Abnormality detection method, abnormality detection device, and program

Abstract

This abnormality detection method is an abnormality detection method that is executed by a computer and is for detecting an abnormality of an autonomous traveling vehicle autonomously traveling from a departure location to a destination. The abnormality detection method includes a step for acquiring a travel route of the autonomous travel vehicle leaving from a departure location to an arbitrary point in time, and a step (S7300) for reading out information regarding a plurality of routes over which the autonomous traveling vehicle can travel from the departure location to the destination, using information regarding the plurality of routes that have been read out as a basis to estimate one of the plurality of routes as an estimated travel route from the departure location to the destination of the autonomous traveling vehicle, calculating an abnormality degree that is a degree indicating to what extent the travel route to the arbitrary point in time is deviated from the estimated travel route, and detecting occurrence of an abnormality in the autonomous travel vehicle when the calculated abnormality degree exceeds a predetermined threshold value.

Description

Anomaly detection method, anomaly detection device, and program

[0001] The present disclosure relates to an anomaly detection method for detecting an anomaly related to an autonomous vehicle.

[0002] In recent years, various efforts have been made to put autonomous vehicles into practical use. For example, demonstration experiments are being conducted on services using autonomous vehicles, such as indoor and outdoor delivery of goods and transportation of people, as well as cleaning robots and security robots.

[0003] To provide safe services for autonomous vehicles, they must communicate with remote monitoring points to monitor their status and perform emergency operations. However, the ability to communicate with remote monitoring points exposes them to the risk of cyberattacks. Therefore, to reduce the risk of cyberattacks, anomaly detection methods that assume cyberattacks on autonomous vehicles have been proposed.

[0004] For example, Patent Document 1 discloses a method for detecting abnormalities in a vehicle that makes use of location characteristics such as current position information of the vehicle.

[0005] International Publication No. 2021 / 149340

[0006] According to the anomaly detection method disclosed in Patent Document 1, an optimal anomaly detection model that suits the characteristics of the location (for example, the difference between an ordinary road and an expressway) is selected based on the vehicle's current location information, and the selected anomaly detection model is used to detect an anomaly in the vehicle, on the assumption that the vehicle's current location information is correct.

[0007] However, if the vehicle's current location information is incorrect, for example, if the location information has been tampered with by a cyber attacker, it may be impossible to select the optimal anomaly detection model, making it impossible to detect anomalies. Even if the vehicle's current location information is correct and has not been tampered with, if an attacker illicitly operates the vehicle and sends it to a location different from its original destination, it may be difficult to confirm whether the vehicle is in an abnormal location. As such, there are some aspects of this technology that make it inappropriate for anomaly detection in autonomous vehicles.

[0008] Therefore, the present disclosure provides an anomaly detection method and the like that can perform anomaly detection more appropriately.

[0009] An anomaly detection method according to one aspect of the present disclosure is an anomaly detection method executed by a computer for detecting an anomaly in an autonomous vehicle that drives autonomously from a starting point to a destination, the anomaly detection method including the steps of: acquiring a driving route of the autonomous vehicle from departing the starting point to an arbitrary time point; reading information about a plurality of routes that the autonomous vehicle may take from the starting point to the destination; estimating one of the plurality of routes as an estimated driving route of the autonomous vehicle from the starting point to the destination based on the read information about the plurality of routes; calculating an anomaly degree that indicates how much the driving route up to the arbitrary time point deviates from the estimated estimated driving route; and detecting that an anomaly has occurred in the autonomous vehicle if the calculated anomaly degree exceeds a predetermined threshold.

[0010] Furthermore, an anomaly detection device according to one aspect of the present disclosure is an anomaly detection device that detects an anomaly in an autonomous vehicle that drives autonomously from a starting point to a destination, and includes: an acquisition unit that acquires a driving route of the autonomous vehicle from departing the starting point to an arbitrary point in time; a memory unit that stores information about a plurality of routes that the autonomous vehicle may drive from the starting point to the destination; and an anomaly detection unit that estimates one of the plurality of routes as an estimated driving route of the autonomous vehicle from the starting point to the destination based on the information about the plurality of routes read from the memory unit, calculates an anomaly degree that indicates how much the driving route up to the arbitrary point in time deviates from the estimated estimated driving route, and detects that an anomaly has occurred in the autonomous vehicle if the calculated anomaly degree exceeds a predetermined threshold.

[0011] Furthermore, a program according to one aspect of the present disclosure is a program for causing the computer to execute the anomaly detection method described above.

[0012] According to the above aspect, abnormality detection of an autonomous vehicle can be performed more appropriately.

[0013] FIG. 1 is a schematic diagram showing an example of route selection for an autonomous vehicle. FIG. 2 is a schematic diagram showing an example of route selection for an autonomous vehicle. FIG. 3 is a configuration diagram showing an example of the overall configuration of an anomaly detection system according to an embodiment. FIG. 4 is a schematic diagram showing an example of map information according to an embodiment. FIG. 5 is a configuration diagram showing an example of the configuration of an anomaly detection unit according to an embodiment. FIG. 6 is a schematic diagram showing an example of the data structure of vehicle information according to an embodiment. FIG. 7 is a schematic diagram showing an example of the data structure of information related to transported goods according to an embodiment. FIG. 8 is a schematic diagram showing an example of the data structure of weather information according to an embodiment. FIG. 9 is a schematic diagram showing an example of the data structure of road traffic information according to an embodiment. FIG. 10 is a schematic diagram showing an example of the data structure of operation management information according to an embodiment. FIG. 11 is a flowchart showing an example of the operation of an anomaly detection process according to an embodiment. FIG. 12 is a schematic diagram showing an example of a combination of route determination parameters according to an embodiment. FIG. 13A is a schematic diagram showing an example of the data structure of an estimation model according to an embodiment. FIG. 13B is a schematic diagram showing an example of the data structure of an estimation model according to an embodiment. FIG. 14A is a schematic diagram showing an example of a data structure of an anomaly detection model according to an embodiment. FIG. 14B is a schematic diagram showing an example of a data structure of an anomaly detection model according to an embodiment. FIG. 14C is a schematic diagram showing an example of a data structure of an anomaly detection model according to an embodiment. FIG. 15 is a flowchart showing an example of an operation of an anomaly detection model selection process according to an embodiment. FIG. 16 is a flowchart showing an example of an operation of an anomaly determination process according to an embodiment. FIG. 17A is a schematic diagram showing an example of a process of an anomaly determination process according to an embodiment. FIG. 17B is a schematic diagram showing an example of a process of an anomaly determination process according to an embodiment. FIG. 18 is a schematic diagram showing an example of a data structure of an anomaly detection result according to an embodiment. FIG. 19 is a configuration diagram showing an example of a configuration of a model learning unit according to an embodiment. FIG. 20 is a flowchart showing an example of an operation of a model learning process according to an embodiment. FIG. 21A is a schematic diagram showing an example of information regarding driving routes during past driving according to an embodiment.FIG. 21B is a schematic diagram showing an example of information related to driving routes during past driving in an embodiment. FIG. 22 is a flowchart showing an example of the operation of driving route determination parameter grouping processing in an embodiment. FIG. 23A is an example of a histogram showing the frequency of use of each driving route during past driving in an embodiment. FIG. 23B is an example of a histogram showing the frequency of use of each driving route during past driving in an embodiment. FIG. 23C is an example of a histogram showing the frequency of use of each driving route during past driving in an embodiment. FIG. 23D is an example of a histogram showing the frequency of use of each driving route during past driving in an embodiment. FIG. 24 is a schematic diagram showing an example of the process of driving route determination parameter grouping processing in an embodiment. FIG. 25 is a schematic diagram showing an example of the process of driving route determination parameter grouping processing in an embodiment. FIG. 26 is a flowchart showing an example of the operation of driving route determination parameter importance calculation processing in an embodiment. FIG. 27A is a schematic diagram showing an example of the process of driving route determination parameter importance calculation processing in an embodiment. FIG. 27B is a schematic diagram showing an example of the process of driving route determination parameter importance calculation processing in an embodiment. Fig. 28 is a flowchart showing an example of the operation of the anomaly detection model learning process according to the embodiment. Fig. 29A is a schematic diagram showing an example of the process of the anomaly detection model learning process according to the embodiment. Fig. 29B is a schematic diagram showing an example of the process of the anomaly detection model learning process according to the embodiment.

[0014] (Background to the development of one aspect of the present disclosure) While services using autonomous vehicles are expanding as a solution to labor shortages, IoT-enabled autonomous vehicles face the risk of attackers illegally accessing the control systems inside the autonomous vehicles from the outside by connecting via a network or directly connecting to devices. If the control system is compromised, a replay attack can be established with a simple command, making the autonomous vehicles at high risk of becoming targets of cyberattacks. As a countermeasure, many anomaly detection methods and devices have been proposed for remotely monitoring autonomous vehicles for abnormalities.

[0015] However, in remote anomaly detection for autonomous vehicles, the high incidence of overdetection due to the diverse driving patterns of autonomous vehicles has been seen as a problem. Taking into account the large number of targets for anomaly monitoring, suppressing overdetection in anomaly detection has become a challenge. Generally, in a target area or location, an autonomous vehicle determines its own driving route to a destination, recognizes and avoids obstacles using information from Lidar (Light Detection and Ranging) or cameras, and performs operations such as changing speed, turning, and turning on blinkers. Therefore, driving patterns are difficult to determine, and methods that simply perform threshold judgments on speed information, etc., often result in overdetection.

[0016] Therefore, as disclosed in Patent Literature 1, an anomaly detection method has been developed that suppresses overdetection by changing the anomaly detection model based on the current vehicle position information and performing anomaly detection that is more suited to the characteristics of the location. However, with the above-mentioned anomaly detection method, there is a risk that detection will be avoided and an anomaly will not be determined to be an anomaly in cases such as when the vehicle position information has been tampered with by an attacker, or when the vehicle is being operated fraudulently by an attacker but the driving at that location is not significantly different from normal operation.

[0017] In view of the above-mentioned problems, the inventors of the present application have come to the conclusion that it is necessary to determine whether or not there is an abnormality in the current driving route and position of an autonomous vehicle traveling toward a predetermined destination.

[0018] On the other hand, in order to determine whether the current driving route of an autonomous vehicle is normal or abnormal, information is required on the logic by which the route that the autonomous vehicle will actually drive is selected and determined from a large number of possible route candidates. As described above, an autonomous vehicle automatically determines the optimal driving route from the departure point to the destination using map information of the target area and location for the set destination. However, information such as the logic involved in determining the driving route is generally used only by the autonomous vehicle vendor and is often a black box that is not disclosed, posing a challenge in that it is difficult for operators who remotely monitor and detect abnormalities in vehicles to obtain this information.

[0019] Therefore, the inventors of the present application wondered whether it might be possible to estimate the logic by which an autonomous vehicle's route is selected and determined, and conducted a demonstration experiment of the driving of an autonomous vehicle. In the course of the demonstration experiment, the inventors discovered that even for the same combination of departure point and destination, the driving route of an autonomous vehicle is not necessarily the same each time.

[0020] As one of the demonstration experiments, Figures 1 and 2 show schematic diagrams showing an example of route selection when an autonomous vehicle used in goods delivery services is driven toward a destination.

[0021] FIG. 1 shows two routes (a first route and a second route) leading from a starting point to a destination, one above the other, and their branching points. The first route on the top side is closer to the destination but includes a bridge over a river and an unpaved road along the way, while the second route on the bottom side is farther to the destination but is mostly paved. In a demonstration experiment, it was confirmed that an autonomous vehicle would travel both the first route and the second route shown in FIG. 1 even for the same combination of starting point and destination. For example, the autonomous vehicle traveled the first route on sunny days or when the transported goods contained perishable goods, and the autonomous vehicle traveled the second route on rainy days or when the transported goods contained fragile goods.

[0022] Here, since the first route is closer to the destination, if the first route is selected, it is considered that the driving route of the autonomous vehicle was determined with an emphasis on the short distance to the destination or the short time required to reach the destination. On the other hand, if the second route is selected, it is considered that the driving route of the autonomous vehicle was determined with an emphasis on whether the driving route includes a dangerous area where river flooding may occur in bad weather, for example, by not selecting the first route that includes a bridge over a river. Alternatively, it is considered that the driving route of the autonomous vehicle was determined with an emphasis on the reduction of vibrations caused by road surface conditions during driving, for example, by not selecting the first route that includes an unpaved road.

[0023] FIG. 2 also shows two routes (Route 3 and Route 4) leading from the starting point to the destination, and their branches. Route 3 on the top is closer to the destination but includes a school route, while Route 4 on the bottom is farther to the destination but does not include a school route. In the example of FIG. 2, as in the example of FIG. 1, even with the same combination of starting point and destination, it was confirmed in demonstration experiments that the autonomous vehicle would travel both Route 3 and Route 4 of FIG. 2. For example, when traveling at night or when the transported goods contained perishable goods, the autonomous vehicle traveled Route 3, and when traveling in the morning or evening or when the transported goods contained fragile goods, the autonomous vehicle traveled Route 4.

[0024] Here, because the third route is closer to the destination, if the third route is selected, it is thought that the route of the autonomous vehicle was determined with an emphasis on the short distance to the destination or the short time required to reach the destination. On the other hand, if the fourth route is selected, it is thought that the route of the autonomous vehicle was determined with an emphasis on the low possibility of an accident due to a collision or the low possibility of delays in arriving at the destination due to route congestion, and for example, the third route, which includes a school route, was not selected during school hours.

[0025] In this way, the driving route of an autonomous vehicle is not determined simply based on the distance to the destination, but is also likely to be determined taking into account information such as the congestion level on the route at different times of the day, the risk of accidents, near-miss locations, whether there are road construction works and the condition of the pavement, and the impact of weather on the route, as well as information such as the type of cargo being transported and the transport conditions, when the autonomous vehicle is used for goods transport services.

[0026] As a result of extensive research, the inventors of the present application have come up with the idea of ​​predicting the driving route that will be selected and decided while the autonomous vehicle is traveling by estimating information or elements that are thought to be taken into account when the autonomous vehicle decides on its driving route, in other words, driving route decision parameters, as described above.The inventors have then come up with an anomaly detection method for detecting anomalies in an autonomous vehicle by selecting an optimal anomaly detection model for detecting anomalies based on the predicted driving route.

[0027] An anomaly detection method according to a first aspect of the present disclosure is an anomaly detection method executed by a computer for detecting an abnormality in an autonomous vehicle that drives autonomously from a starting point to a destination, the anomaly detection method including the steps of acquiring a driving route of the autonomous vehicle from leaving the starting point to an arbitrary time point, reading out information about a plurality of routes that the autonomous vehicle may take from the starting point to the destination, estimating one of the plurality of routes as an estimated driving route from the starting point to the destination of the autonomous vehicle based on the read information about the plurality of routes, calculating an abnormality degree that indicates how much the driving route up to the arbitrary time point deviates from the estimated driving route, and detecting that an abnormality has occurred in the autonomous vehicle if the calculated abnormality degree exceeds a predetermined threshold.

[0028] According to this, since the estimated driving route of the autonomous vehicle is estimated, anomalies can be detected based on the estimated driving route even if the position information of the autonomous vehicle has been tampered with or if unauthorized operations have been performed. In other words, if the route that the autonomous vehicle has traveled up to a given point in time deviates from the estimated driving route, it is possible to detect that an anomaly has occurred.

[0029] Furthermore, the anomaly detection method according to the second aspect is the anomaly detection method described in the first aspect, and further comprises acquiring information about a plurality of routes when a driving route up to an arbitrary point in time is acquired or each time a predetermined period of time has passed, and the information about the plurality of routes read to estimate an estimated driving route includes information about a plurality of routes acquired in the past.

[0030] This allows information about multiple routes to be updated, and enables estimation of an estimated travel route that corresponds to the route conditions that may change over time.

[0031] Furthermore, the anomaly detection method according to the third aspect is the anomaly detection method described in the first or second aspect, and further includes a step of, when an anomaly in the autonomous vehicle is detected in the detection step, displaying the detection result that an anomaly has occurred in the autonomous vehicle, and predetermined map information that maps the driving route from the departure point to an arbitrary point in time and the estimated driving route.

[0032] This makes it possible to output information relating to the driving route of the autonomous vehicle and information relating to the abnormality detection result.

[0033] Furthermore, an anomaly detection method according to a fourth aspect is an anomaly detection method described in any one of the first to third aspects, wherein the information about the multiple routes is a driving route determination parameter, and the driving route determination parameter is a parameter consisting of a combination of values ​​indicating the presence or absence of each of multiple events that may affect the determination of a driving route from a departure point to a destination by an autonomous vehicle, and in the detection step, an estimated driving route is estimated based on the driving route determination parameter.

[0034] This makes it possible to estimate an estimated driving route based on the presence or absence of each of a number of events that may affect the determination of a driving route by an autonomous vehicle, as information about the multiple routes.

[0035] Furthermore, an anomaly detection method according to a fifth aspect is the anomaly detection method described in the fourth aspect, wherein the information about the multiple routes includes at least one of the distance from the starting point to the destination of each of the multiple routes and the time required to travel from the starting point to the destination when an autonomous vehicle travels along each of the multiple routes.

[0036] This allows estimation of an estimated travel route based on information about distance or required time as information about a plurality of routes.

[0037] Furthermore, an anomaly detection method according to a sixth aspect is an anomaly detection method according to the fourth or fifth aspect, wherein the information about the multiple routes includes information indicating events that may become obstacles when an autonomous vehicle travels along each of the multiple routes at any point in time.

[0038] This makes it possible to estimate an estimated driving route based on information about events that could hinder the autonomous vehicle's driving on each route.

[0039] In addition, the anomaly detection method according to the seventh aspect is the anomaly detection method described in the sixth aspect, wherein the information indicating events that may be obstacles when an autonomous vehicle travels along each of the multiple routes includes information indicating whether or not the autonomous vehicle is capable of traveling along each of the multiple routes.

[0040] This makes it possible to estimate an estimated driving route based on information indicating routes among multiple routes that the autonomous vehicle cannot travel.

[0041] Furthermore, an anomaly detection method according to an eighth aspect is an anomaly detection method described in any one of the fourth to seventh aspects, wherein the autonomous vehicle is a vehicle used to transport goods, the acquiring step further acquires information about the goods to be transported by the autonomous vehicle, the information about the goods being a travel route determination parameter including at least one of one or more pieces of information indicating attributes of the goods and one or more pieces of information indicating transportation conditions for the goods, and the detecting step estimates an estimated travel route based on the information about the multiple routes and the information about the goods.

[0042] According to this, when the autonomous vehicle is a vehicle used to transport goods, it is possible to estimate an estimated travel path based on information about the goods in addition to information about the route.

[0043] In addition, an anomaly detection method according to a ninth aspect is an anomaly detection method described in any one of the fourth to eighth aspects, and further includes a step of selecting and reading out one anomaly detection model from a plurality of anomaly detection models corresponding to each of a plurality of driving route determination parameters, wherein one route from a plurality of routes is defined in each of the plurality of anomaly detection models, and in the reading step, the anomaly detection model corresponding to the driving route determination parameters is selected and read out, and in the detecting step, the one route defined in the read anomaly detection model is estimated as an estimated driving route.

[0044] This allows one route determined by the anomaly detection model corresponding to the travel route determination parameters to be estimated as the estimated travel route.

[0045] Furthermore, an anomaly detection method according to a tenth aspect is the anomaly detection method according to the ninth aspect, in which one path defined for each of a plurality of anomaly detection models is updated to any one of a plurality of paths at predetermined intervals.

[0046] Furthermore, an anomaly detection method according to an eleventh aspect is the anomaly detection method described in the tenth aspect, wherein each of the plurality of anomaly detection models is assigned a plurality of routes and a route weight, which is a degree indicating the probability that an autonomous vehicle will travel on each of the plurality of routes, the route weights are updated at predetermined intervals, and one of the plurality of routes is assigned as the updated route in each of the plurality of anomaly detection models based on the route weight of each of the plurality of routes.

[0047] According to these methods, one route updated every predetermined period can be estimated as the estimated travel route.

[0048] Furthermore, the anomaly detection method according to the twelfth aspect is the anomaly detection method according to the tenth or eleventh aspect, and further includes a step of learning each of a plurality of anomaly detection models, and in the learning step, for each predetermined period, a past driving route is acquired, which is the driving route in each of a plurality of past drivings of the autonomous vehicle from the departure point to the destination, in which the combination of values ​​of the driving route determination parameters was the same, and from each of the acquired plurality of past driving routes, a frequency with which the autonomous vehicle traveled each of the plurality of routes is calculated, and based on the calculated frequency, one of the plurality of routes is determined as one route in the anomaly detection model corresponding to the driving route determination parameters.

[0049] This allows the route determined by the anomaly detection model to be updated based on multiple past driving records of the autonomous vehicle, making it possible to estimate an estimated driving route that reflects the trends in driving routes selected by the autonomous vehicle in the past.

[0050] Furthermore, an anomaly detection method according to a thirteenth aspect is an anomaly detection method described in any one of the ninth to twelfth aspects, wherein each of the plurality of routes consists of a combination of one or more sections connecting two of a plurality of points arbitrarily determined on the plurality of routes, and in each of the plurality of anomaly detection models, a section weight, which indicates the degree of probability that the autonomous vehicle will travel through each of the one or more sections, is determined for each of the one or more sections, and in the detection step, a degree of anomaly of the travel route up to the arbitrary time point is calculated based on the ratio between the sum of the section weights determined for each of the one or more sections included in the travel route from the departure point to the arbitrary time point and the section weights determined for each of the one or more sections included in the estimated travel route, and if the calculated degree of anomaly exceeds a predetermined threshold, it is detected that an abnormality has occurred in the autonomous vehicle.

[0051] This makes it possible to calculate the degree of abnormality of the driving route for each section included in the driving route of the autonomous vehicle.

[0052] Furthermore, an abnormality detection method according to a fourteenth aspect is the abnormality detection method described in the thirteenth aspect, in which, in the detection step, an event that will cause the autonomous vehicle to decide on an exceptional driving route that is different from normal is determined, and if the driving route determination parameter is a parameter that includes in its combination a value indicating the existence of the event, a correction is made to reduce the degree of abnormality, and if the degree of abnormality after correction exceeds a predetermined threshold value, an abnormality is detected to have occurred in the autonomous vehicle.

[0053] This allows corrections to be made to reduce the degree of abnormality when an autonomous vehicle is operated in an exceptional manner that differs from normal operations, thereby reducing the occurrence of false detections.

[0054] Furthermore, an anomaly detection method according to a fifteenth aspect is an anomaly detection method according to any one of the ninth to fourteenth aspects, in which a plurality of parameter groups are set for classifying any driving route determination parameter into a plurality of parameter groups, one or more driving route determination parameters are set for each of the plurality of parameter groups, and each of the plurality of anomaly detection models corresponds to a respective one of the plurality of parameter groups, and in the detection step, it is determined to which of the plurality of parameter groups the driving route determination parameter is set, an anomaly detection model corresponding to the parameter group in which the driving route determination parameter is set is selected and read out, and one route determined by the read out anomaly detection model is estimated as the estimated driving route.

[0055] This allows the travel route determination parameters to be classified into multiple parameter groups, and one route determined by an anomaly detection model corresponding to each parameter group to be estimated as the estimated travel route, thereby reducing the number of anomaly detection models that need to be prepared.

[0056] Furthermore, an anomaly detection method according to a 16th aspect is an anomaly detection method according to the 15th aspect, wherein an importance level representing the tendency of the values ​​of one or more predetermined driving route determination parameters is assigned to each of the plurality of parameter groups, and the importance level consists of a combination of multiple values ​​corresponding to the combination of values ​​indicated by the driving route determination parameters, and in the detection step, if the driving route determination parameter is not assigned to any of the plurality of parameter groups, an anomaly detection model corresponding to the parameter group whose importance level is most similar to the driving route determination parameter is selected and read out from the plurality of parameter groups.

[0057] This allows the importance to be determined for each parameter group, and even if the combination of values ​​of the travel route determination parameters is new, the travel route determination parameters can be classified into parameter groups.

[0058] Furthermore, the anomaly detection method according to the seventeenth aspect is the anomaly detection method according to the fifteenth or sixteenth aspect, and further includes a step of learning each of the plurality of anomaly detection models, and in the learning step, a plurality of driving route determination parameters obtained in each of a plurality of past drivings of the autonomous vehicle from a departure point to a destination, and a plurality of past driving routes which are the driving routes of the autonomous vehicle during the plurality of past drivings, are obtained, and a plurality of parameter groups are created by grouping the plurality of driving route determination parameters determined in each of one or more past drivings in which the plurality of past driving routes are similar into the same parameter group.

[0059] This allows the driving route determination parameters that resulted in similar driving routes from multiple past driving trips to be classified into the same parameter group, thereby allowing the driving route determination parameters to be classified into each parameter group more accurately.

[0060] Furthermore, an anomaly detection method according to an 18th aspect is the anomaly detection method according to the 16th aspect, further including a step of learning each of a plurality of anomaly detection models, and in the learning step, in each of a plurality of parameter groups, when a value indicating the presence of one event among a plurality of events included in one or more driving route determination parameters characterizes the parameter group, the importance value corresponding to the value indicating the presence or absence of the event is determined to be higher than in other cases, and when a value indicating the absence of one event among a plurality of events included in one or more driving route determination parameters characterizes the parameter group, the importance value corresponding to the value indicating the presence or absence of the event is determined to be lower than in other cases.

[0061] This allows the characteristics or tendencies of one or more driving route determination parameters contained in each parameter group to be reflected in the importance, thereby enabling the driving route determination parameters to be classified into each parameter group with greater accuracy.

[0062] Furthermore, an anomaly detection method according to a 19th aspect is the anomaly detection method described in the 17th aspect, in which in the learning step, a frequency at which the autonomous vehicle has traveled each of the multiple routes is calculated using each of multiple similar past driving routes corresponding to one of the grouped parameter groups, and a route weight for each of the multiple routes is determined based on the calculated frequency, thereby further creating an anomaly detection model.

[0063] This allows an anomaly detection model to be created based on multiple past driving records of an autonomous vehicle, making it possible to estimate an estimated driving route that reflects past history information.

[0064] In addition, an anomaly detection method according to a twentieth aspect of the present disclosure is an anomaly detection device that detects an anomaly in an autonomous vehicle that drives autonomously from a departure point to a destination, and includes an acquisition unit that acquires the driving route of the autonomous vehicle from leaving the departure point to an arbitrary point in time, a memory unit that stores information about multiple routes that the autonomous vehicle may take from the departure point to the destination, and an anomaly detection unit that estimates one of the multiple routes as an estimated driving route from the departure point of the autonomous vehicle to the destination based on the information about the multiple routes read from the memory unit, calculates an anomaly level that indicates how much the driving route up to the arbitrary point in time deviates from the estimated driving route, and detects that an anomaly has occurred in the autonomous vehicle if the calculated anomaly level exceeds a predetermined threshold.

[0065] According to this, since the estimated driving route of the autonomous vehicle is estimated, anomalies can be detected based on the estimated driving route even if the position information of the autonomous vehicle has been tampered with or if unauthorized operations have been performed. In other words, if the route that the autonomous vehicle has traveled up to a given point in time deviates from the estimated driving route, it is possible to detect that an anomaly has occurred.

[0066] A program according to a twenty-first aspect of the present disclosure is a program for causing a computer to execute the anomaly detection method according to any one of the first to nineteenth aspects.

[0067] According to this, since the estimated driving route of the autonomous vehicle is estimated, anomalies can be detected based on the estimated driving route even if the position information of the autonomous vehicle has been tampered with or if unauthorized operations have been performed. In other words, if the route that the autonomous vehicle has traveled up to a given point in time deviates from the estimated driving route, it is possible to detect that an anomaly has occurred.

[0068] The embodiments described below each illustrate a specific example of the present disclosure. The numerical values, shapes, components, component arrangements, steps, and step order shown in the following embodiments are merely examples and are not intended to limit the present disclosure. Furthermore, among the components in the following embodiments, components that are not recited in independent claims are described as optional components. Furthermore, in all embodiments, some or all of the contents of each can be combined.

[0069] (Embodiment) An abnormality detection device according to an embodiment will be described below.

[0070] The anomaly detection device detects abnormalities in driving routes by inputting past driving conditions, current driving conditions, and driving route determination parameters that are external factors taken into account when determining a driving route, such as current weather and road traffic information, to determine whether the current driving position is abnormal due to an attacker or a system abnormality, and presents this information to the monitor.

[0071] <Configuration of anomaly detection system> Fig. 3 is a diagram showing the configuration of an anomaly detection system according to an embodiment. In Fig. 3, the anomaly detection system is composed of a monitored object 1000 (autonomous vehicle), a monitoring center 5000, an external information source 3000, and a network 4000.

[0072] The monitoring center 5000 is an on-premises analysis environment including an anomaly detection device 2000 and a remote monitoring device 5100. The monitoring center 5000 acquires information exchanged in the autonomous vehicle network 1100 within the monitored object 1000 located in a remote location via the network 4000, and performs remote monitoring of the monitored object 1000, performs analysis such as anomaly detection based on the acquired information, and stores the information. The monitoring center 5000 may also include a monitor 5200 that controls the monitored object 1000 based on the results of the remote monitoring and anomaly detection of the monitored object 1000. At least a portion of the configuration of the anomaly detection device 2000 and the remote monitoring device 5100 included in the monitoring center 5000 may reside on the cloud. The anomaly detection device 2000 may also reside outside the monitoring center 5000.

[0073] The anomaly detection device 2000 acquires information related to the traveling of the monitored object 1000 from the autonomous vehicle network 1100 within the monitored object 1000 located in a remote location via the network 4000 and the monitoring center 5000, and performs analysis of the acquired information, such as anomaly detection, and stores the information. Note that the anomaly detection device 2000 may also acquire information related to the traveling of the monitored object 1000 directly from the monitored object 1000 via the network 4000.

[0074] The remote monitoring device 5100 acquires information related to the traveling of the monitored object 1000, which is exchanged in the autonomous vehicle network 1100 within the monitored object 1000, via the network 4000 and the monitoring center 5000, and monitors the state of the monitored object 1000 based on the acquired information. Note that the remote monitoring device 5100 may transmit information related to the control of the monitored object 1000 to the monitored object 1000 via the network 4000. In other words, the remote monitoring device 5100 may remotely control the monitored object 1000.

[0075] The monitored object 1000 is an autonomous vehicle, such as a self-driving vehicle, a goods transport vehicle, a cleaning robot, or a security robot. The monitored object 1000 first determines a driving route from a departure point where the monitored object 1000 is located to a set destination using map information of the target area and location, and then drives autonomously along the determined driving route. In this embodiment, the monitored object 1000 is described as a goods transport vehicle, which is an autonomous vehicle used in a goods transport service.

[0076] FIG. 4 is an example of a graph representing map information used by the monitoring target 1000 when determining a driving route. FIG. 4 uses circular nodes and solid-line edges to facilitate handling of the map information in the anomaly detection device 2000. In FIG. 4, the circular node represented by S1 indicates the starting point of the monitoring target 1000, the circular node represented by G1 indicates the destination of the monitoring target 1000, and the circular nodes represented by W1 to W4 indicate intermediate points corresponding to each point previously set in the map information. The intermediate points may be set, for example, at intersections in the map information, or at any point on the map that the autonomous vehicle can reach. Note that the starting point and the destination may be one of the intermediate points, or may be a point separate from the intermediate point, such as an address. The number of destinations is not limited to one, and multiple destinations may be set. Furthermore, the solid-line edges represented by E1 to E9 indicate sections, each of which is a route that the monitoring target 1000 can actually travel among the routes connecting each intermediate point, the starting point, and the destination. Note that each edge is not limited to a straight line, and may include a curved line to correspond to the actual map information. In this embodiment, the number of edges between each node is one or less. In other words, there is only one route connecting each intermediate point and the starting point and destination.

[0077] In FIG. 4 , the monitored object 1000 determines a travel route, for example, traveling from a starting point S1 along route E1, passing through intermediate point W2, and then traveling along route E8 to reach destination G1, and performs autonomous travel. Note that in FIG. 4 , each intermediate point, starting point, and destination are represented by a circular node, and each route is represented by an edge of a circle. However, the map information used by the monitored object 1000 when determining the travel route is not limited to this and may be represented by other methods, such as an actual map or an aerial photograph. The map information may also include the distance of each route and the average time required for the monitored object 1000 to travel each route. The map information used in the following embodiments is the map information shown in FIG. 4 , and each point and each route in the embodiments are assumed to be each point and each route included in the map information in FIG. 4 . The map information shown in FIG. 4 is stored in the data storage unit 2300, but may also be acquired from an external information source 3000 by the external information acquisition unit 2120 (described later in FIG. 5 ).

[0078] The monitored object 1000 has an autonomous vehicle network 1100, and transmits information related to the driving of the monitored object 1000 exchanged in the autonomous vehicle network 1100 to the anomaly detection device 2000 via the network 4000. Note that the monitored object 1000 is not limited to one autonomous vehicle, but may be multiple autonomous vehicles. When there are multiple monitored objects 1000, each of the autonomous vehicles transmits information about itself to the anomaly detection device 2000 via the network 4000.

[0079] The network 4000 is, for example, a Virtual Private Network (VPN) communication using a general internet line, and securely transmits information about the monitoring target 1000 to the anomaly detection device 2000. Note that the network 4000 may be configured from multiple networks.

[0080] The external information source 3000 provides the anomaly detection device 2000 with various information related to vehicle operation, such as weather information, delivery information, and road traffic information, via the network 4000. The external information source 3000 includes, for example, a weather information server 3200 (an external service) that handles weather information, a road traffic information server 3300 (an external service) that handles road traffic information, a delivery management server 3100 that manages information on goods transported by goods delivery vehicles, and an operation management server 3400 that manages the operation of the autonomous vehicle. Depending on the usage and service form of the autonomous vehicle, the external information source 3000 may include servers or information sources that manage or provide information other than those described above. For example, if the monitored object 1000 is a security robot that guards a specific facility or location, the external information source 3000 may include a server that manages or provides information on the facility, information on the density of people, and the like.

[0081] <Configuration of Each Unit of Anomaly Detection Device 2000 > The anomaly detection device 2000 is configured from an anomaly detection unit 2100 , an output unit 2200 , a data storage unit 2300 , an estimation model storage unit 2400 , an anomaly detection model storage unit 2500 , and a model learning unit 2600 .

[0082] The anomaly detection unit 2100 detects anomalies based on information acquired from the monitored object 1000 .

[0083] The output unit 2200 includes a display means such as a UI (User Interface), and displays visualized information obtained by mapping the results of the abnormality detection by the abnormality detection unit 2100 onto map information or the like to the monitor 5200. When an abnormality is detected in the travel route of the monitoring target 1000, the output unit 2200 may output a control instruction or notification for the monitoring target 1000 to the remote monitoring device 5100. For example, the output unit 2200 may output a control instruction to make an emergency stop of the monitoring target 1000 to the remote monitoring device 5100. In this embodiment, the output unit 2200 corresponds to a display unit.

[0084] The data storage unit 2300 saves and accumulates data relating to the travel of the monitored object 1000, which is acquired by the anomaly detection unit 2100. The data storage unit 2300 may save map information, or may save information relating to the determination of the travel route of the monitored object 1000, including information relating to the state of each route in the map information, which is acquired by the anomaly detection unit 2100. In this embodiment, the data storage unit 2300 corresponds to a memory unit.

[0085] The estimation model storage unit 2400 stores an estimation model and parameter information that defines the types of travel route determination parameters that make up a combination of travel route determination parameters included in each travel route determination parameter group in the estimation model. The estimation model, the travel route determination parameter group, and the travel route determination parameters will be described later.

[0086] The anomaly detection model storage unit 2500 stores an anomaly detection model corresponding to each of the travel route determination parameter groups. The anomaly detection models will be described later. In this embodiment, the estimation model storage unit 2400 and the anomaly detection model storage unit 2500 also correspond to storage units.

[0087] The model learning unit 2600 performs a model learning process based on data related to the travel of the monitored object 1000 stored in the data storage unit 2300, and learns an estimation model and an anomaly detection model. The configuration of the model learning unit 2600 and the model learning process will be described later.

[0088] 5 is a diagram showing the configuration of the anomaly detection unit 2100 included in the anomaly detection device 2000 according to the embodiment. In FIG. 5, the anomaly detection unit 2100 includes a vehicle data acquisition unit 2110, an external information acquisition unit 2120, a travel route determination parameter estimation unit 2130, an anomaly detection model selection unit 2140, an anomaly determination unit 2150, and a display control unit 2160.

[0089] The vehicle data acquisition unit 2110 receives vehicle information related to the current travel of the monitored object 1000 from the monitored object 1000 via the network 4000. The vehicle information is information generated based on vehicle control signals, which are signals exchanged in the autonomous vehicle network 1100 of the monitored object 1000. The vehicle control signals referred to here refer to signals used to control the operation and behavior of the vehicle, such as the engine, brake, accelerator, steering, etc. The received vehicle information is also transmitted to and stored in the data storage unit 2300.

[0090] FIG. 6 is a diagram showing an example of the data structure of the vehicle information of the monitoring target 1000 received by the vehicle data acquisition unit 2110. As shown in FIG.

[0091] As shown in FIG. 6 , the generated vehicle information includes a timestamp indicating the date and time, a vehicle ID that is an identifier for the monitored object 1000, a traveling speed and angular velocity that are information indicating the traveling state of the monitored object 1000, location information for the monitored object 1000, passed intermediate points that indicate intermediate points passed along the way so far during the traveling, and the starting point and destination of the monitored object 1000. For example, the monitored object 1000 indicated by the vehicle ID "V001" indicates that the starting point is "S1" and the destination is "G1," and on the date and time indicated by the timestamp "1612304523," the monitored object 1000 traveled at a point at latitude "32.123..." and longitude "131.532..." at a traveling speed of "3.423..." and an angular velocity of "0.23...," indicating that the monitored object 1000 passed intermediate point W2 so far during the traveling. Note that the vehicle information is not limited to the information shown in FIG. 6 and may include information other than the above. For example, the vehicle information may include information relating to device operation while driving, such as information on the use of turn signals and hazard lights.

[0092] The vehicle data acquisition unit 2110 may receive a vehicle control signal from the monitored object 1000 and generate vehicle information related to driving by analyzing the received vehicle control signal. That is, the vehicle data acquisition unit 2110 receives the vehicle control signal in a data format of a protocol such as CAN (Controller Area Network) or FlexRay, or in a format such as ROS, and generates vehicle information through analysis and data processing. In this case, the vehicle information is generated each time the vehicle data acquisition unit 2110 receives and analyzes a vehicle control signal.

[0093] The external information acquisition unit 2120 acquires external information related to determining a driving route from the external information source 3000. The external information includes, for example, information about the transported goods, information about the status of each route in the map information such as weather information and road traffic information, and operation and management information for the autonomous vehicle. The received external information is also transmitted to and stored in the data storage unit 2300. In this embodiment, the vehicle data acquisition unit 2110 and the external information acquisition unit 2120 correspond to acquisition units.

[0094] 7 to 9 are diagrams showing examples of external information related to determining a driving route, which the external information acquisition unit 2120 acquires from the external information source 3000. FIG.

[0095] 7 is a diagram showing an example of information acquired by the external information acquisition unit 2120 regarding an item transported by the monitored object 1000 when the monitored object 1000 is an autonomous vehicle for transporting goods. Information regarding the item is stored in the delivery management server 3100, which is one of the external information sources 3000, by, for example, a goods transport service provider or a system at the shipping location of the item, and the external information acquisition unit 2120 acquires the information regarding the item from the delivery management server 3100.

[0096] In the example shown in FIG. 7 , information about the transported items includes a transported item ID, which is an identifier assigned to each transported item; information indicating the type and transport conditions of the transported item (whether the transported item is perishable or fragile); information indicating whether the transport condition requires punctuality; the weight of the transported item; the shipping location; and the destination (i.e., the destination). The perishable item, fragile item, and punctuality items can take values ​​of 0 or 1. That is, the perishable item and fragile item items are 1 if the transported item is perishable or fragile, respectively, and 0 if not. The punctuality item is 1 if the transport condition requires punctuality and 0 if no punctuality transport condition is specified. In FIG. 7 , for example, a transported item with a transported item ID of "001" indicates that the type of the transported item is perishable, the transport condition requires punctuality, the weight is "3 kg," the shipping location is "S1," and the destination (i.e., G1). Note that the information about the transported item is not limited to the information shown in FIG. 7 and may include information other than those described above. For example, the type of the transported item may include information indicating whether it is moisture-sensitive, chilled, or frozen, and a delivery time may be specified as a transport condition.

[0097] 8 is a diagram showing an example of weather information acquired by the external information acquisition unit 2120. The external information acquisition unit 2120 acquires weather information from, for example, a weather information server 3200, which is one of the external information sources 3000 and is an external service that provides weather information.

[0098] In the example shown in FIG. 8 , the weather information is based on a weather forecast and is composed of date and time, weather, expected precipitation, and area information. In FIG. 8 , for example, the weather forecast for area "Area A" on date and time "2022-08-01-AM" is "fine" and the expected precipitation is "0 mm." Note that the external information acquisition unit 2120 normally acquires weather information for the area where the monitored object 1000 is located, but may also simultaneously acquire weather information for areas surrounding the area where the monitored object 1000 is located. The weather information is not limited to the information shown in FIG. 8 and may also include information other than the above. For example, it may include information such as wind speed, snowfall, temperature, and humidity based on the weather forecast.

[0099] 9 is a diagram showing an example of road traffic information acquired by the external information acquisition unit 2120. The external information acquisition unit 2120 acquires road traffic information from, for example, a road traffic information server 3300, which is one of the external information sources 3000 and is an external service that provides road traffic information.

[0100] In the example shown in FIG. 9 , the road traffic information is composed of details of events that affect traffic on the road, the date and time of the event, and information about the location or section where the event is occurring. FIG. 9 shows, for example, that the event "construction" occurred in section "E2, E4" on date and time "2022-08-01." The external information acquisition unit 2120 typically acquires road traffic information about sections where the monitored object 1000 can travel, but may also simultaneously acquire road traffic information about other surrounding sections. Note that the road traffic information typically includes only information about events that are currently occurring and does not include information about events that have been resolved. However, this is not limited to this, and information about all events that have occurred may also be included. Furthermore, the road traffic information may also include, for example, information indicating whether travel is possible in the section where the event occurred, information indicating the duration of the event, or information about the location where the event occurred.

[0101] 10 is a diagram showing an example of operation management information that manages the day-to-day operational status of an autonomous vehicle, acquired by the external information acquisition unit 2120. The operation management information is stored in an operation management server 3400, which is one of the external information sources 3000, operated by, for example, a monitor 5200 or a business operator that remotely monitors and detects anomalies in autonomous vehicles, and the external information acquisition unit 2120 acquires the operation information from the operation management server 3400.

[0102] In the example shown in FIG. 10 , the operation management information includes information about sections registered by the monitor 5200 or a business operator performing remote monitoring and anomaly detection of the monitored object 1000 as having an event that could impede the travel of the monitored object 1000, the date and time of the section registration, section details indicating the type of event in the registered section, and information about the time period during which congestion may occur in the section due to the event. In FIG. 10 , for example, the section details for section "E6, E7" are registered as "School Route" on the date and time "2022-08-01," indicating that the congestion time period for the section is "8:00-10:00, 15:00-17:00." Note that the operation management information is not limited to the information shown in FIG. 10 and may also include information other than the above. For example, the operation management information may include route information about the characteristics of sections on which the monitored object 1000 may travel, information such as near-miss points, and information indicating the duration of the section registration or event. It may also include information indicating whether each section is drivable during a specific time period.

[0103] 4 from the external information source 3000, or may acquire information such as the distance of each route in the map information and the average time required for the monitored object 1000 to travel each route. Alternatively, the external information acquisition unit 2120 may directly acquire the external information in the form of a combination of travel route determination parameters, which will be described later.

[0104] The above is an example of external information related to determining a travel route when the monitored object 1000 is a goods transport vehicle in this embodiment. However, if the monitored object 1000 is an autonomous vehicle used for a service other than a goods transport vehicle, the external information related to determining a travel route may differ from the above. For example, if the monitored object 1000 is a security robot that guards a specific facility or location, the external information does not include information about the transported goods. Instead, the external information may include information about the level of VIPs visiting, information about the rooms being used, or information about the density of people in cooperation with an indoor Local Positioning System (LPS), etc., for determining a priority patrol route. In this way, the external information related to determining a travel route may vary depending on the use of the monitored object 1000.

[0105] Returning to the description of the configuration of the anomaly detection unit 2100, the following will be described.

[0106] The travel route determination parameter estimation unit 2130 acquires vehicle information of the monitored object 1000 from the vehicle data acquisition unit 2110, and also acquires various external information related to determining a travel route from the external information acquisition unit 2120. Furthermore, it acquires grouping information of combinations of travel route determination parameters related to determining a travel route from the estimation model storage unit 2400. Then, based on the acquired vehicle information and various external information, it determines a combination of travel route determination parameters for the current travel of the monitored object 1000. The travel route determination parameters and combinations of travel route determination parameters will be described later. In this embodiment, the current travel will be described hereinafter as referring to a series of travels from departing from a preset departure point to arriving at all of the preset destinations.

[0107] The anomaly detection model selection unit 2140 acquires a combination of travel route determination parameters that affect the current travel route determination from the travel route determination parameter estimation unit 2130. It also acquires an estimation model from the estimation model storage unit 2400. Then, based on the combination of travel route determination parameters, it performs an anomaly detection model selection process (described later) to select an appropriate anomaly detection model for detecting an anomaly on the current travel route, and acquires the selected model from the anomaly detection model storage unit 2500.

[0108] The abnormality determination unit 2150 performs an abnormality determination process on the current vehicle information of the monitored object 1000 acquired by the vehicle data acquisition unit 2110, using the abnormality detection model selected and acquired by the abnormality detection model selection unit 2140. In the abnormality determination process, the current driving is compared with information on a frequently used driving route linked to a combination of driving route determination parameters included in the abnormality detection model, i.e., an estimated driving route, to determine an abnormality. Details of the abnormality determination process will be described later.

[0109] When the abnormality determination unit 2150 determines that there is an abnormality in the determination result as a result of the abnormality determination process, the display control unit 2160 transmits the determination result to the output unit 2200 and causes it to be displayed, in order to notify the monitor 5200 that there is a possibility that the monitoring target 1000 is following an abnormal route. Map information may be transmitted together with the determination result.

[0110] <Operation of Abnormality Detection Processing> Next, a description will be given of the abnormality detection processing executed by the abnormality detection device 2000. Fig. 11 is a flowchart showing an example of the operation of the abnormality detection processing according to the embodiment.

[0111] In the anomaly detection process, first, the driving route determination parameter estimation unit 2130 acquires vehicle information of the monitored object 1000 from the vehicle data acquisition unit 2110, and acquires various external information related to determining a driving route from the external information acquisition unit 2120. It also acquires parameter information from the estimation model storage unit 2400. The parameter information is information that defines the types of driving route determination parameters included in the combination of driving route determination parameters. Then, based on the acquired parameter information, the combination of driving route determination parameters for the current driving of the monitored object 1000 that is the target of anomaly detection is determined from the vehicle information and various external information (S7100).

[0112] Here, the travel route determination parameter is an index representing the presence or absence of each of multiple factors indicated by various external information that may affect the determination of the travel route of the monitored object 1000, and can take a value of 0 or 1. 0 represents the absence of the factor, and 1 represents the presence of the factor. Factors that may affect the determination of the travel route of the monitored object 1000 include, for example, the fact that the goods being transported by the monitored object 1000 are fragile, the weather is rainy, and the expected congestion in the section.

[0113] Furthermore, the combination of travel route determination parameters includes each set of travel route determination parameters and information about the departure point and destination included in the vehicle information of the monitored object 1000. An example of the combination of travel route determination parameters determined by the travel route determination parameter estimation unit 2130 in step S7100 is shown in FIG.

[0114] FIG. 12 shows a combination of travel route determination parameters when the monitoring target 1000 shown in FIG. 6 and represented by the vehicle ID "V001" transports the transported item shown in FIG. 7 and represented by the transported item ID "001" on the date and time "2022-08-01-AM." In the example shown in FIG. 12, the shipping location of the transported item is the departure point of the monitoring target 1000, and the destination of the transported item is the destination of the monitoring target 1000. Here, information that the monitoring target 1000 represented by the vehicle ID "V001" transports the transported item represented by the transported item ID "001" is input to the external information source 3000 or the monitoring target 1000 by, for example, the monitor 5200 or the business operator loading the transported item onto the monitoring target 1000, and the travel route determination parameter estimator 2130 determines a combination of travel route determination parameters based on the information of the transported item associated with each monitoring target 1000.

[0115] In Figure 12, the combination of travel route determination parameters is composed of vehicle information and travel route determination parameters assigned based on the respective contents of information related to the transported goods, weather information, road traffic information, and operation management information. The vehicle information shows the departure point and destination of the monitored object 1000, whose vehicle ID is "V001" shown in Figure 6. In the information related to the transported goods, the values ​​of the items "perishable goods," "fragile goods," and "punctuality" for the transported goods ID is "001" shown in Figure 7 are shown as travel route determination parameters. In addition, the "cargo presence" item shows a value indicating whether the monitored object 1000 is transporting the transported goods as a travel route determination parameter. Here, if the monitored object 1000 is transporting the transported goods, the travel route determination parameter for the cargo presence item is 1, and if the monitored object 1000 is not transporting the transported goods, the travel route determination parameter for the cargo presence item is 0. For example, when the monitoring target 1000 is returning home after delivering goods or traveling from the garage to a collection point (shipping location), the monitoring target 1000 is not carrying any goods, and therefore the travel route determination parameter is 0. In the weather information, the weather corresponding to the time when the monitoring target 1000 is traveling, as shown in FIG. 8, is indicated as a travel route determination parameter. Here, the travel route determination parameter is 1 when the weather is "clear." In the road traffic information and operation management information, information indicating whether each section is a congested section where congestion is expected, based on the information about each section shown in FIGS. 9 and 10, is indicated as a travel route determination parameter. Here, if each route is a congested section on the date and time when the monitoring target 1000 travels each route, the travel route determination parameter is 1, and if the route is not a congested section, the travel route determination parameter is 0. Note that the method for determining the travel route determination parameter and the combination of travel route determination parameters are not limited to those described above. For example, the travel route determination parameter may include a parameter indicating a section where the monitoring target 1000 is temporarily unable to travel, a parameter indicating that the section is a school route, a parameter based on precipitation, etc.In addition, when the external information acquisition unit 2120 acquires a combination of driving route determination parameters as external information, it determines the acquired combination of driving route determination parameters as the combination of driving route determination parameters for the current driving of the monitored object 1000.

[0116] 11 , next, in step S7100, the anomaly detection model selection unit 2140 acquires the combination of travel route determination parameters determined by the travel route determination parameter estimation unit 2130. The anomaly detection model selection unit 2140 also acquires an estimation model from the estimation model storage unit 2400. The anomaly detection model selection unit 2140 then executes an anomaly detection model selection process to select, from the estimation models, a travel route determination parameter group according to the content of the combination of travel route determination parameters determined in step S7100, and acquires an anomaly detection model corresponding to the selected travel route determination parameter group from the anomaly detection model storage unit 2500 (S7200).

[0117] Here, the estimation model stored in the estimation model storage unit 2400 will be described. Figures 13A and 13B are diagrams showing examples of the estimation model. In Figures 13A and 13B, the estimation model is composed of multiple travel route determination parameter groups, such as travel route determination parameter group A (Figure 13A) and travel route determination parameter group B (Figure 13B). The estimation model also includes third and subsequent travel route determination parameter groups, such as travel route determination parameter group C (not shown). One travel route determination parameter group includes a combination of one or more travel route determination parameters and travel route determination parameter importance.

[0118] A combination of one or more driving route determination parameters included in one driving route determination parameter group has at least the same starting point and destination, but the combined driving route determination parameters are different. Each combination of driving route determination parameters is grouped into one of the driving route determination parameter groups by a driving route determination parameter grouping process described below. When the monitored object 1000 determines driving routes based on combinations of driving route determination parameters grouped into the same driving route determination parameter group, the resulting driving routes are similar. Although only a portion of the driving route determination parameter group with a starting point S1 and a destination G1 is shown in FIGS. 13A and 13B , the estimation model may include driving route determination parameter groups in which other starting points or other destinations are set.

[0119] The importance of each travel route determination parameter indicates the degree of influence that each travel route determination parameter has on the travel route determination of the monitored object 1000. A travel route parameter with a high importance indicates that it has a significant influence on the travel route determination for the group. The parameter importance calculation process for determining the importance will be described later. The importance of a travel route determination parameter with a large number of common values ​​of "1" across multiple combinations within the same group is calculated to be high. In addition, a travel route determination parameter with a high percentage of values ​​of "1" in other groups is calculated to be lower in importance than a travel route determination parameter with a low percentage of values ​​of "1" in other groups. Conversely, a travel route determination parameter with a low percentage of values ​​of "1" in other groups is calculated to be higher in importance than a travel route determination parameter with a high percentage of values ​​of "1" in other groups. In the travel route determination parameter group A in FIG. 13A , the importance of the travel route determination parameter "cargo presence / absence" is 0.5, the importance of "perishable goods" is 0.9, and the importance of "punctuality" is 0.8. The importance of these three travel route determination parameters is higher than the importance of the other travel route determination parameters in the group. This is because many of these travel route determination parameters have a common value of "1" in the group. Furthermore, the importance of the travel route determination parameters "perishable goods" and "punctuality" is higher than the importance of "cargo presence / absence" because the proportion of the values ​​of "1" in other groups is low. In other words, the importance within a group is a representative value that represents the tendency or characteristics of the ease of combination of the multiple travel route determination parameters included in the travel route determination parameter group.

[0120] Next, the anomaly detection models stored in the anomaly detection model storage unit 2500 will be described. FIGS. 14A to 14C are diagrams showing examples of anomaly detection models. In FIGS. 14A to 14C, one anomaly detection model is associated with each travel route determination parameter group. The anomaly detection model is composed of a departure point, a destination, and a weight for each route. The weight for each route is a degree based on the probability that each route will be selected when the monitored object 1000 determines a travel route based on a combination of each travel route determination parameter in the group. A small weight is set for a route that has a high probability of being selected, and a large weight is set for a route that has a low probability of being selected.

[0121] The weight of each route is determined by the anomaly detection model learning process described below. In Fig. 14A, in the anomaly detection model corresponding to travel route determination parameter group A, the weight of route "E1" is 2, and the weight of route "E2" is 6. This indicates that route E1 is more likely to be selected than route E2 when determining a travel route.

[0122] The anomaly detection model also includes an average model for each combination of departure point and destination. The average model is an anomaly detection model formed by the average of all anomaly detection models associated with all driving route determination parameter groups having the same combination of departure point and destination. Figure 14C shows an average model formed by the average of all anomaly detection models for which the departure point is S1 and the destination is G1.

[0123] Next, the anomaly detection model selection process will be described.

[0124] The anomaly detection model selection process is executed by the anomaly detection model selection unit 2140, and is a process of selecting an anomaly detection model that is suitable for the combination of travel route determination parameters determined by the travel route determination parameter estimation unit 2130, i.e., the combination of travel route determination parameters related to route determination for the current travel, from the anomaly detection models stored in the anomaly detection model storage unit 2500. Figure 15 shows a flowchart of an example of the operation of the anomaly detection model selection process.

[0125] In the anomaly detection model selection process, first, it is determined whether any of the multiple travel route determination parameter groups included in the estimation model stored in the estimation model storage unit 2400 includes a travel route determination parameter combination identical to the current travel route determination parameter combination (S7210). For example, in this embodiment, if the current travel route determination parameter combination is one of the combinations shown in Figure 12, it is determined whether that combination exists in any of the travel route determination parameter groups. If it is determined that the same combination exists (YES in step S7210), proceed to step S7220. If it is determined that the same combination does not exist (NO in step S7210), proceed to step S7230.

[0126] The anomaly detection model selection unit 2140 acquires, from the anomaly detection model storage unit 2500, an anomaly detection model associated with a travel route determination parameter group that includes the same travel route determination parameter combination as the current travel route determination parameter combination (S7220), and then terminates the anomaly detection model selection process.

[0127] Alternatively, to determine which of the travel route determination parameter groups the current travel route determination parameter combination is similar to, the anomaly detection model selection unit 2140 calculates the similarity between the current travel route determination parameter combination and the travel route determination parameter importance of each travel route determination parameter group included in the estimation model (S7230). Note that the similarity is calculated here with travel route determination parameter groups that have the same combination of origin and destination included in the current travel route determination parameter combination. For example, a method using cosine similarity is used to calculate the similarity. In the following explanation of how to calculate the similarity, for simplicity, the similarity for group A and group B is calculated using only the travel route determination parameter combinations and travel route determination parameter importance for four categories: "presence / absence of cargo," "perishable goods," "fragile goods," and "punctuality." However, in reality, the similarity for all groups is calculated based on the combinations and importance of all travel route determination parameters.

[0128] For example, in Figures 13A and 13B, the importance of four travel route determination parameters in travel route determination parameter group A, namely, "Cargo," "Perishable," "Fragile," and "Punctual," is [0.5, 0.9, 0.1, 0.8], and the importance of these four travel route determination parameters in travel route determination parameter group B is [0.5, 0.1, 0.9, 0.1]. In this case, if the current combination of four travel route determination parameters is [1, 1, 0, 1] as shown in Figure 12, the similarities between the current travel route determination parameter set and groups A and B are calculated using cosine similarity, resulting in similarity to group A: 0.97 > similarity to group B: 0.39. Therefore, in the above example, it can be determined that the current travel route determination parameter combination is more similar to travel route determination parameter group A, which has a higher similarity.

[0129] Next, it is determined whether any of the similarities calculated in step S7230 with each driving route determination parameter group having the same combination of departure point and destination exceeds a preset threshold (S7240). If it is determined that any of the similarities exceeds the threshold (YES in step S7240), the process proceeds to step S7250, and if it is determined that any of the similarities does not exceed the threshold (NO in step S7240), the process proceeds to step S7260. The threshold is set to any value between 0 and 1, for example, 0.7. The threshold is set by the monitor 5200, for example.

[0130] The anomaly detection model selection unit 2140 acquires from the anomaly detection model storage unit 2500 the anomaly detection model associated with the travel route determination parameter group that has the highest similarity to the current travel route determination parameter combination (S7250). For example, if the similarity to group A is the highest, the anomaly detection model associated with group A is acquired. Then, the anomaly detection model selection process ends.

[0131] Alternatively, since there is no group among the driving route determination parameter groups stored in the anomaly detection model storage unit 2500 that is strongly similar to the current driving route determination parameter combination so as to exceed the threshold, the anomaly detection model selection unit 2140 acquires from the anomaly detection model storage unit 2500 an average model corresponding to the combination of departure point and destination included in the current driving route determination parameter combination (S7260).Then, the anomaly detection model selection process ends.

[0132] Returning to the explanation of the anomaly detection process.

[0133] The abnormality determination unit 2150 performs an abnormality determination process on the current vehicle information of the monitoring target 1000 acquired by the vehicle data acquisition unit 2110 using the abnormality detection model selected and acquired by the abnormality detection model selection unit 2140 (S7300). Furthermore, the abnormality determination result and map information are transmitted to and output by the output unit 2200 via the display control unit 2160. Then, the abnormality detection process ends.

[0134] The abnormality determination process will now be described. The abnormality determination process is a process executed by the abnormality determination unit 2150, and is a process for determining the abnormality (degree of abnormality) of the current travel of the monitoring target 1000 based on the weight of each route included in the abnormality detection model. In this embodiment, for example, the abnormality detection model acquired by the abnormality detection model selection unit 2140 is an abnormality detection model corresponding to travel route determination parameter group A, and the vehicle information acquired by the vehicle data acquisition unit 2110 is vehicle information of the monitoring target 1000 whose vehicle ID is "V001" shown in FIG. 6 . FIG. 16 is a flowchart showing an example of the operation of the abnormality determination process. Also, an example of the process of the abnormality determination process is shown in FIGS. 17A and 17B.

[0135] The anomaly determination unit 2150 acquires the current vehicle information of the monitoring target 1000 acquired by the vehicle data acquisition unit 2110. Next, based on the information on the passing waypoints included in the vehicle information, the route from the departure point that the monitoring target 1000 has traveled up to now is identified. Then, for the identified route, the weight of each route included in the anomaly detection model is used to calculate the sum of the weights of each route that has been traveled up to now (S7310). For example, as shown in FIG. 17B , the passing waypoint of the monitoring target 1000 at the date and time represented by the timestamp "1612304523" is "W2." Because the departure point is "S1," it is identified that only "E1" has been traveled up to now, as indicated by the dashed-dotted arrow in FIG. 17A . Therefore, the sum of the weights of each route that has been traveled up to that date is calculated to be "2" based on the weight value of route E1. Similarly, since the intermediate points passed by the monitoring target 1000 at the date and time indicated by the timestamp "1612304527" are "W2 → W3 → W1," the route traveled up to now is identified as "E1 → E7 → E4," as shown by the straight arrow in Fig. 17A. Therefore, the total weight of each route traveled up to that date is calculated as 2 + 3 + 7 = 12, which is "12."

[0136] The anomaly determination unit 2150 uses the weights of each route included in the anomaly detection model to identify an estimated travel route along which the monitored object 1000 is estimated to travel. The unit 2150 also calculates the sum of the weights for the identified estimated travel route (S7320). The estimated travel route is, for example, a combination of multiple routes from the departure point to the destination, with the smallest sum of the weights for each route. In FIG. 17A , there are multiple combinations of routes from the departure point to the destination, but the combination with the smallest sum of the weights is the route from S1 through E1 to E8 to G1, as indicated by the dotted arrow. In this embodiment, this route is the estimated travel route, and the sum of the weights is calculated as "3" since 2 + 1 = 3. The estimated travel route is identified based on the anomaly detection model. However, since the anomaly detection model is selected based on the combination of travel route parameters determined in step S7100, the estimated travel route can also be said to be identified according to the combination of travel route parameters.

[0137] Next, the abnormality determination unit 2150 compares the total weight of the route traveled up to the present calculated in step S7310 with the total weight of the estimated travel route calculated in step S7320 to calculate the abnormality level (S7330). The abnormality level is the degree to which the route traveled up to the present by the monitoring target 1000 is abnormal. The abnormality level is calculated as the ratio of the total weight of the route traveled up to the present to the total weight of the estimated travel route. In FIG. 17B , the total weight of the route traveled up to the date and time represented by the timestamp "1612304523" is 2, so the abnormality level is calculated as 2 / 3 = 0.666..., or "0.67." Similarly, the total weight of the route traveled up to the date and time represented by the timestamp "1612304527" is 12, so the abnormality level is calculated as 12 / 3 = 4, or "4." The calculation of the degree of abnormality is not limited to the above. For example, if the total weight of the route traveled up to now is greater than the total weight of the estimated driving route, the degree of abnormality may simply be calculated as the difference between the total values ​​of the two weights.

[0138] The anomaly determination unit 2150 determines whether to correct the anomaly level calculated in step S7330 (S7340). For example, when a driving route determination parameter that is a factor in the monitoring target 1000 selecting an exceptional driving route that differs from the normal driving route has been identified, and the driving route determination parameter that is the factor is present among the driving route determination parameters related to the current driving route determination of the monitoring target 1000, the anomaly determination unit 2150 determines to perform correction. Factors that cause exceptional driving route selection include, for example, driving in maintenance mode for operational reasons or driving in an area where manual driving is always required. In such cases, a driving route may be selected based on the maintenance content or the experience of the manual driving driver, resulting in the selection of an unusual driving route. As a result, a deviation from the estimated driving route determined by the learned anomaly detection model may occur, potentially reducing the reliability of the calculated anomaly level. Therefore, correcting the anomaly level to lower it in such cases can prevent overdetection. Additionally, the anomaly determination unit 2150 may determine to perform correction when the anomaly detection model used is of a specific type or when a correction value β for correcting the degree of anomaly is preset in the anomaly determination unit 2150. If it is determined that a correction should be performed on the degree of anomaly because a parameter for which the degree of anomaly should be corrected exists (YES in step S7340), the process proceeds to step S7350, and if it is determined that a correction should not be performed because a parameter for which the degree of anomaly should be corrected does not exist (NO in step S7340), the process proceeds to step S7360.

[0139] The abnormality determination unit 2150 corrects the degree of abnormality calculated in step S7330 using the correction value β (S7350), and then proceeds to step S7360. For example, the abnormality determination unit 2150 may perform a correction such as dividing the degree of abnormality calculated in step S7330 by the correction value β to lower the degree of abnormality, or subtracting the correction value β from the degree of abnormality calculated in step S7330.

[0140] The anomaly determination unit 2150 determines whether the degree of anomaly calculated in step S7330 or the degree of anomaly corrected in step S7350 exceeds a preset threshold θ (S7360). For example, as shown in FIG. 17B , if the threshold θ is set to "θ = 3," and the degree of anomaly is greater than "3," the route traveled up to the present time of the monitored object 1000 is determined to be abnormal. In FIG. 17B , the degree of anomaly of the route traveled up to the date and time represented by the timestamp "1612304523" is 0.67, and therefore the route is determined to be normal. However, the degree of anomaly of the route traveled up to the date and time represented by the timestamp "1612304527" is 4, which exceeds the threshold θ, and therefore the route is determined to be abnormal. Note that, although the threshold θ is "θ = 3" in this embodiment, the threshold θ may be a value other than "θ = 3." Furthermore, the threshold θ may be set for each anomaly detection model, or a uniform value may be set commonly across all the anomaly detection models.

[0141] Furthermore, the abnormality determination unit 2150 transmits the abnormality determination result and the map information to the output unit 2200 via the display control unit 2160 and causes them to be output. Then, the abnormality determination process ends. Note that the abnormality determination unit 2150 may transmit the abnormality determination result to the output unit 2200 and cause it to be output only when the degree of abnormality of the route exceeds the threshold value θ. In this way, it is also possible for the output unit 2200 to not output anything if the degree of abnormality of the route is equal to or less than the threshold value θ.

[0142] The abnormality determination unit 2150 transmits and outputs to the output unit 2200 via the display control unit 2160 a diagram including the abnormality determination result and map information, such as a combination of the contents shown in FIGS. 17A and 17B . Alternatively, for example, information such as each point, each route, and the weight of each route in the map information of FIG. 17A may be mapped onto an actual map and then output to the output unit 2200. This provides information that allows an unusual driving route to be immediately identified visually, helping the monitor 5200 or the like to identify any abnormalities in the current driving of the monitored object 1000 and facilitating subsequent analysis. Furthermore, information about the estimated driving route identified in step S7320 may be mapped together with map information and then output to the output unit 2200.

[0143] In addition to the abnormality determination result, the current vehicle information of the monitored object 1000 acquired by the vehicle data acquisition unit 2110 in step S7310 may be output. For example, data merging the abnormality determination result and the vehicle information acquired by the vehicle data acquisition unit 2110, as shown in FIG. 18, may be output. The data shown in FIG. 18 includes each piece of data included in the current vehicle information of the monitored object 1000 shown in FIG. 6 and the abnormality determination result. The abnormality determination result may include, for example, the abnormality degree calculated in step S7330 or corrected in step S7350, the abnormality determination result determined in step S7360 based on the abnormality degree, and a data series ID indicating that an abnormality was determined using a continuous series. The data series ID is assigned the same value if it is an abnormality determination result based on vehicle information from a continuous series of travels by the monitored object 1000 from the departure point to the destination. For example, in Figure 18, the timestamps from "1612304523" to "1612304527" indicate that the same monitoring target 1000 is moving from the same departure point to the destination, and therefore are the same data series, and the same value (S001) is assigned as the data series ID. On the other hand, the timestamp "1612304530" indicates that the same monitoring target 1000 is moving from the same departure point to a different destination, and therefore is a different data series, and a different value (S002) is assigned as the data series ID. The data series ID is assigned to the abnormality determination result when the abnormality detection process is executed, and is output.

[0144] <Configuration of Model Learning Unit 2600> Next, the configuration of model learning unit 2600 will be described.

[0145] 19 is a diagram showing the configuration of a model learning unit 2600 included in anomaly detection device 2000. In FIG. 19, model learning unit 2600 is composed of a data acquisition unit 2610, a parameter combination grouping unit 2620, a travel route determination parameter importance calculation unit 2630, and an anomaly detection model learning unit 2640.

[0146] The data acquisition unit 2610 acquires from the data storage unit 2300 vehicle information for past travels of the monitored object 1000 and a combination of travel route determination parameters related to determining the travel route during each past travel, based on information for an arbitrarily set learning period.

[0147] The parameter combination grouping unit 2620 acquires data from the data acquisition unit 2610. Next, the parameter combination grouping unit 2620 groups combinations of travel route determination parameters that have resulted in similar travel, based on the acquired vehicle information from past travels and the combinations of travel route determination parameters related to travel route determination for each past travel. The parameter combination grouping unit 2620 then stores the grouping results as an estimation model in the estimation model storage unit 2400.

[0148] The travel route determination parameter importance calculation unit 2630 acquires grouping information of travel route determination parameter combinations from the parameter combination grouping unit 2620. Then, the travel route determination parameter importance calculation unit 2630 calculates, for each travel route determination parameter group, the travel route determination parameter importance, which is the degree to which the group contributes to travel route determination, and adds the calculation result to the estimation model stored in the estimation model storage unit 2400.

[0149] The anomaly detection model learning unit 2640 acquires past driving data and driving route determination parameters related to driving route determination for each driving from the data storage unit 2300. The anomaly detection model learning unit 2640 also acquires driving route determination parameter groups and information indicating the frequency of use of driving routes corresponding to the driving route determination parameter groups from the parameter combination grouping unit 2620. The anomaly detection model learning unit 2640 then learns an anomaly detection model for each group and stores the learned anomaly detection model in the anomaly detection model storage unit 2500.

[0150] <Operation of Model Learning Process> Next, the model learning process will be described.

[0151] FIG. 20 is a flowchart showing an example of the operation of the model learning process.

[0152] The anomaly detection device 2000 defines travel route determination parameters that are considered to be related to travel route determination from among multiple parameters that may be included in the information acquired from the external information source 3000 (S6100). Specifically, the anomaly detection device 2000 defines parameters that are set in advance by the monitor 5200 as parameters related to travel route determination as travel route determination parameters. Information on the defined travel route determination parameters is stored in the data storage unit 2300, the anomaly detection unit 2100, the model learning unit 2600, etc.

[0153] The parameters related to driving route determination are set after consideration by the monitor 5200 based on the type of service and the content of the service in which the autonomous vehicle, which is the monitoring target 1000, is used. For example, if the monitoring target 1000 is an autonomous vehicle for transporting goods, parameters indicating information about the type of cargo to be transported and information about the weather may be set as parameters related to driving route determination. In addition, parameters indicating information about road conditions and road traffic information may be set as parameters related to driving route determination. Furthermore, in various services using autonomous vehicles, an optimal driving route determination algorithm that does not interfere with operation is generally introduced by the autonomous vehicle service provider, and therefore various parameters may be set as parameters related to driving route determination depending on the case in which the autonomous vehicle is used.

[0154] The data acquisition unit 2610 acquires information about the driving route included in the vehicle information for past driving of the monitoring target 1000 from the data storage unit 2300 (S6200). The information about the driving route includes at least the starting point, each route traveled by the monitoring target 1000, and the destination. The information about the driving route may also include information about each intermediate point passed by the monitoring target 1000. FIG. 21A shows an example of the acquired information about the driving route for past driving. In FIG. 21A, the driving route information is composed of a driving number, which is an identifier assigned to each driving of the monitoring target 1000 from the starting point to the destination, a date and time indicating the start date and time or end date and time of the driving, the starting point, the destination, and information indicating whether or not each route has been driven. The information indicating whether or not each route has been driven takes a value of 0 or 1, indicating that the monitoring target 1000 has driven each route assigned a value of 1 and has not driven each route assigned a value of 0. For example, the driving number may be In the trip indicated by "1", the monitored object 1000 started from "S1", traveled along the route "E3, E4, E5, E9", and arrived at the destination "G1".

[0155] The data acquisition unit 2610 acquires from the data storage unit 2300 a combination of travel route determination parameters related to travel route determination for past travels corresponding to the travel route information acquired in step S6200 (S6300). The combinations of travel route determination parameters for each past travel shown in FIG. 21A are shown in FIG. 21B. In FIG. 21B, the combination of travel route determination parameters is composed of a travel number corresponding to the travel number in the travel route information and each travel route determination parameter. Note that the following description will be given assuming that the travel route determination parameters for each of the congested sections with travel numbers 1 to 4 shown in the figure are the same.

[0156] The parameter combination grouping unit 2620 acquires the information on past travel routes and the corresponding combinations of travel route determination parameters acquired by the data acquisition unit 2610 in steps S6100 and S6200. Then, the parameter combination grouping unit 2620 performs a travel route determination parameter grouping process based on the information on the past travel routes, and performs grouping of the combinations of travel route determination parameters (S6400). As a result, travel route determination parameter groups are created, and each group is linked to the information on the past travel routes.

[0157] 22 is a flowchart showing an example of the operation of the parameter grouping process for determining a driving route, which is performed by the parameter combination grouping unit 2620.

[0158] The parameter combination grouping unit 2620 creates a histogram of the frequency of route use for each past trip based on the information about the past trips (S6410). The histogram for each past trip shown in FIG. 21A is, for example, a histogram with the vertical axis indicating frequency of use, as shown in FIGS. 23A to 23D. Note that for simplicity, FIGS. 23A to 23D only show the combinations of route determination parameters corresponding to the information about the past trips for the following items: "Cargo," "Perishables," "Fragile Items," "Punctual Schedule," and "Weather."

[0159] Next, the parameter combination grouping unit 2620 groups histograms with similar trends based on the histograms representing the frequency of use of each travel route during each past travel (S6420). The grouped histograms are then aggregated into a single histogram. The parameter combination grouping unit 2620 also groups the combinations of travel route determination parameters corresponding to each grouped histogram into respective travel route determination parameter groups, and then terminates the travel route determination parameter grouping process.

[0160] To calculate the degree of similarity between the trends of each histogram, for example, the Earth Mover's Distance (EMD) algorithm can be used to calculate the distance between two histograms. Then, agglomerative clustering is used to sequentially group and aggregate histograms with close EMD distances, i.e., similar histograms, so that histograms with similar trends are aggregated into multiple groups. The aggregated results are then stored in the estimation model storage unit 2400 and transmitted to the anomaly detection model learning unit 2640.

[0161] 24 and 25 show an example of the process of grouping and aggregating the four histograms shown in FIGS. 23A to 23D. First, the parameter combination grouping unit 2620 calculates the similarity between two of the four histograms shown in FIGS. 23A to 23D for all combinations. Next, the parameter combination grouping unit 2620 groups and aggregates the histograms for the pair whose similarity between the histograms is equal to or greater than a threshold and has the highest similarity. In this embodiment, as shown in FIG. 24, the pair whose histograms have travel No. 1 and 4 have the highest similarity, so these are grouped and the histograms are aggregated as group 1. At the same time, combinations of travel route determination parameters are also grouped, and there are two combinations of travel route determination parameters in that group.

[0162] Next, the parameter combination grouping unit 2620 calculates the similarity between two histograms for all combinations: the histogram aggregated into group 1, the ungrouped histogram indicated by the travel number 2, and the histogram indicated by the travel number 3. If there are multiple histogram groups, the similarity between the groups may be calculated, and if the similarity between the groups is equal to or greater than a threshold, the groups may be grouped into one group. The parameter combination grouping unit 2620 then groups and aggregates the histograms for the pair whose similarity between the histograms is equal to or greater than a threshold and has the highest similarity. In this embodiment, as shown in FIG. 25 , the histogram indicated by the travel number 2 and the histogram in group 1 have the highest similarity, so these are grouped. That is, the histogram indicated by the travel number 2 is aggregated into group 1. At the same time, the combinations of travel route determination parameters are also grouped, and there are three combinations of travel route determination parameters in that group.

[0163] The parameter combination grouping unit 2620 sequentially performs similar grouping until the similarity between the histograms becomes less than the threshold. Then, when the calculated similarity becomes less than the threshold, the parameter combination grouping unit 2620 groups the ungrouped histograms into independent groups and terminates grouping. In this embodiment, as shown in FIG. 25 , since the similarity between the histograms aggregated as group 1 and the ungrouped histogram indicated by run No. 3 becomes less than the threshold, the histogram indicated by run No. 3 is not grouped into group 1 but becomes an independent group 2.

[0164] In this way, each group, each of which has a combination of one or more travel route determination parameters, is stored as a travel route determination parameter group in the estimation model storage unit 2400. In addition, an aggregated histogram corresponding to each travel route determination parameter group is transmitted to the anomaly detection model learning unit 2640.

[0165] In this way, by performing the travel route determination parameter grouping process, it is possible to aggregate combinations of travel route determination parameters that use similar travel routes, and therefore it is possible to suppress an increase in the number of anomaly detection models even if the number of travel route determination parameter combinations increases. This makes it possible to suppress problems such as a small amount of data available for learning and the risk of overlearning, compared to when an anomaly detection model is simply selected for each combination of travel route determination parameters.

[0166] Returning to the description of the model learning process.

[0167] The travel route determination parameter importance calculation unit 2630 acquires grouping information of the travel route determination parameter combinations grouped in step S6400 from the parameter combination grouping unit 2620. Next, the travel route determination parameter importance calculation unit 2630 performs a travel route determination parameter importance calculation process to calculate, for each travel route determination parameter group, the travel route determination parameter importance related to travel route determination for each group (S6500). Then, the travel route determination parameter importance calculation unit 2630 adds the calculated travel route determination parameter importance to the travel route determination parameter group, thereby creating an estimation model and storing it in the estimation model storage unit 2400.

[0168] The route determination parameter importance calculation process will be described.

[0169] Fig. 26 is a flowchart showing an example of the operation of a travel route determination parameter importance calculation process executed by the travel route determination parameter importance calculation unit 2630. Also, Figs. 27A and 27B show an example of the process of calculating the travel route determination parameter importance for a travel route determination parameter group including a combination of three travel route determination parameters, which is shown as group 1 in Fig. 25.

[0170] First, the travel route determination parameter importance calculation unit 2630 calculates the frequency with which each travel route determination parameter in each group obtained in step S6400 has a value of 1 (S6510). Furthermore, the travel route determination parameter importance calculation unit 2630 identifies travel route determination parameters that frequently have a value of 1. For example, in the travel route determination parameter group shown in FIG. 27A , the travel route determination parameters that frequently have a value of 1 are identified as "presence / absence of cargo," "fragile item," and "(congested section) E2."

[0171] Next, for each travel route determination parameter, the proportion of travel route determination parameters that are identified as travel route determination parameters that frequently have a value of 1 in other groups is calculated. Then, based on the calculated proportion and the frequency of travel route determination parameters that are 1 calculated in step S6510, the importance of the travel route determination parameter in each group is calculated using, for example, the TF-IDF (Team Frequency-Inverse Document Frequency) method (S6520). That is, a travel route determination parameter that is frequently set to 1 in the group and that is identified as a travel route determination parameter that is frequently set to 1 in a low proportion in other groups can be said to be a travel route determination parameter unique to the group, and is therefore calculated to have a high importance. Furthermore, even if a travel route determination parameter is frequently set to 1 in the group, if the proportion of travel route determination parameters that are identified as a travel route determination parameter that is frequently set to 1 in other groups is high, the importance is calculated to be lower than the above. Furthermore, a travel route determination parameter that is rarely set to 1 in the group is calculated to have an even lower importance than the above. The importance level can take a value ranging from 0 to 1. The number of group-specific travel route determination parameters is not limited to one, and there may be a plurality of travel route determination parameters calculated to have high importance.

[0172] For example, in the travel route determination parameter group shown in Fig. 27A, the importance of the travel route determination parameters "cargo presence / absence," "fragile item," and "(congested section) E2," which have a high frequency of being set to the value 1, is calculated to be high. Furthermore, if the travel route determination parameter for "fragile item" is identified as a travel route determination parameter that has a high frequency of being set to the value 1 in other groups at a low rate, the travel route determination parameter can be said to be a travel route determination parameter unique to the group, and is therefore calculated to have a higher importance of 0.9 as shown in Fig. 27B.

[0173] In this way, by setting the importance of each travel route determination parameter, even when a new travel route determination parameter combination that has not been observed in the past appears in the anomaly detection model selection process, it is possible to calculate a similarity that indicates to which group the new travel route determination parameter combination is similar. Therefore, even for a new travel route determination parameter combination, an appropriate travel route determination parameter loop can be selected.

[0174] Returning to the description of the model learning process.

[0175] The anomaly detection model learning unit 2640 acquires the driving route determination parameter groups grouped in step S6400 from the driving route determination parameter combination grouping unit 2620. The anomaly detection model learning unit 2640 also acquires an aggregated histogram indicating the frequency of use of the driving route corresponding to each driving route determination parameter group. Then, the anomaly detection model learning unit 2640 executes an anomaly detection model learning process (S6600), stores the driving route determination parameter group including the learned anomaly detection model in the anomaly detection model storage unit 2500, and terminates the model learning process.

[0176] The anomaly detection model learning process will be described.

[0177] 28 is a flowchart showing an example of the operation of the anomaly detection model learning process executed by the anomaly detection model learning unit 2640. In the anomaly detection model learning process, learning of an anomaly detection model that determines the validity or abnormality of a driving route is carried out for each driving route determination parameter group.

[0178] One anomaly detection model is associated with each travel route determination parameter group, and the model determines the usage frequency for each of multiple travel routes based on each travel route determination parameter in the travel route determination parameter group. By applying the anomaly detection model in the anomaly detection process, the validity or abnormality of the travel route selected during the travel of the monitored object 1000 can be determined. For example, as shown in FIGS. 14A to 14C , the anomaly detection model includes a departure point, a destination, and weights set based on the usage frequency of each route from the departure point to the destination. An example of the learning process for the anomaly detection model is shown in FIGS. 29A and 29B . FIG. 29 illustrates the learning process for the anomaly detection model associated with the travel route determination parameter group grouped as group 1 in step S6420.

[0179] The anomaly detection model learning unit 2640 acquires a histogram of the travel routes for each travel route determination parameter group aggregated in step S6400 (S6610). For example, a histogram showing the frequency of use of each travel route, as shown in FIG. 29A, is acquired.

[0180] Next, the anomaly detection model learning unit 2640 calculates the usage probability for each route based on the histogram of travel routes acquired in step S6610 (S6620). For example, in FIG. 29B , the usage probability of routes E3 and E9 is calculated to be 1, the usage probability of routes E4 and E5 is 0.67, and the usage probability of route E6 is 0.33. Furthermore, if there is a route that has never been used and has a calculated usage probability of 0, the anomaly detection model learning unit 2640 performs smoothing by adding a small value α to all routes. α is, for example, 0.1 as shown in FIG. 29B .

[0181] The anomaly detection model learning unit 2640 obtains the number of past trips of the monitored object 1000 in which the combination of origin and destination indicated by each travel route determination parameter group matches the combination of origin and destination for each trip. The number of matches indicates the number of times the monitored object 1000 has traveled using the combination of origin and destination indicated by each travel route determination parameter group. The anomaly detection model learning unit 2640 then determines whether the number of trips for each travel route determination parameter group that is the target of the anomaly detection model learning process is equal to or less than a threshold (S6630). If it is determined that the number of trips is not equal to or less than the threshold (NO in step S6630), the process proceeds to step S6640. If it is determined that the number of trips is equal to or less than the threshold (YES in step S6630), the process proceeds to step S6650.

[0182] The anomaly detection model learning unit 2640 calculates the cost of using each route from the utilization probability of each route calculated in step S6620 and sets a weight for each route based on the cost of using each route (S6640). The cost of using each route is calculated, for example, by taking the negative logarithmic likelihood of the utilization probability of each route. Furthermore, the weight of each route is set, for example, based on the ratio of the costs of using each route. In FIG. 29B , the weight of route E1 is set to 9, the weight of route E2 is set to 9, the weight of route E3 is set to 3, and so on. Note that the weight values ​​of each route may be appropriately scaled and set as long as the ratios do not change. For example, scaling the weight values ​​based on the maximum total value of the weights of the routes traveled from a certain departure point to a certain destination can facilitate the setting of a threshold for anomaly detection in step S7360.

[0183] Alternatively, the anomaly detection model learning unit 2640 calculates the cost of using each route based on the value obtained by adding the average value of the usage probabilities of each route in all travel route determination parameter groups in which the departure point and destination match to the usage probability of each route calculated in step S6620, and sets a weight for each route (S6650). This allows for setting weights based on the average value of the usage probabilities for all data, even if the number of trips from the departure point to the destination indicated in the travel route determination parameter group is below a threshold and learning may be insufficient. This makes it possible to prevent a deterioration in anomaly detection accuracy and an increase in overdetection due to insufficient learning.

[0184] The present invention can be widely used in anomaly detection devices that detect anomalies related to the driving of an autonomous vehicle.

[0185] 1000 Monitoring target 1100 Autonomous vehicle network 2000 Anomaly detection device 2100 Anomaly detection unit 2110 Vehicle data acquisition unit 2120 External information acquisition unit 2130 Travel route determination parameter estimation unit 2140 Anomaly detection model selection unit 2150 Anomaly determination unit 2160 Display control unit 2200 Output unit 2300 Data storage unit 2400 Estimation model storage unit 2500 Anomaly detection model storage unit 2600 Model learning unit 2610 Data acquisition unit 2620 Parameter combination grouping unit 2630 Travel route determination parameter importance calculation unit 2640 Anomaly detection model learning unit 3000 External information source 3100 Delivery management server 3200 Weather information server 3300 Road traffic information server 3400 Operation management server 4000 Network 5000 Monitoring center 5100 Remote monitoring device 5200 Monitor

Claims

1. An anomaly detection method executed by a computer for detecting an abnormality in an autonomous vehicle traveling autonomously from a starting point to a destination, comprising the steps of: acquiring a driving route of the autonomous vehicle from departing from the starting point to an arbitrary time point; reading information about a plurality of routes that the autonomous vehicle may travel from the starting point to the destination, estimating one of the plurality of routes as an estimated driving route of the autonomous vehicle from the starting point to the destination based on the read information about the plurality of routes, calculating an abnormality degree which is the degree to which the driving route up to the arbitrary time point deviates from the estimated estimated driving route, and detecting that an abnormality has occurred in the autonomous vehicle if the calculated abnormality degree exceeds a predetermined threshold.

2. The anomaly detection method according to claim 1, further comprising obtaining information about the plurality of routes when the driving route up to the arbitrary time point is obtained or each time a predetermined period of time elapses, and the information about the plurality of routes read to estimate the estimated driving route includes information about the plurality of routes obtained in the past.

3. The anomaly detection method according to claim 1, further comprising a step of, if an abnormality in the autonomous vehicle is detected in the detection step, displaying a detection result indicating that an abnormality has occurred in the autonomous vehicle, as well as predetermined map information mapping the driving route from the departure point to the arbitrary time point and the estimated driving route.

4. The anomaly detection method described in claim 1, wherein the information about the multiple routes is a driving route determination parameter, the driving route determination parameter is a parameter consisting of a combination of values ​​indicating the presence or absence of each of multiple events that may affect the determination of a driving route from the departure point to the destination by the autonomous vehicle, and in the detection step, the estimated driving route is estimated based on the driving route determination parameter.

5. The anomaly detection method described in claim 4, wherein the information about the multiple routes includes at least one of the distance from the starting point to the destination for each of the multiple routes and the time required to travel from the starting point to the destination when the autonomous vehicle travels each of the multiple routes.

6. The anomaly detection method described in claim 4, wherein the information about the multiple routes includes information indicating events that may be obstacles to the autonomous vehicle traveling along each of the multiple routes at any given point in time.

7. The anomaly detection method described in claim 6, wherein the information indicating events that may be obstacles when the autonomous vehicle travels along each of the plurality of routes includes information indicating whether the autonomous vehicle is capable of traveling along each of the plurality of routes.

8. The anomaly detection method described in claim 4, wherein the autonomous vehicle is a vehicle used for transporting transported goods, the acquiring step further acquires information about the transported goods transported by the autonomous vehicle, the information about the transported goods being the travel route determination parameter including at least one of one or more pieces of information indicating attributes of the transported goods and one or more pieces of information indicating transportation conditions of the transported goods, and the detecting step estimates the estimated travel route based on the information about the multiple routes and the information about the transported goods.

9. The anomaly detection method according to claim 4, further comprising the step of selecting and reading out one of a plurality of anomaly detection models corresponding to each of a plurality of the driving route determination parameters, wherein one of the plurality of routes is defined in each of the plurality of anomaly detection models, and in the reading step, the anomaly detection model corresponding to the driving route determination parameters is selected and read out, and in the detecting step, the one route defined in the read anomaly detection model is estimated as the estimated driving route.

10. The anomaly detection method according to claim 9, wherein one route defined for each of the plurality of anomaly detection models is updated to any one of the plurality of routes at each predetermined time period.

11. The anomaly detection method described in claim 10, wherein each of the multiple anomaly detection models is assigned a route weight for each of the multiple routes, the route weight being a degree indicating the probability that the autonomous vehicle will travel along each of the multiple routes, the route weight being updated at predetermined intervals, and one of the multiple routes is determined as the updated route in each of the multiple anomaly detection models based on the route weight for each of the multiple routes.

12. The anomaly detection method according to claim 10, further comprising a step of learning each of the plurality of anomaly detection models, wherein the learning step, for each predetermined period of time, acquires past driving routes which are driving routes in each of a plurality of past driving trips of the autonomous vehicle from the departure point to the destination in which the combination of values ​​of the driving route determination parameters was the same, calculates a frequency of the autonomous vehicle traveling each of the plurality of routes from each of the acquired plurality of past driving routes, and determines, based on the calculated frequency, one of the plurality of routes as one route in the anomaly detection model corresponding to the driving route determination parameters.

13. The anomaly detection method according to claim 9, wherein each of the multiple routes consists of a combination of one or more sections connecting two of a plurality of points arbitrarily determined on the multiple routes, and in each of the multiple anomaly detection models, a section weight indicating a degree of probability that the autonomous vehicle will travel through each of the one or more sections is determined for each of the one or more sections, and in the detection step, the degree of anomaly of the driving route to the arbitrary time point is calculated based on the ratio between the sum of the section weights determined for each of the one or more sections included in the driving route from the departure point to the arbitrary time point and the section weights determined for each of the one or more sections included in the estimated driving route, and if the calculated degree of anomaly exceeds a predetermined threshold, an abnormality occurring in the autonomous vehicle is detected.

14. The anomaly detection method described in claim 13, wherein in the detection step, an event that will cause the autonomous vehicle to determine an exceptional driving route that is different from normal is determined, and when the driving route determination parameters are parameters that include in their combination a value indicating the presence of the event, a correction is made to reduce the degree of abnormality, and when the degree of abnormality after correction exceeds a preset threshold value, an abnormality occurring in the autonomous vehicle is detected.

15. The anomaly detection method according to claim 9, further comprising: setting a plurality of parameter groups for classifying any of the driving route determination parameters into a plurality of parameter groups; one or more of the driving route determination parameters are defined in each of the plurality of parameter groups; each of the plurality of anomaly detection models corresponds to a respective one of the plurality of parameter groups; and in the detecting step, determining in which of the plurality of parameter groups the driving route determination parameter is defined; selecting and reading out an anomaly detection model corresponding to the parameter group in which the driving route determination parameter is defined; and estimating one route defined by the read out anomaly detection model as the estimated driving route.

16. The anomaly detection method described in claim 15, wherein an importance level representing a tendency of values ​​of one or more of the determined driving route determination parameters is determined for each of the plurality of parameter groups, the importance level consisting of a combination of a plurality of values ​​corresponding to a combination of values ​​indicated by the driving route determination parameters, and in the detection step, when the driving route determination parameter is not determined in any of the plurality of parameter groups, the anomaly detection model corresponding to the parameter group whose importance level is most similar to the driving route determination parameter is selected and read out from the plurality of parameter groups.

17. The anomaly detection method according to claim 15, further comprising a step of learning each of the plurality of anomaly detection models, wherein in the learning step, a plurality of the driving route determination parameters acquired in each of a plurality of past drivings of the autonomous vehicle from the departure point to the destination and a plurality of past driving routes which are the driving routes of the autonomous vehicle during the plurality of past drivings are acquired, and the plurality of parameter groups are created by grouping the plurality of driving route determination parameters determined in each of one or more of the past drivings in which the plurality of past driving routes are similar as the same parameter group.

18. The anomaly detection method according to claim 16, further comprising a step of learning each of the plurality of anomaly detection models, wherein in the learning step, for each of the plurality of parameter groups, when a value indicating the presence of one event among the plurality of events included in one or more of the driving route determination parameters characterizes the parameter group, the importance value corresponding to the value indicating the presence or absence of the event is determined to be higher than in other cases, and when a value indicating the absence of one event among the plurality of events included in one or more of the driving route determination parameters characterizes the parameter group, the importance value corresponding to the value indicating the presence or absence of the event is determined to be lower than in other cases.

19. The anomaly detection method described in claim 17, wherein in the learning step, the frequency at which the autonomous vehicle has traveled each of the plurality of routes is calculated using each of the plurality of similar past driving routes corresponding to one of the grouped parameter groups, and the anomaly detection model is further created by determining a route weight for each of the plurality of routes based on the calculated frequency.

20. An anomaly detection device that detects an abnormality in an autonomous vehicle traveling autonomously from a starting point to a destination, comprising: an acquisition unit that acquires a driving route of the autonomous vehicle from departing from the starting point to an arbitrary time point; a memory unit that stores information about a plurality of routes that the autonomous vehicle may travel from the starting point to the destination; and an anomaly detection unit that estimates one of the plurality of routes as an estimated driving route of the autonomous vehicle from the starting point to the destination based on the information about the plurality of routes read from the memory unit, calculates a degree of anomaly that indicates how much the driving route up to the arbitrary time point deviates from the estimated estimated driving route, and detects that an abnormality has occurred in the autonomous vehicle if the calculated degree of anomaly exceeds a predetermined threshold.

21. A program for causing a computer to execute the anomaly detection method according to any one of claims 1 to 19.

Images