WO2023286866A1 - Calculation device, calculation method, and program - Google Patents

Abstract

The present invention inhibits a decrease in inference accuracy of a road surface condition. This calculation device comprises: a first determination unit (34) that, on the basis of behavior information detected in a predetermined position and first reference behavior information detected in a first region overlapping the predetermined position, determines whether the behavior information detected in the predetermined position corresponds to the first reference behavior information; a second determination unit (36) that, if the behavior information detected in the predetermined position is determined not to correspond to the first reference behavior information, on the basis of behavior information detected in a second region encompassing the entire first region and broader than the first region and second reference behavior information detected in the second region, determines whether the behavior information detected in the second region corresponds to the second reference behavior information; and a learning unit (30) that causes an AI model to machine-learn. If the behavior information detected in the second region is determined not to correspond to the second reference behavior information, the learning unit (30) updates the AI model with the behavior information detected in the second region as teacher data.

Description

Arithmetic device, arithmetic method and program

The present invention relates to an arithmetic device, an arithmetic method and a program.

For example, as shown in Patent Document 1, there is known a technique for estimating road surface conditions by detecting the acceleration of a vehicle traveling on a road and inputting the acceleration data into a learning model.

Japanese Patent Application Laid-Open No. 2020-86960

However, if the characteristics on the side that detects vehicle behavior change, there is a risk that the accuracy of road surface condition estimation using the learning model will decrease. SUMMARY OF THE INVENTION It is an object of the present invention to provide an arithmetic device, an arithmetic method, and a program capable of suppressing deterioration in accuracy of road surface condition estimation.

In order to solve the above-described problems and achieve the object, an arithmetic device according to the present disclosure acquires behavior information indicating the behavior of the vehicle detected by a behavior sensor provided in the vehicle traveling on the road. an acquisition unit, the behavior information detected at a predetermined position, and a first reference that is behavior of the vehicle detected in a first region overlapping the predetermined position before the behavior information detected at the predetermined position. a first determination unit that determines whether vehicle behavior indicated by the behavior information detected at the predetermined position corresponds to vehicle behavior indicated by the first reference behavior information, based on the behavior information; When it is determined that the behavior of the vehicle indicated by the behavior information detected in the above does not correspond to the behavior of the vehicle indicated by the first reference behavior information, the entire area of the first area is included, and more than the first area. Based on the behavior information detected within a wide second area and second reference behavior information detected within the second area prior to the behavior information detected within the second area, the a second judgment unit for judging whether the behavior of the vehicle indicated by the behavior information detected in the second area corresponds to the behavior of the vehicle indicated by the second reference behavior information; a learning unit that causes the AI model to perform machine learning of a correspondence relationship with a road surface state, wherein the learning unit is configured such that the behavior of the vehicle indicated by the behavior information detected in the second area and the second reference behavior information are If it is determined that the behavior of the vehicle shown does not correspond, the AI model is updated using the behavior information detected in the second area as teacher data.

In order to solve the above-described problems and achieve the object, a calculation method according to the present disclosure includes a step of obtaining behavior information indicating the behavior of the vehicle detected by a behavior sensor provided in the vehicle traveling on the road; , the behavior information detected at a predetermined position; and first reference behavior information, which is behavior of the vehicle detected in a first region overlapping the predetermined position before the behavior information detected at the predetermined position. determining whether the behavior of the vehicle indicated by the behavior information detected at the predetermined position corresponds to the behavior of the vehicle indicated by the first reference behavior information based on the behavior detected at the predetermined position; When it is determined that the information and the first reference behavior information do not correspond, the behavior information detected in a second area that includes the entire first area and is wider than the first area; Vehicle behavior indicated by the behavior information detected in the second area based on second reference behavior information detected in the second area prior to the behavior information detected in the second area. determining whether the behavior of the vehicle indicated by the second reference behavior information corresponds to the behavior of the vehicle; and causing an AI model to perform machine learning of the correspondence relationship between the behavior of the vehicle and the road surface condition of the road. and in the step of causing the AI model to perform machine learning, when it is determined that the behavior of the vehicle indicated by the behavior information detected in the second region does not correspond to the behavior of the vehicle indicated by the second reference behavior information. updates the AI model using the behavior information detected in the second area as teacher data.

In order to solve the above-described problems and achieve the object, a program according to the present disclosure acquires behavior information indicating the behavior of the vehicle detected by a behavior sensor provided in the vehicle traveling on the road; the behavior information detected at a predetermined position; and first reference behavior information, which is vehicle behavior detected in a first region overlapping the predetermined position, before the behavior information detected at the predetermined position. determining whether the vehicle behavior indicated by the behavior information detected at the predetermined position and the vehicle behavior indicated by the first reference behavior information correspond to each other; and the behavior information detected at the predetermined position. and the first reference behavior information do not correspond, the behavior information detected in a second region that includes the entire first region and is wider than the first region, and the second vehicle behavior indicated by the behavior information detected in the second area based on second reference behavior information detected in the second area prior to the behavior information detected in the area; A computer executes a step of determining whether or not the behavior of the vehicle indicated by the second reference behavior information corresponds, and a step of allowing an AI model to perform machine learning of the correspondence relationship between the behavior of the vehicle and the surface condition of the road. In the step of causing the AI model to perform machine learning, it is determined that the behavior of the vehicle indicated by the behavior information in the second region does not correspond to the behavior of the vehicle indicated by the second reference behavior information. In this case, the AI model is updated using the behavior information detected in the second area as teacher data.

　According to the present invention, it is possible to suppress the deterioration of the estimation accuracy of the road surface condition.

FIG. 1 is a schematic block diagram of a detection system according to this embodiment. FIG. 2 is a schematic diagram of a vehicle. FIG. 3 is a schematic block diagram of an arithmetic device. FIG. 4 is a schematic diagram showing an example of types of the first AI model. FIG. 5A is a schematic diagram showing an example of types of the second AI model. FIG. 5B is a schematic diagram illustrating an example of the first area and the second area. FIG. 6 is a flowchart for explaining the processing flow of the arithmetic device.

Preferred embodiments of the present invention will be described in detail below based on the drawings. In addition, this invention is not limited by embodiment described below.

(detection system)
FIG. 1 is a schematic block diagram of a detection system according to this embodiment. As shown in FIG. 1 , the detection system 1 according to this embodiment includes a vehicle 10 , a measurement data acquisition device 12 and an arithmetic device 14 . The detection system 1 is a system that determines the road surface condition based on the behavior information of the vehicle 10 traveling on the road. Behavior information is information that indicates the behavior of the vehicle 10 traveling on a road, and will be described later in detail. The detection system 1 estimates the road surface condition by calculating the road surface condition based on the behavior information using the arithmetic unit 14 . The road surface condition is an index indicating the degree of unevenness of the road surface in this embodiment. More specifically, in this embodiment, the road surface condition is the IRI (International Roughness Index). However, the road surface condition is not limited to the IRI, and may be any index indicating the road surface condition. For example, the road surface condition may be at least one of IRI, road surface flatness, cracks, rutting, and MCI (Maintenance Control Index).

In the detection system 1 , the vehicle 10 detects behavior information and position information while traveling on a road, and transmits the detected behavior information and position information to the measurement data acquisition device 12 . The measurement data acquisition device 12 is, for example, a device (computer) managed by an entity that manages roads. The measurement data acquisition device 12 transmits the behavior information and position information transmitted from the vehicle 10 to the arithmetic device 14 . The calculation device 14 estimates the road surface condition based on the behavior information and position information transmitted from the measurement data acquisition device 12 . The calculation device 14 then transmits the estimation result of the road surface condition to the measurement data acquisition device 12 . In this way, the computing device 14 acquires behavior information and position information via the measurement data acquisition device 12, but is not limited to this. For example, the detection system 1 may not be provided with the measurement data acquisition device 12 and the arithmetic device 14 may directly acquire behavior information and position information from the vehicle 10 .

(vehicle)
FIG. 2 is a schematic diagram of a vehicle. As shown in FIG. 2, the vehicle 10 includes a position sensor 10A, a behavior sensor 10B, and a measuring device 10C. The position sensor 10A is a sensor that acquires position information of the vehicle 10 . The positional information of the vehicle 10 is information indicating the earth coordinates of the vehicle 10 . Position sensor 10A is a module for GNSS (Global Navigation Satellite System) in this embodiment. Note that the Z direction in FIG. 2 indicates upward in the vertical direction, and FIG. 2 can be said to be a schematic diagram when the vehicle 10 is viewed from above in the vertical direction.

The behavior sensor 10B is a sensor that detects behavior information that indicates the behavior of the vehicle 10. The behavior information may be any information as long as it indicates the behavior of the vehicle 10 traveling on the road. In this embodiment, the behavior sensor 10B preferably detects the acceleration of the vehicle 10 as behavior information. In this case, the behavior sensor 10B is an acceleration sensor that detects acceleration, more preferably an acceleration sensor that detects acceleration on three axes. Also, the behavior information detected by the behavior sensor 10B is not limited to acceleration. It may be at least one of an amount of braking of the vehicle 10, an operation amount of the wipers of the vehicle 10, and an operation amount of the suspension of the vehicle 10. Since the image data around the vehicle 10 changes according to the movement of the vehicle 10 , it can be said that the image data represents the behavior of the vehicle 10 . The behavior sensor 10B that detects the captured image around the vehicle 10 is, for example, a camera. The behavior sensor 10B that detects the speed of the vehicle 10 is, for example, a speed sensor. The behavior sensor 10B that is a gyro sensor and detects the steering angle of the vehicle 10 is, for example, a steering sensor, and the behavior sensor 10B that detects the braking amount of the vehicle 10 is, for example, a brake sensor that detects the operation of the wipers of the vehicle 10. An example of the behavior sensor 10B is a wiper sensor, and an example of the behavior sensor 10B that detects the amount of operation of the suspension of the vehicle 10 is a suspension sensor.

In this embodiment, the vehicle 10 is equipped with a plurality of behavior sensors 10B. Each behavior sensor 10B is mounted at a position different from each other in the vehicle 10 . In the example of FIG. 2, as the behavior sensor 10B, a behavior sensor 10B1 is provided on the Z direction side (vertical direction upward side) of the left front wheel tire TR1, and a behavior sensor 10B1 is provided on the Z direction side of the right front wheel tire TR2. a behavior sensor 10B3 provided on the Z direction side of the left rear wheel tire TR3; and a behavior sensor 10B4 provided on the Z direction side of the right rear wheel tire TR4. However, the position where the behavior sensor 10B is provided is arbitrary. Also, the number of behavior sensors 10B is not limited to four, and may be any number. Moreover, although the number of tires TR is four in the example of FIG. Also, in the example of FIG. 2, the behavior sensors 10B1 to 10B4 detect the same type of behavior information (here, acceleration), but each behavior sensor 10B detects different types of behavior information. It's okay. For example, multiple behavior sensors 10B (for example, multiple acceleration sensors) that detect the same type of behavior information and behavior sensors 10B (for example, speed sensors) that detect different behavior information may be provided.

The behavior information detected by the behavior sensors 10B1, 10B2, 10B3, and 10B4 is hereinafter referred to as behavior information D1, D2, D3, and D4 as appropriate.

The measuring device 10C is a device that controls the position sensor 10A and the behavior sensor 10B to detect position information and behavior information of the vehicle 10, and records the detected position information and behavior information. That is, the measuring device 10C functions as a data logger that records position information and behavior information of the vehicle 10 . The measuring device 10C can also be said to be a computer, and includes a control section 10C1, a storage section 10C2, and a communication section 10C3. The control unit 10C1 is an arithmetic unit, that is, a CPU (Central Processing Unit). The storage unit 10C2 is a memory that stores various types of information such as the calculation contents and programs of the control unit 10C1, the position information and behavior information of the vehicle 10, and includes, for example, a RAM (Random Access Memory) and a ROM (Read Only Memory). and at least one of non-volatile storage devices such as flash memory and HDD (Hard Disk Drive). Note that the program for the control unit 10C1 stored in the storage unit 10C2 may be stored in a recording medium readable by the measuring device 10C. The communication unit 10C3 is a communication module that communicates with an external device, such as an antenna.

The control unit 10C1 reads the program stored in the storage unit 10C2 and controls the position sensor 10A and the behavior sensor 10B. Control unit 10C1 causes behavior sensor 10B to detect behavior information of vehicle 10 at predetermined time intervals while vehicle 10 is traveling on the road. Further, the control unit 10C1 causes the position sensor 10A to detect the position information of the vehicle 10 at the timing when the behavior sensor 10B detects the behavior information. That is, control unit 10C1 causes behavior sensor 10B to detect behavior information of vehicle 10 and causes position sensor 10A to detect position information of vehicle 10 each time vehicle 10 travels for a predetermined period of time. The predetermined time here is preferably a fixed time such as one minute, but the predetermined time is not limited to a fixed time and may be any length. That is, the predetermined time may change each time.

The control unit 10C1 transmits the behavior information and the position information to the measurement data acquisition device 12 via the communication unit 10C3. The measurement data acquisition device 12 transmits the behavior information and position information received from the vehicle 10 to the arithmetic device 14 . If the measurement data acquisition device 12 is not provided, the control section 10C1 may directly transmit the behavior information and the position information to the arithmetic device 14. FIG.

FIG. 3 is a schematic block diagram of an arithmetic unit. As shown in FIG. 3, the computing device 14 is, for example, a computer, and includes a communication section 20, a storage section 22, and a control section 24. The communication unit 20 is a communication module that communicates with an external device, such as an antenna. The storage unit 22 is a memory that stores various information such as calculation contents and programs of the control unit 24, a first AI model M and a second AI model N, which will be described later. At least one of non-volatile storage devices such as flash memory and HDD is included. Note that the program for the control unit 24 and the first AI model M and the second AI model N stored in the storage unit 22 may be stored in a recording medium readable by the computing device 14 .

The control unit 24 is an arithmetic device and includes an arithmetic circuit such as a CPU (Central Processing Unit). Control unit 24 includes learning unit 30 , behavior information acquisition unit 32 , first determination unit 34 , second determination unit 36 , and calculation unit 38 . The control unit 24 implements the learning unit 30, the behavior information acquisition unit 32, the first determination unit 34, the second determination unit 36, and the calculation unit 38 by reading and executing a program (software) from the storage unit 22. to perform those operations. Note that the control unit 24 may execute these processes by one CPU, or may be provided with a plurality of CPUs and may execute the processes by the plurality of CPUs. Moreover, at least a part of the learning unit 30, the behavior information acquisition unit 32, the first determination unit 34, the second determination unit 36, and the calculation unit 38 may be realized by hardware.

(learning department)
(Learning of the first AI model)
The learning unit 30 generates a learned first AI model M by subjecting the untrained first AI model M to machine learning. When the behavior information of the vehicle 10 is input to the first AI model M, the learning unit 30 performs machine learning on the first AI model M so that information indicating whether the behavior information is normal is output from the first AI model M. Let That is, the first AI model M is an AI (Artificial Intelligence) model that detects anomalies in behavior information. Any model may be used as the first AI model M. For example, a model that performs logistic regression analysis, a CNN (Conventional Neural Network) model, or the like may be used.

When the behavior information is normal, the behavior of the vehicle 10 indicated by the behavior information corresponds to the behavior of the vehicle 10 indicated by the reference behavior information, which is the behavior information of the vehicle 10 detected before the behavior information. means that there is In other words, that the behavior information is normal means that the behavior of the vehicle 10 indicated by the behavior information is of the same quality as the behavior of the vehicle 10 indicated by the reference behavior information. means that the behavior of the vehicle 10 is similar to that indicated by the reference behavior information. Further, behavior information is not normal (abnormal) means that the behavior information does not correspond to the reference behavior information, in other words, that the behavior information is different from the reference behavior information. , for example, means that the behavior information is not similar to the reference behavior information. The reference behavior information is data (acceleration in this embodiment) detected by causing the vehicle 10 to travel on the same road before detecting the behavior information and estimating the road surface condition. It can be said that it is the initial behavior information before deterioration.

The learning unit 30 sets a data set of reference behavior information and a label indicating that the reference behavior information is normal as teacher data. That is, the learning unit 30 sets normal data as teacher data. The learning unit 30 prepares a plurality of data sets of reference behavior information and labels as teacher data. The learning unit 30 inputs a data set (teacher data) of reference behavior information and labels to the unlearned first AI model M, and when behavior information is input, information as to whether or not the behavior information is normal. machine learning of the first AI model M so as to be able to output (to be able to label the behavior information as to whether it is normal or not).

FIG. 4 is a schematic diagram showing an example of the types of the first AI model. In this embodiment, as shown in FIG. 4, the learning unit 30 prepares the first AI model M for each combination of the unit area, the motion state range, and the behavior sensor 10B that detected the behavior information. A unit area refers to one divided area when the area in which the vehicle 10 travels is divided into predetermined areas. Further, the motion state range refers to the range of the motion state of the vehicle, and in this embodiment refers to the speed range of the vehicle. For example, the range between the assumed upper limit value and lower limit value of the vehicle speed may be divided into a plurality of speed ranges, and each divided speed range may be set as the exercise state range.

When the learning unit 30 machine-learns the first AI model M corresponding to the combination of the predetermined behavior sensor 10B, the predetermined unit area, and the predetermined motion state range, the vehicle position is within the predetermined unit area. The first AI model M to learn For example, in the example of FIG. 4, the first AI model M1a can be said to be a model corresponding to a combination of the unit area UR1, the exercise state range C1, and the behavior sensor 10B1 (input data is the behavior information D1). In this case, the learning unit 30 determines that the position of the vehicle when the reference behavior information is detected is within the unit region UR1, the vehicle speed is within the motion state range C1 when the reference behavior information is detected, and , the reference behavior information detected by the behavior sensor 10B1 is extracted. The learning unit 30 machine-learns the first AI model M1a using the extracted reference behavior information as teacher data for the first AI model M1a. Note that in FIG. 4, a first AI model M1b corresponding to the combination of the unit region UR1, the motion state range C2, and the behavior information D1, and a first AI model M1b corresponding to the combination of the unit region UR1, the motion state range C1, and the behavior information D2 A first AI model M2b corresponding to the combination of the model M2a, the unit region UR1, the motion state range C2, and the behavior information D2, and a first AI model M3a corresponding to the combination of the unit region UR1, the motion state range C1, and the behavior information D3. , a first AI model M3b corresponding to the combination of the unit region UR1, the motion state range C2, and the behavior information D3, and a first AI model M4a corresponding to the combination of the unit region UR1, the motion state range C1, and the behavior information D4; A first AI model M4b corresponding to the combination of the unit area UR1, the motion state range C2, and the behavior information D4 is illustrated.

(Learning of the second AI model)
The learning unit 30 causes the unlearned second AI model N to machine-learn the correspondence relationship between the behavior of the vehicle and the road surface condition, thereby learning the correspondence relationship between the behavior of the vehicle and the road surface condition. Generate 2AI model N. The learning unit 30 may perform machine learning of the correspondence relationship between the vehicle behavior detection results obtained by one behavior sensor 10B and the road surface conditions, or may perform machine learning of the vehicle behavior detection results obtained by a plurality of behavior sensors 10B and the road surface conditions. The correspondence may be machine-learned. The second AI model N may be any model capable of outputting road surface conditions when behavior information is input, but may be, for example, a CNN model.

The learning unit 30 sets a data set of the reference behavior information and the road surface condition at the position where the reference behavior information is detected as teacher data. A value measured in advance may be used for the road surface condition at the position where the reference behavior information is detected. The learning unit 30 sets a plurality of data sets of reference behavior information and road surface conditions as teacher data. The learning unit 30 inputs the data set of the reference behavior information and the road surface condition set as teacher data to the unlearned second AI model N, and transfers the correspondence relationship between the behavior information and the road surface condition to the second AI model N. machine learning.

FIG. 5A is a schematic diagram showing an example of types of the second AI model. In the present embodiment, as shown in FIG. 5A, the learning unit 30 prepares a plurality of second AI models N in which at least some of the behavior sensors 10B that detect behavior information as input data are different. That is, each second AI model N differs in at least part of the behavior information input as input data. In FIG. 5A, a second AI model N1 to which behavior information D1, D2, D3, and D4 detected by the behavior sensors 10B1, 10B2, 10B3, and 10B4 are input, and behavior information D1 and D2 detected by the behavior sensors 10B1, 10B2, and 10B3. , D3 are input, the second AI model N3 is input with behavior information D1, D2, and D4 detected by the behavior sensors 10B1, 10B2, and 10B4, and the behavior detected by the behavior sensors 10B1, 10B3, and 10B4. A second AI model N4 to which information D1, D3, and D4 are input, a second AI model N5 to which behavior information D2, D3, and D4 detected by the behavior sensors 10B2, 10B3, and 10B4 are input, and the behavior detected by the behavior sensor 10B4. and a second AI model Nn to which information D4 is input.

When the learning unit 30 machine-learns the second AI model N to which the behavior information of the predetermined behavior sensor 10B is input, the learning unit 30 uses the reference behavior information detected by the predetermined behavior sensor 10B as teacher data to generate the second AI model. Let N be machine-learned. In other words, for example, the learning unit 30 acquires the reference behavior information detected by the behavior sensors 10B1, 10B2, 10B3, and 10B4 and the road surface conditions at the positions where the reference behavior information is detected as a teacher of the second AI model N1. As data, the second AI model N1 is machine-learned.

(behavior information acquisition unit)
The behavior information acquisition unit 32 acquires behavior information detected by the behavior sensor 10B. More specifically, the behavior information acquisition unit 32 acquires behavior information and position information indicating the position of the vehicle 10 when the behavior information is detected. The behavior information acquisition unit 32 acquires behavior information and position information from the measurement data acquisition device 12 via the communication unit 20 , but may also acquire the behavior information and the position information directly from the vehicle 10 .

(First judgment unit)
The first determination unit 34 determines behavior information detected at a predetermined position and reference behavior information detected in a first region overlapping the predetermined position where the behavior information is detected (more than the behavior information detected at the predetermined position). First reference behavior information previously detected in the first area), whether the behavior information detected at the predetermined position corresponds to the reference behavior information detected in the first area, that is, whether the behavior information is normal determine whether it is The predetermined position here may be any point (position) on the route when the vehicle 10 travels on the road. That is, the first determination unit 34 may extract behavior information detected at an arbitrary position as behavior information detected at a predetermined position from a plurality of pieces of behavior information detected at different positions. The first area is an area of a predetermined area that overlaps the position (predetermined position) of the vehicle 10 when the behavior information is detected. good. In this embodiment, among the unit areas obtained by dividing the area in which the vehicle 10 travels, the unit area overlapping the position of the vehicle 10 when the behavior information is detected may be the first area. That is, the first determination unit 34 selects a unit area overlapping the position of the vehicle 10 indicated by the position information acquired by the behavior information acquisition unit 32 (the position information of the vehicle 10 when the behavior information is detected) as the first area. Based on the extracted reference behavior information and the behavior information detected in the first area, it is determined whether the behavior information is normal.

In the present embodiment, the first determination unit 34 selects a learned first AI model corresponding to the first area (a unit area that overlaps with the position of the vehicle 10 when the behavior information is detected) from among the plurality of first AI models M. 1 AI model M is read, and behavior information detected at a predetermined position is input to the read 1 AI model M. The first determination unit 34 acquires information indicating whether the behavior information output from the first AI model M is normal, and based on the acquired information, determines whether the behavior information detected at the predetermined position is normal. to judge. For example, when a label indicating that the behavior information is normal is output from the first AI model M1a, the first determination unit 34 determines that the behavior information is normal, and the label indicating that the behavior information is not normal is output to the first AI model M1a. When output from the 1AI model M1a, it may be determined that the behavior information is not normal (abnormal). Further, in a case where the probability that the behavior information is normal is output from the first AI model M1a, the first determination unit 34 determines that the behavior information is normal when the probability that the behavior information is normal is equal to or greater than a predetermined value. If it is determined that the behavior information is normal and the probability that the behavior information is normal is less than a predetermined value, it may be determined that the behavior information is not normal.

More specifically, the behavior information acquisition unit 32 also acquires the motion state of the vehicle 10 when the behavior information is detected, along with the behavior information and the position information. The motion state is the speed of the vehicle 10 in this embodiment, and is detected by a speed sensor mounted on the vehicle 10 . The first determination unit 34 extracts a motion state range (vehicle speed range) that overlaps with the motion state (vehicle speed) of the vehicle 10 when the behavior information is detected. Then, the first determination unit 34 reads out the first AI model M1 corresponding to the combination of the behavior sensor 10B that detected the behavior information, the unit area extracted as the first area, and the extracted motion state range. The first determination unit 34 inputs the behavior information to the read first AI model M and determines whether the behavior information is normal. That is, for example, when the behavior information is detected by the behavior sensor 10B1, the position of the vehicle 10 when the behavior information is detected overlaps the unit area UR1, and the motion state when the behavior information is detected overlaps the motion state range C1. 4, the first determination unit 34 reads the first AI model M1a (see FIG. 4), inputs the behavior information to the first AI model M1a, and determines whether the behavior information is normal. Even when the first AI model M is used in this way, since the first AI model M is learned based on the reference behavior information, it can be said that whether the behavior information is normal is determined based on the reference behavior information.

Here, since a plurality of behavior sensors 10B are provided, the behavior information acquisition unit 32 acquires behavior information detected by each behavior sensor 10B at the same position. The first judgment unit 34 judges whether or not each piece of behavior information detected by each behavior sensor 10B at the same predetermined position is normal. That is, for example, the first determination unit 34 determines the behavior information D1 detected by the behavior sensor 10B1 at a predetermined position, the behavior information D2 detected by the behavior sensor 10B2 at a predetermined position, and the behavior information D3 detected by the behavior sensor 10B3 at a predetermined position. , and the behavior information D4 detected by the behavior sensor 10B4 at a predetermined position, it is determined whether they are normal.

It should be noted that the determination of whether the behavior information is normal by the first determination unit 34 is not limited to using the first AI model M. For example, the first determination unit 34 may set a range of behavior information to be normal (normal range) based on the reference behavior information, and determine whether the behavior information is normal based on the normal range. The first determination unit 34 determines that the behavior information is normal when the behavior information acquired by the behavior information acquisition unit 32 falls within the normal range, and determines that the behavior information is not normal when the behavior information does not fall within the normal range. I judge. The normal range may be set by any method. For example, a 3σ range calculated from a plurality of pieces of reference behavior information based on a normal distribution method may be set as the normal range. Also, the normal range may be set for each combination of the unit area, the exercise state range, and the behavior sensor 10B that detected the behavior information, as in the first AI model M. In this case, the first determination unit 34 reads the normal range set for the combination of the unit area corresponding to the behavior information, the exercise state range, and the behavior sensor 10B, and if the behavior information falls within the normal range, the behavior Judge the information as normal.

(Second judgment unit)
The second determination unit 36 determines behavior information detected in a second area that is wider than the first area and reference behavior information detected in the second area (before the behavior information detected in the second area). , second reference behavior information detected in the second region), and whether the behavior information detected in the second region corresponds to the reference behavior information, that is, the behavior detected in the second region Determine if the information is correct. The second determination unit 36 determines whether the behavior information detected in the second area is normal when the first determination unit 34 determines that the behavior information detected at the predetermined position is not normal. In the present embodiment, when it is determined that all the behavior information detected by each behavior sensor 10B at the predetermined position is not normal, the second determination unit 36 determines that the behavior information detected within the second area is normal. determine whether it is That is, in the present embodiment, when it is determined that only the behavior information by some of the behavior sensors 10B at the predetermined position is not normal, it is not necessary to determine whether the behavior information detected within the second area is normal. good.

FIG. 5B is a schematic diagram explaining an example of the first area and the second area. The second area may be an area at any position in the area where the vehicle 10 travels, as long as it is wider than the first area. In the present embodiment, the second area preferably includes the entire first area and is wider than the first area. For example, the second area may be the entire area (entire section) traveled by the vehicle 10 . For example, as shown in FIG. 5B, it is assumed that behavior information is obtained at predetermined positions Pa, Pb, Pc, and Pd in the area AR in which the vehicle 10 travels. In this case, for example, the first regions (unit regions) overlapping the predetermined positions Pa, Pb, Pc, and Pd can be said to be the first regions UR1a, UR1b, UR1c, and UR1d, respectively. In this case, the second region UR2 is wider than the first regions UR1a, UR1b, UR1c, and UR1d, and is the entire first region UR1a, the entire first region UR1b, the entire first region UR1c, and It may be an area that includes the entire first area UR1d, or an area that includes the entire area AR in which the vehicle 10 travels. However, the present invention is not limited to this, and for example, the second region UR2 may include only the entire regions of the first region UR1a and the second region UR1b.

Within the second area, multiple pieces of behavior information are detected for each position. The second determination unit 36 determines whether the behavior information is normal for each piece of behavior information detected at different positions in the second region, and determines whether the behavior information is normal based on the determination result. 2 It is determined whether the behavior information detected in the area is normal. The second determination unit 36 may determine whether the behavior information for each position is normal by the same method as the determination method by the first determination unit 34 . That is, for example, the second determination unit 36 determines the behavior sensor 10B that detected the behavior information, the unit area that overlaps the position where the behavior information was detected, and the motion state range that overlaps the motion state when the behavior information was detected. 1AI model M1 is read, behavior information is input to the read 1AI model M, and information indicating whether the behavior information is normal is acquired. The second determination unit 36 acquires information indicating whether the behavior information is normal by the same method for each piece of behavior information detected at different positions in the second area. The second determination unit 36 determines whether the behavior information detected within the second area is normal based on the information indicating whether each behavior information is normal. For example, the second determination unit 36 determines that the behavior information detected in the second region, if the ratio of the behavior information determined to be normal in the first AI model M is equal to or greater than a predetermined value, If the behavior information detected in the first AI model M is determined to be normal, and if the ratio of the behavior information determined to be normal in the first AI model M is less than a predetermined value, the behavior information detected in the second region may be judged to be abnormal.

The second determination unit 36 may determine whether the behavior information detected by each behavior sensor 10B in the second region is normal, or may determine whether some of the behavior sensors 10B ( For example, only the behavior information detected by the behavior sensor 10B1) may be judged to be normal. Further, in the present embodiment, the second determination unit 36 detects behavior within the second region based on a plurality of pieces of behavior information detected within the second region and a plurality of pieces of reference behavior information detected within the second region. Determines whether the received behavior information is normal. However, the second determination unit 36 is not limited to using a plurality of pieces of behavior information and a plurality of pieces of reference behavior information. Based on one piece of reference behavior information, it may be determined whether the behavior information detected in the second area is normal.

Here, while the behavior information determined to be normal by the first determination unit 34 is data for one point at a predetermined position, the behavior information determined to be normal by the second determination unit 36 is , data of a plurality of points in the second region. That is, it can be said that the second determination unit 36 determines whether a larger number of pieces of behavior information are normal than the first determination unit 34 does. However, the behavior information that is judged to be normal by the first judging unit 34 is not limited to data of one point, and may be data of a plurality of points. The number of pieces of behavior information for which it is determined whether or not is larger.

The determination of whether the behavior information is normal by the second determination unit 36 is not limited to using the first AI model M. The second determination unit 36 may set a normal range and determine whether the behavior information is normal based on the normal range, as described for the first determination unit 34 . Further, the second determination unit 36 uses an AI model different from the first AI model M, which can determine whether the behavior information in the second area is normal, and determines whether the behavior information is normal. good.

An example of behavior information detected at a predetermined position and reference behavior information detected in the first area and the second area will be described with reference to FIG. 5B. In FIG. 5B, behavior information is acquired at predetermined positions Pa, Pb, Pc, and Pd, and first regions (unit regions) UR1a, UR1b, UR1c, and UR1d overlap predetermined positions Pa, Pb, Pc, and Pd. This is the first area. In this case, the first determination unit 34 treats the behavior information detected in the past at the position Pa′ in the first region UR1a as the reference behavior information detected in the first region UR1a, and treats the behavior information detected at the predetermined position Pa as It is determined whether the behavior information detected at the position Pa' corresponds to the reference behavior information detected at the position Pa'. Similarly, the first determination unit 34 determines whether the behavior information detected at each of the predetermined positions Pb, Pc, and Pd corresponds to the reference behavior information detected at each of the positions Pb', Pc', and Pd'. to judge. For example, when a plurality of pieces of reference behavior information are detected in one first area, the average value of the pieces of reference behavior information may be treated as the reference behavior information of the first area. That is, for example, the first determination unit 34 may determine whether the behavior information detected at the predetermined position Pa and the average value of the plurality of reference behavior information detected within the first region UR1a correspond to each other. .

In addition, in the example of FIG. 5B, the second determination unit 34 combines the behavior information detected in the past at each of the positions Pa', Pb', Pc', and Pd' with the reference behavior information detected in the second region UR2. , and it is determined whether the behavior information detected at the predetermined positions Pa, Pb, Pc, and Pd correspond to the reference behavior information detected at the positions Pa', Pb', Pc', and Pd'. In this way, by comparing a plurality of pieces of behavior information with a plurality of pieces of reference behavior information detected within the second region UR2, it is possible to appropriately determine whether the behavior information is normal over a wide region. Also, as described above, one piece of behavior information may be compared with one piece of reference behavior information detected within the second region UR2. In this case, for example, the second determination unit 34 calculates the average value of the reference behavior information detected at a plurality of positions within the second region UR2, such as the position Pa' and the position Pb'. It may be determined whether the behavior information detected at the predetermined position Pa and the average value of the reference behavior information correspond to each other by treating the reference behavior information as one piece of reference behavior information. In this way, by comparing one piece of behavior information and one piece of reference behavior information, it is possible to appropriately determine whether the behavior information is normal in an area close to the position Pa while considering a wide area.

(Calculation part)
The calculation unit 38 calculates the road surface condition by inputting the behavior information to the second AI model N that has undergone machine learning. By inputting the behavior information as input data to the second AI model N, the calculation unit 38 obtains, as output data from the second AI model N, an estimation result of the road surface condition at the position where the behavior information is detected. . The calculation unit 38 calculates the road surface condition based on the estimation result of the road surface condition acquired from the second AI model N. The calculation unit 38 calculates the road surface state by different methods according to the determination results of the first determination unit 34 and the second determination unit 36 . Each case will be described below.

(When it is determined that all behavior information detected at a predetermined position is normal)
When the first determination unit 34 determines that all of the behavior information detected by each behavior sensor 10B at the predetermined position is normal, the calculation unit 38 uses the existing second AI model N to determine the existing By inputting the behavior information to the second AI model N of , the road surface condition is calculated. That is, in this case, without updating the second AI model N using the behavior information detected this time, the existing second AI model N generated based on the previously detected reference behavior information is used to change the road. Calculate the road surface condition. In this way, when it is determined that all of the behavior information detected at the predetermined position is normal, the characteristics of the vehicle 10 and the behavior sensor 10B are not changed, and the second AI model N is updated. Instead, the existing learned second AI model N is used to estimate the road surface condition.

In this embodiment, a plurality of second AI models N with different input behavior information are prepared. The calculation unit 38 inputs the behavior information to each of the second AI models N, and acquires the estimation result of the road surface state output from each of the second AI models N. The calculation unit 38 calculates the road surface condition based on the estimation result of the road surface condition output from each of the second AI models N. As shown in FIG. That is, it can be said that the calculation unit 38 is performing ensemble learning. In the example of FIG. 5A, the calculation unit 38 inputs the behavior information D1, D2, D3, and D4 to the second AI model N1, and acquires the estimated value O1 of the road surface state, which is output data. Similarly, the computing unit 38 adds behavior information D1, D2, D3, behavior information D1, D2, D4, behavior information D1, D3, D4, behavior information D2, D3, D4, and behavior information D4 are input, and estimated road surface state values O2, O3, O4, O5, and On are output data. to get Then, the calculation unit 38 calculates the road surface condition based on each of the estimated values O1 to On. For example, the calculation unit 38 may calculate an estimated value R, which is the final estimation result of the road surface state, based on the following equation (1).

　R=(r1・O1+r2・O2+r3・O3+r4・O4+r5・O5+...rn・On)/n...(1)

Here, r1, r2, r3, r4, r5, . 2 is the number of AI models N; The weighting coefficients r1 . . . rn may be set arbitrarily. As described above, in the present embodiment, the average value of the values obtained by multiplying the estimated values by each second AI model N by the weighting factor is calculated as the estimated value of the road surface condition. Any method using the model N estimated value may be used to calculate the estimated value of the road surface condition. Hereinafter, when the estimated values O1 to On for each of the second AI model N2 are not distinguished, they are referred to as an estimated value O, and when the weighting coefficients r1 to rn for each of the second AI model N2 are not distinguished, they are referred to as a weighting coefficient r. do.

(When it is determined that only part of the behavior information detected at a predetermined position is not normal)
When the first determination unit 34 determines that only a part of the behavior information detected at the predetermined position is not normal, the calculation unit 38 selects the behavior sensor 10B (abnormal sensor) determined to be abnormal. The existing second AI model N is changed so that the degree of influence on the road surface state by the behavior information detected by is lower than the degree of influence on the road surface state by the behavior information detected by the behavior sensor 10B determined to be normal. The road surface condition is calculated while using For example, when it is determined that the behavior information detected by the behavior sensor 10B1 is not normal and the behavior information detected by the behavior sensors 10B2, 10B3, and 10B4 is determined to be normal, the calculation unit 38 detects the behavior information of the behavior sensor 10B1. The degree of influence of D1 on the estimated value R of the road surface condition is lower than the degree of influence of the estimated value O of the road surface condition by the behavior information D2, D3, and D4 of the behavior sensors 10B2, 10B3, and 10B4. An estimated value R is calculated. In this way, when it is determined that only a part of the behavior information is abnormal, the degree of influence of the behavior information of the behavior sensor 10B is reduced on the assumption that the characteristics of the behavior sensor 10B have changed, and the road surface condition is determined. is estimated. As a result, the road surface state is estimated with a reduced degree of influence from the behavior sensor 10B whose characteristics have changed, so that it is possible to suppress deterioration in the estimation accuracy of the road surface state. Even if only a part of the behavior information is determined to be abnormal, the second AI model N is updated using the behavior information detected this time in the same way as when all the behavior information is determined to be normal. Instead, an existing second AI model N generated based on previously detected reference behavior information is used.

Here, the output data (estimated value O) from the second AI model N (first model) whose input data is the behavior information of the behavior sensor 10B determined to be abnormal is set as the first estimation result, and if it is not normal, The output data (estimated value O) from the second AI model N (second model) whose input data is not the determined behavior information of the behavior sensor 10B is set as the second estimation result. In this case, the calculation unit 38 obtains the first estimation result by inputting the behavior information of the behavior sensor 10B determined to be abnormal to the first model, and the behavior detected by the behavior sensor 10B determined to be abnormal. Based on the second estimation result obtained by inputting the behavior information, the degree of influence of the first estimation result on the estimated value R of the road surface condition is the degree of influence of the second estimation result on the estimated value R of the road surface condition. An estimated value R of the road surface condition is calculated so as to be lower than .

Specifically, the calculation unit 38 calculates the weighting factor r for the estimated value O (first estimation result) of the second AI model N whose input data is the behavior information of the abnormal behavior sensor 10B. A smaller value than the weighting factor r for the estimated value O (second estimation result) of the second AI model N whose behavior information is not input data. That is, for example, when it is determined that only the behavior sensor 10B1 is not normal, the weighting factors r1, r2, r3, and r4 in Equation (1) are made smaller than the weighting factors r5 and rn. As a result, the degree of influence on the estimated value R by the estimated values O1 to O4 of the second AI models N1 to N4 to which the behavior information of the behavior sensor 10B1 is input is determined by the second AI models N5 and Nn to which the behavior information of the behavior sensor 10B1 is not input. is lower than the degree of influence on the estimated value R due to the estimated value O5, On.

The method of reducing the degree of influence of the behavior information of the behavior sensor 10B determined to be abnormal is not limited to reducing the weighting factor r, and any method may be used. For example, when the behavior information is input to the second AI model N, the calculation unit 38 may correct the behavior information of the behavior sensor 10B determined to be abnormal, thereby reducing the degree of influence of the behavior information. . In this case, for example, the calculation unit 38 may input the behavior information of the behavior sensor 10B determined to be abnormal to the second AI model N as an infinite value. As a result, the calculation of the second AI model N to which the behavior information of the behavior sensor 10B determined to be abnormal is input becomes an error, and the estimated value O of the second AI model N becomes invalid, so the degree of influence should be reduced. becomes possible. Alternatively, for example, without using the second AI model N whose input data is the behavior information of the abnormal behavior sensor 10B, only the second AI model N whose input data is the behavior information of the abnormal behavior sensor 10B is used, An estimated value R may be calculated.

(When behavior information detected within the second area is determined to be normal)
As described above, when the first determination unit 34 determines that the behavior information detected at the predetermined position is not normal, the second determination unit 36 determines that the behavior information detected within the second area is normal. to judge whether When the behavior information detected in the second region is determined to be normal by the second determination unit 36, the calculation unit 38 inputs the behavior information to the existing second AI model N to determine the road surface condition. calculate. That is, in this case, similarly to the case where it is determined that all the behavior information detected at the predetermined position is normal, the behavior information detected this time is not used to update the second AI model N, and the second AI model N is updated before that. The road surface condition is calculated using the existing second AI model N generated based on the detected reference behavior information. As described above, when it is determined that the behavior information detected in the wide second area is normal although the behavior information at the predetermined position is abnormal, the characteristics of the vehicle 10 and the behavior sensor 10B change. The road surface condition is estimated by using the existing learned second AI model N without updating the second AI model N by judging that the road surface condition is changing at the point of the predetermined position. be done.

(When it is determined that the behavior information detected within the second area is not normal)
The learning unit 30 updates the second AI model N when the second determination unit 36 determines that the behavior information detected in the second area is not normal. Specifically, the learning unit 30 performs machine learning on the second AI model N using the behavior information as teacher data, and uses the behavior information detected in the second region as teacher data for the model used for estimating the road surface condition. Update to the second AI model N. The behavior information used to update the second AI model N is a group of behavior information acquired when behavior information determined to be abnormal by the second determination unit 36 is detected (i.e., during the current run of the vehicle 10). A group of detected behavioral information), and can be said to be behavioral information detected in the second region acquired after reference behavioral information used for learning of the second AI model N before updating.

More specifically, the learning unit 30 causes the second AI model N to perform machine learning by inputting a data set of behavior information and road surface conditions as teacher data into the unlearned second AI model N, thereby creating a new learned model. generate a second AI model N of That is, the learning unit 30 does not re-learn the learned second AI model N, which has been machine-learned using the reference behavior information, using the behavior information. It can be said that a new second AI model N is generated as data. The road surface condition used as teaching data for the new second AI model N may be the measured road surface condition used as teaching data for the second AI model N before updating, instead of the newly measured values. However, among the measured road surface conditions, road surface conditions that are judged to be unusable due to reasons such as old data need not be used as teacher data.

The calculation unit 38 uses the new second AI model N updated using the behavior information to estimate the road surface condition. In this case, the behavior information detected this time is used to generate a new second AI model N, so it is not used to estimate the road surface state. That is, the calculation unit 38 inputs new behavior information detected after updating the second AI model N to the updated second AI model N, thereby estimating the road surface state.

In this way, when the behavior information detected in the second area is not normal, the existing second AI model N is used because the characteristics of the entire vehicle 10 or the behavior sensor 10B have changed instead of the individual behavior sensor 10B. Therefore, the second AI model N is updated using the latest behavior information. Therefore, it becomes possible to use the second AI model N adapted to changes in the characteristics of the vehicle 10, and it is possible to suppress deterioration in the estimation accuracy of the road surface state.

As described above, in the present embodiment, when a part of the behavior information detected at a predetermined position is abnormal, the processing to reduce the degree of influence of the abnormal behavior sensor 10B is performed, and the behavior information detected at the predetermined position is If all of the behavior information is abnormal and the behavior information detected within the second area is not normal, a process of updating the second AI model N is performed. However, it is not limited to performing all of these processes, and at least some of these processes may be performed. That is, for example, only the process of updating the second AI model N may be performed when the behavior information detected at the predetermined position is not normal and the behavior information detected within the second area is not normal. Conversely, only the process of reducing the degree of influence of the abnormal behavior sensor 10B when part of the behavior information detected at a predetermined position is abnormal may be performed.

(processing flow)
Next, the processing flow of the computing device 14 will be described. FIG. 6 is a flowchart for explaining the processing flow of the arithmetic device. As shown in FIG. 6, the computing device 14 acquires the behavior information by the behavior information acquisition unit 32 (step S10), and the first determination unit 34 determines that all the behavior information detected at the predetermined position is normal. (step S12). If it is determined that all the behavior information detected at the predetermined position is normal (step S12; Yes), the arithmetic device 14 allows the arithmetic unit 38 to update the existing second AI model N without updating the second AI model N. By using the model N, that is, by inputting the behavior information to the existing second AI model N, the estimation result of the road surface state is obtained (step S14). On the other hand, if all the behavior information detected at the predetermined position is not normal (step S12; No) and only some of the behavior information at the predetermined position is abnormal (step S16; Yes), the calculation unit 38 , using the existing second AI model N, that is, by inputting the behavior information to the existing second AI model N, while reducing the degree of influence on the road surface state of the behavior information of the behavior sensor determined to be abnormal. A state estimation result is acquired (step S18). Further, when all the behavior information detected at the predetermined position is not normal (step S12; No) and only a part of the behavior information at the predetermined position is not abnormal (step S16; No), that is, when it is detected at the predetermined position If all of the behavior information obtained is abnormal, the arithmetic device 14 determines whether the behavior information detected in the second area is normal by the second determination unit 36 (step S20). If the behavior information detected in the second area is determined to be normal (step S20; Yes), the process proceeds to step S14, and the calculation unit 38 uses the existing second AI model N to acquire the estimation result of the road surface condition. (step S14). On the other hand, when it is determined that the behavior information detected in the second area is not normal (step S20; No), the arithmetic unit 14 updates the second AI model N based on the behavior information (step S22).

(effect)
The computing device 14 according to the present embodiment includes a behavior information acquisition section 32 , a first determination section 34 , a second determination section 36 and a learning section 30 . The behavior information acquisition unit 32 acquires behavior information indicating the behavior of the vehicle 10 detected by the behavior sensor 10B provided in the vehicle 10 traveling on the road. Based on the behavior information detected at the predetermined position and the reference behavior information detected in the first region overlapping the predetermined position, the first determination unit 34 determines the behavior of the vehicle indicated by the behavior information detected at the predetermined position and the first determination unit 34 . It is determined whether or not the behavior of the vehicle indicated by the reference behavior information detected in one area corresponds. Note that the reference behavior information refers to the behavior of the vehicle 10 detected prior to the behavior information. When it is determined that the behavior of the vehicle indicated by the behavior information detected at the predetermined position does not correspond to the behavior of the vehicle indicated by the reference behavior information detected in the first region, the second determination unit 36 selects the first region. vehicle behavior indicated by the behavior information detected in the second region based on a plurality of pieces of behavior information detected in a second region wider than the second region and a plurality of pieces of reference behavior information detected in the second region; It is determined whether or not the behavior of the vehicle indicated by the reference behavior information corresponds. The learning unit 30 causes the second AI model N (AI model) to machine-learn the correspondence relationship between the behavior of the vehicle 10 and the road surface condition. When it is determined that the behavior of the vehicle indicated by the behavior information detected in the second area does not correspond to the behavior of the vehicle indicated by the reference behavior information, the learning unit 30 changes the behavior information detected in the second area. The second AI model N is updated as teacher data.

Here, when estimating the road surface state from the vehicle behavior using the AI model, if the characteristics of the side that detects the vehicle behavior (here, the vehicle 10 and the behavior sensor 10B) change, the AI model is learned. Since the tendency of the vehicle behavior will change, there is a possibility that the estimation accuracy of the road surface state using the AI model will decrease. On the other hand, if behavior information detected at a predetermined position is not normal and a plurality of pieces of behavior information detected in a wider second area are not normal, the arithmetic device 14 according to the present embodiment The behavior information detected at that time is used to update the second AI model N. Therefore, according to the present embodiment, it is possible to appropriately detect that the characteristics of the vehicle 10 have changed, and to update the second AI model N when the characteristics of the vehicle 10 have changed. It is possible to suppress the deterioration of the estimation accuracy of the road surface condition.

In addition, when it is determined that the behavior information detected at the predetermined position corresponds to the reference behavior information detected in the first area, and when the behavior information detected in the second area and the reference behavior information If it is determined that the second AI model N corresponds to the behavior information, the update of the second AI model N using the behavior information is not executed. That is, in the present embodiment, when the behavior information detected at the predetermined position and the second area is normal, the existing second AI model N is used because the characteristics of the vehicle 10 have not changed, and If the behavior information detected in the second area is not normal, it is determined that the characteristics of the vehicle 10 have changed, and the second AI model N is updated. Therefore, according to the present embodiment, it is possible to use the second AI model N adapted to changes in the characteristics of the vehicle 10, thereby suppressing a decrease in the estimation accuracy of the road surface state.

Further, the first determination unit 34 and the second determination unit 36 input the behavior information to the first AI model M (learning model) capable of determining whether the behavior information corresponds to the reference behavior information, thereby Determine whether the information corresponds to reference behavior information. By using the first AI model M for behavior information abnormality determination, the accuracy of behavior information abnormality determination can be improved, and characteristic changes of the vehicle 10 can be detected with high accuracy.

In addition, the arithmetic device 14 further includes an arithmetic unit 38 that inputs the behavior information to the second AI model N and obtains the estimation result of the road surface condition. By using the second AI model N, it is possible to suppress a decrease in the estimation accuracy of the road surface state.

Also, the behavior information acquisition unit 32 acquires behavior information detected by each of the plurality of behavior sensors 10B provided in the vehicle 10 . The first determination unit 34 determines whether the behavior information detected at the predetermined position and the reference behavior information detected in the first area correspond to each of the plurality of behavior sensors 10B. When only part of the behavior information detected at a predetermined position does not correspond to the reference behavior information, the calculation unit 38 detects the road surface state based on the behavior information of the behavior sensor 10B (abnormal sensor) whose behavior information does not correspond to the reference behavior information. The estimation result of the road surface condition is obtained so that the degree of influence on the road surface condition is lower than the degree of influence on the road surface condition by the behavior information of the behavior sensor 10B other than the abnormality sensor. According to the present embodiment, when the characteristics of some of the behavior sensors 10B are changed, the degree of influence of the behavior sensors 10B is reduced, so it is possible to suppress deterioration in the estimation accuracy of the road surface state.

The second AI model N is the first model to which the behavior information detected by the behavior sensor 10B (abnormal sensor) whose behavior information is not normal is input, and the behavior information detected by the behavior sensor 10B other than the abnormal sensor is input. and a second model that is The calculation unit 38 inputs the first estimation result of the road surface state obtained by inputting the behavior information detected by the abnormality sensor into the first model, and the behavior information detected by the behavior sensor other than the abnormality sensor into the second model. The degree of influence of the first estimation result on the estimation result of the road surface condition is higher than the degree of influence of the second estimation result on the estimation result of the road surface condition based on the second estimation result of the road surface condition acquired by inputting The estimation result of the road surface condition is calculated so as to be low. According to the present embodiment, when the characteristics of some of the behavior sensors 10B have changed, the road surface condition It is possible to suppress the decrease in the estimation accuracy of

Although the embodiments and examples of the present invention have been described above, the embodiments are not limited by the contents of these embodiments. In addition, the components described above include those that can be easily assumed by those skilled in the art, those that are substantially the same, and those within the so-called equivalent range. Furthermore, the components described above can be combined as appropriate. Further, various omissions, replacements, or modifications of components can be made without departing from the scope of the above-described embodiments.

1 detection system 10 vehicle 10B behavior sensor 12 measurement data acquisition device 14 arithmetic device 30 learning unit 32 behavior information acquisition unit 34 first judgment unit 36 second judgment unit 38 arithmetic unit M first AI model (learning model)
N 2nd AI model (AI model)

Claims (5)

1. a behavior information acquisition unit that acquires behavior information indicating the behavior of the vehicle detected by a behavior sensor provided in the vehicle traveling on the road;
   the behavior information detected at a predetermined position; and first reference behavior information, which is vehicle behavior detected in a first region overlapping the predetermined position, before the behavior information detected at the predetermined position. a first determination unit that determines whether the behavior of the vehicle indicated by the behavior information detected at the predetermined position corresponds to the behavior of the vehicle indicated by the first reference behavior information,
   If it is determined that the behavior of the vehicle indicated by the behavior information detected at the predetermined position does not correspond to the behavior of the vehicle indicated by the first reference behavior information, the first The behavior information detected within a second area wider than the area, and the second reference behavior information detected within the second area before the behavior information detected within the second area. a second determination unit that determines whether the behavior of the vehicle indicated by the behavior information detected in the second area corresponds to the behavior of the vehicle indicated by the second reference behavior information,
   a learning unit that causes an AI model to perform machine learning of a correspondence relationship between the behavior of the vehicle and the road surface condition of the road;
   has
   If it is determined that the behavior of the vehicle indicated by the behavior information detected in the second area does not correspond to the behavior of the vehicle indicated by the second reference behavior information, the learning unit updating the AI model using the detected behavior information as teacher data;
   Arithmetic unit.
2. The behavior information acquisition unit acquires the behavior information detected by each of the plurality of behavior sensors provided in the vehicle,
   The first determination unit determines whether the behavior of the vehicle indicated by the behavior information detected at the predetermined position corresponds to the behavior of the vehicle indicated by the first reference behavior information for each of the plurality of behavior sensors. death,
   The computing unit determines that, when only the behavior of the vehicle indicated by the part of the behavior information detected at the predetermined position does not correspond to the behavior of the vehicle indicated by the first reference behavior information, the behavior information indicates the first reference behavior. The degree of influence on the road surface state by the behavior information of the abnormal sensor that is the behavior sensor that does not correspond to the information is lower than the degree of influence on the road surface state by the behavior information of the behavior sensor other than the abnormal sensor. 2. The arithmetic unit according to claim 1, wherein the estimation result of the road surface condition is obtained such that
3. The AI model includes a first model to which the behavior information detected by the abnormality sensor is input, and a second model to which the behavior information detected by the behavior sensor other than the abnormality sensor is input,
   The calculation unit inputs the behavior information detected by the abnormality sensor into the first model to obtain a first estimation result of the road surface state, and the behavior sensor other than the abnormality sensor into the second model. and the degree of influence of the first estimation result on the estimation result of the road surface condition, based on the second estimation result obtained by inputting the behavior information detected in 3. The arithmetic unit according to claim 2, wherein the estimation result of the road surface condition is calculated so as to be lower than the degree of influence of the estimation result of the road surface condition.
4. a step of acquiring behavior information indicating the behavior of the vehicle detected by a behavior sensor provided in the vehicle traveling on the road;
   the behavior information detected at a predetermined position; and first reference behavior information, which is vehicle behavior detected in a first region overlapping the predetermined position, before the behavior information detected at the predetermined position. determining whether the vehicle behavior indicated by the behavior information detected at the predetermined position and the vehicle behavior indicated by the first reference behavior information correspond to each other;
   When it is determined that the behavior information detected at the predetermined position does not correspond to the first reference behavior information, detection is performed within a second area that includes the entire first area and is wider than the first area. detected in the second area based on the behavior information detected in the second area and second reference behavior information detected in the second area before the behavior information detected in the second area a step of determining whether the behavior of the vehicle indicated by the behavior information indicated by the second reference behavior information corresponds to the behavior of the vehicle indicated by the second reference behavior information;
   causing an AI model to perform machine learning of a correspondence relationship between the behavior of the vehicle and the surface condition of the road;
   has
   In the step of causing the AI model to perform machine learning, if it is determined that the behavior of the vehicle indicated by the behavior information detected in the second region does not correspond to the behavior of the vehicle indicated by the second reference behavior information, , updating the AI model using the behavior information detected in the second region as teacher data;
   Arithmetic method.
5. a step of acquiring behavior information indicating the behavior of the vehicle detected by a behavior sensor provided in the vehicle traveling on the road;
   the behavior information detected at a predetermined position; and first reference behavior information, which is vehicle behavior detected in a first region overlapping the predetermined position, before the behavior information detected at the predetermined position. determining whether the vehicle behavior indicated by the behavior information detected at the predetermined position and the vehicle behavior indicated by the first reference behavior information correspond to each other;
   When it is determined that the behavior information detected at the predetermined position does not correspond to the first reference behavior information, detection is performed within a second area that includes the entire first area and is wider than the first area. detected in the second area based on the behavior information detected in the second area and second reference behavior information detected in the second area before the behavior information detected in the second area a step of determining whether the behavior of the vehicle indicated by the behavior information indicated by the second reference behavior information corresponds to the behavior of the vehicle indicated by the second reference behavior information;
   causing an AI model to perform machine learning of a correspondence relationship between the behavior of the vehicle and the surface condition of the road;
   A program that causes a computer to execute
   In the step of causing the AI model to perform machine learning, when it is determined that the behavior of the vehicle indicated by the behavior information in the second region does not correspond to the behavior of the vehicle indicated by the second reference behavior information, the second updating the AI model using the behavior information detected in the area as teacher data;
   program.

Images