and the weather information service 14 and the information generated by the operation state estimation server 10, and is managed by the data management unit 106. The time prediction model training data 1D8 will be explained. The delay time prediction model training data 1D8 includes the group ID (1D801), the flight ID (1D802), the first departure date and time (1D803), the weather (1D804), the traffic volume (1D805), and the delay time column (1D806). , is provided. The group ID (1D801) corresponds to the group ID (1D104) of the flight data 1D1, and together with the flight ID (1D802) and the first departure date and time (1D803), the delay time prediction model training data 1D8 is specified. The flight ID (1D802) corresponds to the flight ID (1D101) of the flight data 1D1. The first departure date and time (1D803) is the first departure date and time of the flight. The weather (1D804) is collected from the weather information service 14, and in this embodiment, one of four values related to the flight: sunny, cloudy, rainy, and snowy is registered. Note that the types of weather are not limited to four. Further, more detailed information such as probability of precipitation, temperature, humidity, wind speed, etc. may be registered, or may be registered with finer granularity. Traffic volume (1D805) is collected from the traffic information service 13 and represents the congestion status of the road. In this embodiment, integer values from 0 to 100 are registered for the congestion status related to flights, and the congestion status is indicated by the size of the integer value. Note that the information may be registered with finer granularity. The delay time column (1D806) is data obtained by grouping the delay time data 1D7 for each route number into group IDs (1D801), flight IDs (1D802), and first departure dates and times (1D803).

A detailed explanation of the data shown in FIG. 10 will be supplemented. In this example, the group ID is G10000, the flight ID is XY-LINE-OUT-W600, the first departure date and time is 2021-03-01T06:00:00, and the weather (sunny), traffic volume (9), delay time column ([0.00, 0.04, 0.10, 0.12,...]) is registered.

Next, delay time prediction model data 1D9 generated by the operation state estimation server 10 and managed by the data management unit 106 will be explained with reference to FIG. The delay time prediction model data 1D9 includes a group ID (1D901), a construction date (1D902), and a model parameter (1D903). The group ID (1D901) corresponds to the group ID (1D104) of flight data 1D1, and the route of two flights is the same, and the elapsed time of the timetable (when the first departure time is set to time 0) This is an identifier to ensure that the elapsed time) is also the same. In this embodiment, a delay time prediction model is constructed for each group ID (1D901). The construction date (1D902) is the date on which the delay time prediction model was constructed. The model parameter (1D903) is a parameter for expressing a delay time prediction model.

A detailed explanation of the data shown in FIG. 11 will be supplemented. In this example, a delay time prediction model is constructed on February 1, 2021 regarding group ID "G10000".

Next, the predicted delay time data 1D10 by route order, which is generated by the operation state estimation server 10 and managed by the data management unit 106, will be described with reference to FIG. The predicted delay time data 1D10 by route order includes a flight ID (1D1001), a starting date and time (1D1002), a route number (1D1003), and a predicted delay time (1D1004). The flight ID (1D1001) corresponds to the flight ID (1D101) of the flight data 1D1, and the delay time data 1D7 for each route number is specified using the first departure date and time (1D1002) and the route number (1D1003). The first departure date and time (1D1002) is the first departure date and time of the flight. The route number (1D1003) corresponds to the route number (1D302) of the route data 1D3, and represents the running order of each position on the route, and consecutive numbers such as 0, 1, 2, etc. are registered. be done. The predicted delay time (1D1004) is an estimated value of the delay time at each position calculated using the delay time prediction model. Unlike the delay time (1D704) of the route number delay time data 1D7, a predicted value of future delay time is included.

A detailed explanation of the data shown in FIG. 12 will be supplemented. In this example, route order predicted delay time data 1D10 for flight "XY-LINE-OUT-W600" is registered, and the fourth position (route ID: R0, route number: 3, latitude: 35.399711) is registered. , longitude: 139.539513), the predicted delay time is 0.11 minutes.

Next, the reliability data 1D11 generated by the operation state estimation server 10 and managed by the data management unit 106 will be explained with reference to FIG. The reliability data 1D11 includes a flight ID (1D1101), the first flight date and time (1D1102), a reconstruction error-based reliability (1D1103), and a mask-based reliability (1D1104). The flight ID (1D1101) corresponds to the flight ID (1D101) of the flight data 1D1, and is an identifier that specifies the reliability data 1D11. The restoration error-based reliability (1D1103) is the reliability calculated from the delay time (1D704) of the delay time data for each route number 1D7 and the predicted delay time (1D1004) of the predicted delay time data for each route order 1D10. The mask-based reliability (1D1104) is a reliability based on multiple predictions using a delay time prediction model.

A detailed explanation of the data shown in FIG. 13 will be supplemented. In this example, the reconstruction error-based reliability (1D1103) of the flight "XY-LINE-OUT-W600" that departed at 6:00 on March 1, 2021 is 88, and the mask-based reliability (1D1104) is 89.

Next, the processing flow of the construction phase in this embodiment will be explained using FIG. 14.

First, the collection unit 101 of the operation state estimation server 10 collects flight data 1D1, timetable data 1D2, and route data 1D3 from the traffic information service 13, and stores them in the data management unit 106 (step 1F101). Note that flight data 1D1, timetable data 1D2, and route data 1D3 may be generated from the data collected from the traffic information service 13 and stored.

Next, the collection unit 101 of the operation state estimation server 10 collects real-time position data 1D4 of the vehicle 11 from the vehicle 11, environmental equipment 12, traffic information service 13, etc. over a predetermined period. Estimated passing time and delay time data 1D6 are then generated and stored in the data management unit 106. Further, traffic data is collected from the traffic information service 13 and weather data is collected from the weather information service 14, and these data are stored in the data management unit 106 (step 1F102).

Next, the passage time estimation model construction unit 102 of the operation state estimation server 10 constructs a passage time estimation model for each group ID using the flight data 1D1, timetable data 1D2, and real-time position data 1D4, and estimates the passage time. Model data 1D5 is generated and stored in the data management unit 106 (step 1F103). Note that this process will be explained in detail later.

Next, the delay time prediction model construction unit 103 of the operation state estimation server 10 uses the constructed passing time estimation model data 1D5 and the position information (vehicle latitude, vehicle longitude, passed route number) of the real-time position data 1D4 to The estimated transit time (1D606) for each flight and each date and time is calculated. Then, the delay time prediction model construction unit 103 of the operation state estimation server 10 calculates the delay time (1D607) obtained as the difference from the date and time, generates estimated passage time and delay time data 1D6, and sends it to the data management unit 106. Store (step 1F104).

Next, the delay time prediction model construction unit 103 of the operation state estimation server 10 uses the estimated passage time and delay time data 1D6 and the route data 1D3 to generate a fixed length delay time string for each group ID, Delay time data 1D7 for each route number is generated and stored in the data management unit 106 (step 1F105). Note that this process will be explained in detail later.

Next, the delay time prediction model construction unit 103 of the operation state estimation server 10 combines the delay time data 1D7 for each route number with the traffic volume, weather, and delay time sequences to generate delay time prediction model training data 1D8, The data is stored in the data management unit 106 (step 1F106).

Finally, the delay time prediction model construction unit 103 of the operation state estimation server 10 learns (constructs) a delay time prediction model, generates delay time prediction model data 1D9, and stores it in the data management unit 106 (step 1F107). . Note that the delay time prediction model will be explained in detail later.

Next, with reference to FIG. 15, the processing of the passage time estimation model construction unit 102 of the operation state estimation server 10 in step 1F103 will be described in detail. The passage time estimation model construction unit 102 implements a passage time estimation model construction function that constructs a passage time estimation model using machine learning or probability statistics.

First, the flight data 1D1 and the real-time position data 1D4 are combined using the flight ID (step 1F201).

Next, the passing time (the time elapsed since the first train time is set to 0) is calculated based on the first train time in the timetable data 1D2 (step 1F202). The position at which the passing time is calculated can be any position on the route. Here, the passing time of the stop position may be calculated by referring to the departure time (1D204) of the timetable data 1D2, for example. Further, the passing time of a position different from a stop on the route may be calculated, for example, by linearly interpolating the data of the stop position.

Next, for each group ID, a passing time estimation model is learned by quantile regression of passing time with respect to vehicle latitude, vehicle longitude, and passed route number (step 1F203). In this embodiment, the regression quantile is set to 10 percent, but it can be changed to an appropriate value as needed. In quantile regression, by using a quantile near the lower limit, elapsed time with less influence of delay is determined. As an example, by estimating elapsed time using quantile regression for relatively small percentile values (e.g., the 10th percentile), it is expected that the elapsed time at a stop can be calculated without much difference from that calculated from the timetable. be done. In addition, in this embodiment, an example using quantile regression has been explained, but regarding a cumulative distribution function in which distribution is estimated using a generative model and integrated from the larger to the smaller, the point at which 90% has passed has elapsed. It may be calculated as time.

Finally, the average of the differences between the passing time obtained from the passing time estimation model at each stop and the passing time calculated from the actual timetable is calculated as an offset. Then, passing time estimation model data 1D5 is generated together with the above information and stored in the data management unit 106. By using this offset (1D504) as a correction value, it is possible to cope with constant delays that occur due to circumstances such as not being able to travel earlier than the timetable, as described above. In addition, in this embodiment, the same (single) offset is used in all sections, but for example, an offset is calculated for each stop, and in the section between them, the closer offset is used, or two offsets are used. An interpolated value of may be used.

The delay time prediction model construction unit 103 realizes a delay time prediction model construction function that constructs a delay time prediction model that predicts the future delay time of a mobile object. Next, the processing of the delay time prediction model construction unit 103 of the operation state estimation server 10 in step 1F105 will be described in detail with reference to FIG. 16. Note that in this explanation, processing related to a certain flight ID will be explained.

First, the flight data 1D1 and the route data 1D3 are combined using the route ID, and the route number, route latitude, and route longitude associated with the flight ID of interest are obtained (step 1F301). Note that, in the processing flow of FIG. 16, these are fixed lengths for each flight ID.

Next, from the estimated passage time and delay time data 1D6, five pieces of position data are acquired for each of the aforementioned route numbers, starting from the route number difference of 2 or less and the closest distance (step 1F302). Note that in this embodiment, the difference in route numbers is within 2, but other values may be used. Similarly, although the number of location data to be acquired is set to five, other values may be used.

Finally, the average value of the five acquired data is calculated and taken as the delay time for the corresponding route number. Through this procedure, delay time data 1D7 for each route number is generated and stored in the data management unit 106 (step 1F303).

Through the above processing, the estimated delay time is converted to a fixed length for the flight ID, or more precisely for flights traveling on the same route (that is, the estimated delay time at the time is converted to the estimated delay time at a predetermined position). ) to obtain a fixed-length delay time sequence. This facilitates modeling using statistics and machine learning. More specifically, a structure is realized in which vectors of the same length are accepted for input and output of a delay time prediction model, which will be described later. Furthermore, it can be used to calculate reliability, which will be described later. Note that a fixed length delay time for a position on the route may be obtained using the above method (the method of steps 1F301 to 1F303), or a known data search method such as the k-nearest neighbor method may be used. A fixed length delay time for a position on the route may be obtained.

Next, the delay time prediction model will be explained in detail with reference to FIG. 17. In this embodiment, a delay time prediction model is constructed for each group ID using a transformer, which is a type of neural network.

First, the overall input and output will be explained. The input is a partially masked delay time sequence, weather and traffic, and the output is a (unmasked) delay time sequence. By using a model trained on tasks with such inputs and outputs, it becomes possible to predict future delay times given known delay times and surrounding information during operation.

Next, the mask will be explained. During training, a different mask is randomly generated for each mini-batch and training is performed. At that time, a certain route number on the route is selected and the values after that route number are set to 0.

Next, the embedding process will be explained. First, a convolution layer is applied to a vector representing a delay time sequence. Thereafter, by concatenating with vectors representing weather and traffic volume, a vector containing both delay time and surrounding information is obtained.

Next, the process of position encoding will be explained. In Position Encoding, in order to deal with the loss of position information regarding vector elements in the Self-attention layer, it is possible to add a vector created using a different sine wave (appropriate sine wave) to the input vector. Common. In this embodiment, in addition to this, the information on the month, day of the week, weekday or holiday, and time zone is embedded by adding a predetermined vector as an offset for each month, day of the week, weekday or holiday, and time zone information. It will be done.

Next, self-attention processing will be explained. Basically, a layer in which a general multi-head Self-attention layer and Point Feed Forward are stacked three times is applied. However, since the input is masked, correspondingly, the value immediately before passing through the softmax function (the product of the query and the key) is masked. Note that although the number of stacks is set to three, it may be more or less than this. Furthermore, information on weather and traffic volume may be added as appropriate to the residual block of the query, key, and Point Feed Forward.

Next, the processing of Linear Transform will be explained. Here, the fully connected layer is applied twice. Although the total number is set to 2, it may be more or less than that. Furthermore, in cases where restoration is difficult, such as when the delay time sequence is extremely long, a deconvolution layer or the like may be used.

Finally, loss function calculation processing will be explained. Here, the MSE (Mean Squared Error) of the restored estimated delay time and the original delay time sequence is calculated, and the loss of the model is estimated. The learning of the model progresses by updating the parameters to minimize this loss using a method such as back error propagation. Although MSE is used in this embodiment, MAE (Mean Absolute Error), hinge loss, etc. may also be used.

Next, the processing flow of the inference phase in this embodiment will be described with reference to FIG. It is assumed that prior to this process, the construction phase described in FIG. 14 has been executed (that is, the passage time estimation model and the delay time prediction model have been constructed).

In the inference phase, first, the collection unit 101 of the driving state estimation server 10 collects real-time position data 1D4 regarding the running vehicle 11 from the vehicle 11, environmental equipment 12, traffic information service 13, and the like. Further, traffic data is collected from the traffic information service 13 and weather data is collected from the weather information service 14, and these data are stored in the data management unit 106 (step 1F401).

Next, the inference unit 104 of the operation state estimation server 10 uses the transit time estimation model constructed in the construction phase to calculate the estimated transit time (1D606) and delay time (1D607) from departure to the present for each flight. is calculated (step 1F402). In this way, the inference unit 104 realizes an inference function that calculates the current estimated passing time and delay time.

Next, the inference unit 104 of the operation state estimation server 10 generates a delay time sequence for each group ID using the estimated passing time and delay time data 1D6 and route data 1D3 (step 1F403). Here, by making the length fixed even during prediction, data handling becomes easier. Note that this procedure is basically the same as described above using FIG. 16. However, if there is a part where the condition that the route number is within 2 is not met, the length will not be fixed.

Next, the inference unit 104 of the operation state estimation server 10 uses the delay time prediction model constructed for each group ID in the construction phase to calculate the weather and traffic volume collected in step 1F401 and the delay obtained in step 1F403. From the time sequence, the predicted delay time for each flight ID, first departure date and time, and route number is calculated, and predicted delay time data 1D10 for each route order is generated and stored in the data management unit 106 (step 1F404). Note that, as described above, the delay time sequence for each group ID may not have a fixed length. In this case, the value is set using the same concept as the mask used when learning the delay time prediction model explained using Figure 17 (in other words, by selecting a certain route number and setting the values after that route number to 0). Complete it and make it a fixed length.

Next, the inference unit 104 of the operation state estimation server 10 calculates a delay time string for each group ID of a flight on a certain date and time, and a delay time string of predicted delay time data 1D10 for each route order that the group and date and time overlap (i.e. , the delay time sequence of the predicted delay time (1D1004)), and the average value of the absolute value of the difference (that is, the average value of the restoration error) is calculated. Then, with respect to the predetermined constant D, the result of 100×(D−(the average value))/D is taken as the restoration error-based reliability. Since this reconstruction error-based reliability uses a delay time sequence, it is the average value of the absolute value of the vector difference over a short period at the beginning of the run, but as the run progresses, the absolute value of the vector difference over a longer period becomes This is the average value. However, if the reconstruction error-based reliability is less than 0, the reconstruction error-based reliability is replaced with 0. Here, in this embodiment, the 90th percentile value of the average value of restoration errors during learning regarding the same delay time prediction model is used as the constant D, but other values may be used (step 1F405).

Next, the inference unit 104 of the operation state estimation server 10 creates a delay time sequence by masking the delay time sequence for each group ID by 5 minutes, 10 minutes, and 15 minutes from the current time. That is, using masks of different sizes, a delay time sequence is created that is masked from 5 minutes before, 10 minutes before, and 15 minutes before the current time. Then, the inference unit 104 of the operation state estimation server 10 predicts the predicted delay time by route order for four patterns including no mask (no mask, 5-minute mask, 10-minute mask, and 15-minute mask), and predicts the predicted delay time for each route sequence after the current time. Calculate the standard deviation of predicted delay times for each route order for each pattern. Then, the value of 100×(S−(the standard deviation))/S becomes the mask-based reliability. However, if the mask-based reliability is less than 0, it is replaced with 0, and if the mask-based reliability exceeds 100, it is replaced with 100. Here, in this embodiment, the 90th percentile value of the standard deviation during learning regarding the same delay time prediction model is used as the constant S, but other values may be used (step 1F406).

Next, the inference unit 104 of the operation state estimation server 10 generates reliability data 1D11 based on the obtained restoration error-based reliability and mask-based reliability, and generates the predicted delay by route order generated up to the above steps. The data is stored in the data management unit 106 over time. Then, the publishing unit 105 of the operation state estimation server 10 notifies the user terminal 15 of the updated information. The publishing unit 105 implements a presentation function that presents the operating status of the mobile object to the user. Finally, the user terminal 15 reads the latest route order predicted delay time and reliability data 1D11 based on the notification, and presents these data to the user 16 as the operation status (step 1F407).

Next, the user interface will be explained with reference to the drawings. FIG. 19 is an example of a flight status display screen 1G1 presented to the user 16 by the user terminal 15. Note that the user 16 can operate the user terminal 15 to specify the flight he or she wishes to display.

The operation status display screen 1G1 includes a flight name label (1G101), an arrival time table (1G102), a reliability table (1G103), and a delay status graph (1G104). The flight name of the flight specified by the user 16 is displayed on the flight number label (1G101). The arrival time table (1G102) displays the arrival time and delay time obtained using the timetable information and the delay time prediction model for the flight specified by the user 16. In the example of FIG. 19, if the timetable had been followed, the vehicle would have arrived at Stop G at 6:30, but it is displayed that the vehicle was delayed by 4 minutes at 6:34. Further, if the timetable is as per the timetable, the arrival at Stop F is scheduled to be at 6:40, but it is shown that the arrival is scheduled to be delayed by 5 minutes at 6:45. The reliability table (1G103) displays restoration error-based reliability and mask-based reliability obtained using the delay time prediction model. Note that in this embodiment, both reliability levels are displayed, but it is also possible to display either one or not. The delay status graph (1G104) displays predicted delay times for each route order obtained using the delay time prediction model. Here, the horizontal axis is information including the position of a stop on the route, and the vertical axis is information indicating predicted delay time. Note that the position of the stop can be indicated as appropriate, and as shown in FIG. 19, for example, a symbol may be displayed at the position of the stop. Further, a stop name (corresponding to stop A, stop G, etc. in FIG. 19) corresponding to the route number may be displayed at the stop position. Further, the current position is visualized as a broken line (1G104a). Further, the delay time obtained from the traveled position data (delay time determined without using the delay time prediction model) is visualized as a solid line (1G104b). Further, the predicted (restored) delay time for the traveled area is visualized as a narrow broken line (1G104c). On the other hand, the predicted delay time for the untraveled area is visualized as a wide broken line (1G104d). Note that the restoration error-based reliability is a value calculated based on the average of the absolute values of the differences between the solid line (1G104b) and the narrow broken line (1G104c). Further, the delay status graph (1G104) may be generated by an appropriate method. The delay situation graph (1G104) may be generated, for example, by a regression curve based on the delay time sequence.

As explained above, according to the operation state estimation system in this embodiment, by using a passing time estimation model that estimates the elapsed time from departure for information specifying the position of an arbitrary moving object, it is possible to It becomes possible to estimate the current delay time and predict future delay times using the results of the model.

Furthermore, if information to identify the location of a moving object in operation is obtained, it is possible to predict the delay time of the moving object even if accurate arrival and departure times cannot be obtained. . For example, buses and streetcars may pass through without passengers, and even if it is difficult to obtain accurate arrival and departure times at the stop, the delay time of the moving object is Can be predicted.

It should be noted that the present invention is not limited to the above-described embodiments as they are, but can be embodied by modifying the constituent elements without departing from the scope of the invention. Moreover, various inventions can be formed by appropriately combining the plurality of components disclosed in the above embodiments. For example, some components may be deleted from all the components shown in the embodiments.

The processing main body of the operation state estimation server 10 is a processor, and a CPU can be considered as an example of the processor, but other semiconductor devices (for example, GPU) may be used as long as the processing main body executes predetermined processing.

The transit time estimation model and the delay time prediction model can be generated by a method based on probability statistics or machine learning using acquired or generated data.

In the above explanation, when generating the delay time prediction model, the delay time prediction model construction unit 103 calculates the estimated passing time and the delay time. A delay time may also be used.

In this specification, a "stop" is a place where a moving object stops, and for example, a bus stop for a bus or a station for a train.

10 Operation state estimation server 11 Vehicle 12 Environmental equipment 13 Traffic information service 14 Weather information service 15 User terminal 16 User 101 Collection unit 102 Passage time estimation model construction unit 103 Delay time prediction model construction unit 104 Inference unit 105 Publication unit 106 Data management unit

Claims (8)

Realized by a processor, using at least route information, a timetable, and position information of a mobile body based on a travel history, or information on a position of a stop of the mobile body and a timetable at the stop; information that can be obtained by linearly complementing the information, and in units that have commonality in the operation pattern, in response to input of information on the current position of the mobile object, from the departure of the mobile object. a passage time estimation model construction unit that constructs a passage time estimation model for estimating elapsed time;
an inference unit that is realized by a processor and calculates the current delay time of the mobile object from the elapsed time output from the passage time estimation model;
a publishing unit that is realized by a processor and presents to a user the operating state of the mobile body determined by execution of the inference unit;
Converting delay time information obtained from the elapsed time realized by a processor and output from the passage time estimation model into a delay time string that is fixed length delay time information, and using the delay time string, a delay time prediction model construction unit that constructs a delay time prediction model that predicts the future delay time of the mobile body in units that have commonality in operation patterns in response to input of delay time information of the mobile body; Equipped with
The reasoning unit calculates a future delay time of the mobile body from information on the delay time of the mobile body.
An operation state estimation system characterized by: The operating state estimation system according to claim 1 ,
The passage time estimation model construction unit determines the commonality between moving bodies that have the same route and different start times but have the same elapsed time from the start time to arrival at each stop. Build a transit time estimation model by collecting data in units that have commonality.
An operation state estimation system characterized by: The operating state estimation system according to claim 1 ,
The delay time prediction model construction unit further uses at least one of weather information, traffic information, month information, day of the week information, weekday or holiday information, and time zone information. and construct the delay time prediction model ,
The inference unit calculates a future delay time of the mobile object from information on the delay time of the mobile object and the at least one piece of information used to construct the delay time prediction model.
An operation state estimation system characterized by: The operating state estimation system according to claim 1 ,
The passage time estimation model construction unit constructs the passage time estimation model using quantile regression of the elapsed time with respect to the position of the moving object.
An operation state estimation system characterized by: The operating state estimation system according to claim 1 ,
The delay time prediction model construction unit constructs the delay time prediction model by including data obtained by masking a part of the fixed length delay time sequence ,
The inference unit masks the delay time from the current point to the end point with respect to the delay time string from the starting point to the current point, thereby creating a delay time string having the same length as the fixed length delay time string. and using the generated delay time sequence to restore the delay time from the first train to the current point, and to predict the delay time from the current point to the end point.
An operation state estimation system characterized by: The operating state estimation system according to claim 1 ,
The inference unit predicts the delay time at a predetermined position on the route based on the fixed length delay time sequence,
The publishing unit presents the predicted delay time at a predetermined position on the route to the user.
An operation state estimation system characterized by: The operating state estimation system according to claim 5 ,
The inference unit calculates a reconstruction error-based reliability based on a reconstruction error of a delay time sequence of a mobile object from the first train to the present, and a delay time sequence of the same date and time restored from the delay time prediction model, and/ Alternatively, a mask-based reliability is calculated based on the delay time predicted by the delay time prediction model using a delay time sequence in which a predetermined range from the present to the past is masked,
The publishing unit presents the calculated restoration error-based reliability and/or the calculated mask-based reliability to the user together with a future delay time.
An operation state estimation system characterized by: A method of estimating operating conditions using an electronic computer,
The computer uses at least route information, a timetable, and position information of the mobile body based on the travel history, or information on the position of the mobile body's stop and the timetable at the stop, and the Information that can be obtained by linear interpolation of the information, and the progress since the departure of the mobile object in units that have commonality in the operation pattern according to the input of the information of the current position of the mobile object. Build a transit time estimation model to estimate time ,
The electronic computer calculates the current delay time of the mobile object from the elapsed time output from the passage time estimation model ,
The computer converts delay time information obtained from the elapsed time output from the passage time estimation model into a delay time sequence that is fixed length delay time information,
The electronic computer uses the delay time sequence to predict the future delay time of the mobile object in units that have commonality in the operation pattern, in response to input of information on the delay time of the mobile object. Build a predictive model,
The computer calculates a future delay time of the mobile body from information on the delay time of the mobile body,
The computer outputs the determined operating state of the mobile object to a user terminal.
A driving state estimation method characterized by:

Images