CN115398062A - Method for generating learned model and road surface feature determination device - Google Patents

Abstract

In the method for generating a learned model, the road surface characteristics are learned in advance by machine learning using, as teacher data, sample data including data extracted from vibration information using at least one frequency band, and the learned model is generated. The road surface feature determination device (1) is provided with the learned model, and is provided with a determination unit (5) for determining the road surface feature for the input of new data having the same type as the sample data.

Description

Method for generating learned model and road surface feature determination device

Technical Field

The present invention relates to a method for generating a learned model and a road surface feature determination device.

Background

Roads are damaged by the load effect of the wheels of the vehicle, temperature changes or the influence of rain water, and cracks, ruts, road surface web cracks or potholes are generated. In order to quickly grasp such a damage situation of the road and maintain and manage the road, an operation for grasping road surface characteristics of the road is performed on a daily basis.

As a method for grasping road surface characteristics, the following methods are generally used: an information collecting device for collecting information for grasping irregularities of a road surface is provided in advance in a test vehicle, and when the test vehicle travels on a road, the road surface characteristics are grasped by analyzing the information collected by the information collecting device. For example, as disclosed in JP2019-108755A, an information collecting device is provided with an acceleration sensor for detecting vibration due to a road surface on which a detection vehicle travels, an imaging device such as a digital camera for acquiring scan data of the traveling road surface, and a laser scanner, and collects information about a state of the road surface.

In this way, information obtained when the detection vehicle travels on the road is transmitted to the server via the internet communication network or the like, and is stored in the server and used for the subsequent analysis by the road surface characteristic determination device. In the road surface characteristic determination device, for example, scan data is analyzed to determine an International Roughness Index (hereinafter abbreviated as "IRI") and determine the depth of a rut and the presence or absence of a crack. The international roughness index is an evaluation index of unevenness of a paved road surface, in which a virtual vehicle model obtained by extracting only 1 wheel of a 2-axle 4-wheel vehicle is defined as a quarter-wheel vehicle, and a ratio of an accumulated value of vertical movement displacement received by the vehicle when the quarter-wheel vehicle travels on the road surface at a speed of 80km per hour to a travel distance is defined as roughness of the road surface.

The IRI value is used to determine whether a road with a large road width needs to be repaired, and a local government that manages the road is currently acquiring IRI value data and applying for repair to the country.

Documents of the prior art

Patent document

Patent document 1: JP2019-108755A

Disclosure of Invention

Problems to be solved by the invention

As described above, in order to acquire an IRI value, it is necessary to acquire and analyze scanned image data, and it is costly to acquire and analyze image data.

Therefore, an object of the present invention is to provide a method for generating a learned model and a road surface characteristic determination device that can determine road surface characteristics at low cost and with high accuracy.

Means for solving the problems

In order to achieve the above object, a method for generating a learned model according to the present invention generates a learned model by learning road surface characteristics in advance by machine learning using, as teacher data, sample data including data extracted using at least one or more frequency bands from vibration information detected while a vehicle is traveling. In the method for generating a learned model configured as described above, the vibration in the frequency band that is closely related to the deterioration of the road surface feature is given as the teacher data, and the determination unit performs machine learning on the road surface feature to generate the learned model.

Further, the method for generating a learned model includes: a vibration information collection step of collecting vibration information by a vibration information collection device while the vehicle is traveling; and a road surface feature learning step of generating a learned model by learning road surface features in advance by machine learning using, as teacher data, sample data including data extracted from the vibration information using at least one or more frequency bands. In the learned model generation method configured in this way, the learned model can be generated simply by two devices.

The road surface feature determination device of the present invention is a road surface feature determination device for determining a road surface feature of a road on which a vehicle is traveling from vibration information detected while the vehicle is traveling, and includes a learned model obtained by learning a road surface feature in advance by machine learning using, as teacher data, sample data including data extracted from the vibration information by using at least one or more frequency bands, and a determination unit for determining the road surface feature based on the learned model by inputting new data having the same type as the sample data. In the road surface characteristic determination device configured as described above, the determination unit gives, as the teacher data, the vibration in the frequency band closely related to the deterioration of the road surface characteristic, and the determination unit performs machine learning on the road surface characteristic, so that the determination of the road surface characteristic can be performed with high accuracy after the machine learning. Further, since the road surface feature determination device determines the road surface feature using new data having the same type as the sample data obtained from the vibration information, it is not necessary to perform high-cost analysis such as analysis of image data after scanning the road surface unevenness when performing the road surface feature determination process.

In order to achieve the above object, another road surface characteristic determination device according to the present invention is a road surface characteristic determination device for determining a road surface characteristic of a road on which a vehicle is traveling from vibration information detected while the vehicle is traveling, the road surface characteristic determination device including a learned model generated by machine learning using, as teacher data, sample data including data extracted from the vibration information by using at least one or more frequency bands, and including a determination section for performing determination of the road surface characteristic based on the learned model with respect to input of new data having the same type as the sample data. The road surface characteristic determination device configured as described above includes a learned model, and the learned model determination unit generates vibration in a frequency band closely related to deterioration of the road surface characteristic as teacher data by machine learning, and can perform determination of the road surface characteristic with high accuracy based on the learned model. Further, since the road surface feature determination device determines the road surface feature using new data having the same type as the sample data obtained from the vibration information, it is not necessary to perform high-cost analysis such as analysis of image data after scanning the road surface unevenness and visual work by the inspector when performing the determination processing of the road surface feature.

Further, the determination unit in the road surface characteristic determination device may include a plurality of item learning units that learn the road surface characteristics only for corresponding items, corresponding to each of a plurality of items indicating the road surface characteristics, and the item learning unit may determine the road surface characteristics only for the items corresponding to the input of new data. According to the road surface feature determination device configured as described above, since the item learning unit is provided that performs machine learning and determination exclusively for items corresponding to each item representing road surface features at different scales, it is possible to perform determination with high accuracy for each item of road surface features.

Further, the item learning unit in the road surface characteristic determination device may include a plurality of learning units classified according to the running conditions, and the learning units classified according to the running conditions may machine-learn the road surface characteristics for each running condition of the vehicle and for each corresponding running condition. According to the road surface characteristic determination device configured in this way, since machine learning and determination conforming to the driving condition are performed for each item indicating the road surface characteristic, it is possible to perform highly accurate determination for each item of the road surface characteristic even if the driving condition is different.

The road surface feature determination device may further include a determination unit configured to determine whether or not the new data is abnormal, based on a reference point having an average value of sample data of the same tag as a coordinate and the new data determined to belong to the same tag. The road surface characteristic determination device configured in this way can present the necessity of the determination section to perform machine learning again and the necessity of a new tag to the operator of the road surface characteristic determination device.

The road surface feature determination device may further include an input data calculation unit that cuts out data of a predetermined amount of time from the original data of the vibration information, processes the cut data of the predetermined amount of time to obtain new data, and inputs the new data to the determination unit, and the input data calculation unit may perform an overlap process for overlapping the data cut out when the previous new data is obtained and a part of the data cut out when the next new data is obtained and sequentially obtains the new data. In this way, the road surface feature determination device performs the superimposition processing, and therefore can grasp the road surface feature for each time segment, and perform machine learning and determination using sample data and new data that are not missing even in the vicinity of the time division line. Therefore, the road surface characteristics can be determined with high accuracy.

In the road surface characteristic determination device, the vibration information may include acceleration of a sprung member of the vehicle. According to the road surface characteristic determination device configured in this way, the road surface characteristic can be determined using the sprung acceleration that is directly related to the vehicle riding comfort, and the road surface characteristic can be determined from the viewpoint of the vehicle riding comfort.

In the road surface characteristic determination device, the sample data and the new data may include a total local vibration value extracted using at least one frequency band. According to the road surface feature determination device configured as described above, since the value clearly indicating the in-band vibration intensity, that is, the local vibration total magnitude value having a strong correlation with the road surface feature is included as the feature amount in the sample data and the new data, the accuracy of determining the road surface feature by the determination section is improved.

Further, in the road surface characteristic determination device, the sample data and the new data may include a value obtained by dividing the unsprung acceleration of the vehicle by the speed of the vehicle. According to the road surface characteristic determination device configured in this way, the road surface characteristic can be determined with high accuracy regardless of the traveling speed of the vehicle.

In the road surface characteristic determination device, the sample data and the new data may include a value obtained by dividing a maximum value of the unsprung acceleration of the vehicle by an effective value of the unsprung acceleration. According to the road surface characteristic determination device configured in this way, it is possible to grasp more accurately whether the road surface is a road surface characteristic of global roughness or local roughness, and to determine the road surface characteristic with high accuracy.

Further, in the road surface characteristic determination device, the sample data and the new data may include a maximum value, an average value, a median value, a minimum value, a variance, and a standard deviation of sprung components of the vehicle. According to the road surface characteristic determination device configured in this way, the road surface characteristic can be machine-learned and determined using the index representing the vibration information of the sprung member.

Effects of the invention

According to the method for generating a learned model and the road surface feature determination device of the present invention, the determination of the road surface feature can be performed at low cost and with high accuracy.

Drawings

Fig. 1 is a diagram showing a system configuration of a road surface characteristic determination device in a first embodiment.

Fig. 2 is a diagram illustrating a sensor unit in the vibration information collecting device.

Fig. 3 is a diagram showing a hardware configuration of a controller in the vibration information collecting apparatus.

Fig. 4 is an exemplary diagram showing a data structure that associates positional information stored in the database with road surface characteristic items.

Fig. 5 is a diagram illustrating processing of raw data input to the data calculation unit.

Fig. 6 is a diagram showing an example of a machine learning process of learning with sample data of the road surface characteristic determination device in the first embodiment.

Fig. 7 is an exemplary diagram showing a determination process of processing new data of the road surface characteristic determination device in the first embodiment.

Fig. 8 is a diagram showing a system configuration of an item learning unit of the road surface characteristic determination device according to the second embodiment.

Detailed Description

First embodiment

The present invention will be described based on embodiments shown in the drawings. As shown in fig. 1, a road surface characteristic determination device 1 according to a first embodiment includes: a communication unit 2 that can communicate with a vibration information collection device 10 that collects vibration information; an input data calculation unit 3 that processes raw data of the vibration information received via the communication unit 2; a database 4 that stores road surface characteristics associated with the location information; a determination unit 5 that includes a plurality of item learning units 5a, 5b, 5c, 5d, 5e, and 5f and determines road surface characteristics; a display device 6 that displays the road surface characteristics determined by the determination unit 5; and a printing device 7 capable of printing the determination result of the road surface characteristics determined by the determination unit 5 on a paper medium.

Next, the road surface characteristic determination device 1 will be described in detail. The road surface characteristic determination device 1 processes vibration information transmitted from the vibration information collection device 10 that collects vibration information, and determines a road surface characteristic.

The vibration information collection device 10 is mounted on a vehicle V, which is a four-wheeled vehicle, collects vibration information, and transmits the collected vibration information to the road surface characteristic determination device 1 provided in a base station. As shown in fig. 1, the vibration information collection device 10 includes: a storage unit 11 capable of storing information; a sensor unit 12 that detects vibration information of the vehicle V; a controller 13 that stores vibration information detected by the sensor portion 12 and vehicle information, which is information related to another vehicle, in the storage portion 11; and a communication unit 14 controlled by the controller 13 and transmitting information such as vibration information stored in the storage unit 11 to the road surface characteristic determination device 1.

The vibration information collection device 10 operates by receiving electric power supplied from a battery, not shown, of the vehicle V. When an ignition switch, not shown, is turned on, the vibration information collection device 10 receives electric power supplied from a battery, collects vibration information detected by the sensor unit 12 and the vehicle information, and stores the vibration information and the vehicle information in the storage unit 11; when the ignition switch is turned off, the vibration information and other information stored in the storage unit 11 are transmitted to the road surface characteristic determination device 1 through the communication unit 14.

The storage unit 11 is controlled by the controller 13, and stores the vibration information and information of other vehicles according to instructions from the controller 13. When the start-up processing is completed, the controller 13 associates the data of the vibration information and the vehicle information with the time when the data is obtained every predetermined information storage time, generates a file in which each piece of data collected at the same sampling rate is collected, and stores the file in the storage unit 11.

The storage unit 11 is configured by, for example, a nonvolatile semiconductor memory such as a flash memory, but is not limited to the flash memory and may be a magnetic disk or the like. The storage unit 11 may include an auxiliary storage device including a storage medium such as an optical disk and a drive capable of reading and writing data from and to the storage medium.

The communication unit 14 is controlled by the controller 13, and is capable of communicating with the road surface characteristic determination device 1 provided outside, and transmits the vibration information and the raw data of other vehicle information stored in the storage unit 11 to the road surface characteristic determination device 1. The communication unit 14 includes an antenna unit not shown in the drawings, and in the present embodiment, performs wireless LAN (Local Area Network) communication in the IEEE802.11 standard, but may communicate with the road surface characteristic determination device 1 via a telephone line or an internet communication Network.

As shown in fig. 2, the sensor unit 12 includes, as a sensor for detecting vibration information of the vehicle V: 4 acceleration sensors 15a, 15b, 15c, 15d that respectively detect sprung acceleration as acceleration in the vertical direction directly above each of the four wheels of the vehicle V; 4 acceleration sensors 16a, 16b, 16c, 16d that detect unsprung acceleration as acceleration in the vertical direction of each of the four wheels; stroke sensors 17a, 17b, 17c, 17d that detect stroke displacements of each of the four wheels with respect to the vehicle body; and a three-axis gyro sensor 19 mounted on the body of the vehicle V and detecting angular velocities of the body rotating about three axes of front-rear, left-right, up-down; in addition to the above sensors, a position detection device 18 for detecting position information of the vehicle V is provided. The position information includes a longitude and a latitude, and the position detection device 18 detects the longitude and the latitude as the position information of the vehicle V using the global positioning satellite system.

In the present embodiment, the vibration information collecting device 10 collects, in addition to the vibration information detected by the sensor unit 12, vehicle information such as the speed, the engine water temperature, the accelerator opening degree, the brake operation, and the wiper driving state of the vehicle V detected by an in-vehicle sensor or the like mounted in advance on the vehicle V through an OBD (on-board diagnostic) terminal, not shown, of the vehicle V. The vibration information is information related to vibration of the vehicle V, such as acceleration, velocity, displacement, angular acceleration in the vehicle body rotation direction, angular velocity, and angular displacement of the vehicle body, wheels, and suspension, and does not include image data obtained by scanning irregularities on the road surface. The vibration information collection device 10 may be arbitrarily selected from the plurality of types of vibration information as long as the vibration information of the vehicle V required for the determination by the road surface characteristic determination device 1 can be obtained. Although not included in the vibration information, the vibration information collection device 10 collects the position information as described above in order to grasp the position of the road surface. The vibration information collection device 10 may include a camera, not shown, for capturing an image of the road surface, and may collect image data together with the vibration information so as to assist the operator of the road surface characteristic determination device 1 in confirming the road surface characteristics.

As shown in fig. 3, the controller 13 includes a CPU (Central Processing Unit) 13a, a memory 13b, an interface 13c, and a bus 13d that communicatively connects these devices. The controller 13 is connected to the storage unit 11 and the communication unit 14 via the bus 13d so as to be able to communicate with each other. The controller 13 is connected to the sensor unit 12 via the interface 13c so as to be able to receive information detected by the sensor unit 12, and is also able to receive information detected by an in-vehicle sensor via the interface 13c and an OBD terminal, not shown.

The CPU13a controls the storage unit 11 and the communication unit 14 in the vibration information collection device 10 by executing an operating system and other programs, and processes various information detected by the sensor unit 12 and the in-vehicle sensors. The Memory 13b includes a RAM (Random Access Memory) for providing a Memory area necessary for the arithmetic processing of the CPU13a in addition to a ROM (Read Only Memory), and stores a program used for the arithmetic processing of the CPU13a in the ROM. Further, a program used for the arithmetic processing of the CPU13a may be stored in the storage unit 11.

The controller 13 executes a program necessary for the CPU13a to function as the vibration information collection device 10, and temporarily stores data of various information detected by the sensor unit 12 and the in-vehicle sensor in a buffer memory fixed to the memory 13b, processes the data of the various information stored in the buffer memory to generate a file, stores the file in the storage unit 11, and further transmits the file stored in the storage unit 11 from the communication unit 14 to the external road surface characteristic determination device 1 (vibration information collection step).

As described above, when the ignition switch is turned on and the start-up processing is completed, the controller 13 stores data obtained by the sensors having the same sampling rate among the data of various information detected by the sensor unit 12 and the in-vehicle sensors in one file. More specifically, the controller 13 executes a recorder process of associating data detected by the sensor unit 12 and the in-vehicle sensor with the time when the data is obtained, and generating one file by summarizing data obtained by the sensors at the same sampling rate during each predetermined information storage time for the information storage time, and storing the data in the storage unit 11. In the present embodiment, the information storage time is 1 minute, but may be set to a time other than 1 minute.

The controller 13 in the present embodiment stores each of the data in the vibration information and the vehicle information in association with time in the buffer memory, and stores the bundle of data in the vibration information and the vehicle information obtained within 1 minute among the data stored in the buffer memory as a separate file for each data obtained by the sensors having the same sampling rate in the storage section 11.

Next, when the ignition switch is turned off, the controller 13 ends collection of various information obtained from the sensor unit 12 and the in-vehicle sensor, and transmits a file storing various information to the road surface characteristic determination device 1 via the communication unit 14.

The controller 13 executes the transmission processing of the files in this manner, and when the transmission of all the files stored in the storage section 11 is completed, the processing is ended and closed.

As described above, the road surface characteristic determination device 1 that receives the file from the vibration information collection device 10 is provided in the base of the vehicle V. In the present embodiment, the road surface characteristic determination device 1 processes the file received from the vibration information collection device 10 while storing the file. As described above, the road surface characteristic determination device 1 according to the first embodiment includes: a communication unit 2 that can communicate with a vibration information collection device 10 that collects vibration information; an input data calculation unit 3 that processes raw data of the vibration information received via the communication unit 2; a database 4 that stores road surface characteristics associated with the location information; a determination unit 5 that includes a plurality of item learning units 5a, 5b, 5c, 5d, 5e, and 5f and determines road surface characteristics; a display device 6 that displays the road surface characteristics determined by the determination unit 5; and a printing device 7 capable of printing the determination result of the road surface characteristics determined by the determination unit 5 on a paper medium. The determination unit 5 performs machine learning using sample data obtained in advance by the input data calculation unit 3 as teacher data, and determines road surface characteristics for input of new data after the machine learning using the sample data.

As shown in fig. 1, the hardware of the road surface characteristic determination device 1 is a computer system, and includes a CPU20, a storage device 21, an input device 22 such as a keyboard and a mouse, a display device 6, a printing device 7, a communication unit 2, and a bus 23 for connecting these devices so as to be communicable with each other.

The CPU20 performs an arithmetic process by executing an operating system and other programs, and controls the storage device 21, the input device 22, the communication unit 2, the database 4, the display device 6, and the printing device 7. The CPU20 realizes the input data calculation unit 3 and the determination unit 5 by executing a program for functioning as the road surface characteristic determination device 1.

A parking space for the vehicle V is provided in a base station in which the road surface characteristic determination device 1 is provided, and a communication unit 2 capable of mutually performing wireless LAN communication with the communication unit 14 in the vibration information collection device 10 is provided in the vicinity of the parking space. The communication unit 2 is controlled by the CPU20, can communicate with the vibration information collection device 10, and receives the vibration information and the raw data of the vehicle information transmitted from the vibration information collection device 10. The communication unit 2 includes an antenna unit not shown in the drawings, and performs wireless LAN (Local Area Network) communication in IEEE802.11 standard in the present embodiment, as in the case of the communication unit 14, but may communicate with the vibration information collection device 10 via a telephone line or an internet communication Network. When the distance between the parking space and the road surface characteristic determination device 1 is long, the communication unit 2 may include a relay, not shown, and may be capable of performing wireless communication with the vibration information collection device 10 via the relay.

The storage device 21 includes a hard disk in addition to ROM and RAM. Further, the storage device 21 provides a storage area necessary for processing in the CPU20 while storing programs necessary for control of the database 4 and the road surface characteristic determination device 1. The storage device 21 may include an auxiliary storage device or a semiconductor memory including a storage medium such as a magnetic disk or an optical disk and a drive capable of reading and writing data of the storage medium. The display device 6 includes a screen for displaying data processed by the CPU20, and is, for example, a liquid crystal display. The printing device 7 is, for example, a printer or the like that prints data or the like processed by the CPU20 on a paper medium.

When receiving the file of the data of the vibration information and the vehicle information from the vibration information collection device 10 through the communication section 14, the CPU20 stores the received file in the storage device 21.

Further, the database 4 stored in the storage device 21 stores point information and data of road surface characteristics associated with the point information. The road surface characteristics are a general term of road surface deterioration or damage caused by a vehicle running load or the like over time, and are grasped by various quantitative indices. Specifically, the road surface characteristics can be grasped by various indexes such as road surface characteristic values quantitatively indicating road surface conditions such as roughness, cracks, and ruts, which are different phenomena. For example, the road surface characteristics are obtained by items representing road surface characteristics such as IRI, rut depth, crack rate, flatness, MCI, presence or absence of a pit, and the like. IRI is an evaluation index relating to the unevenness of a paved road surface, and is an index indicating that the larger the value, the larger the unevenness of a paved road surface, and also an index indicating the state of road surface roughness, which is one of road surface characteristics. Generally, on a highway, when the IRI value exceeds 4m/km, it is judged that the road surface is damaged. The rut depth is a value indicating the rut depth of the road surface, and is also a value indicating the rut depth, which is one of road surface characteristics. The crack rate is a value obtained by dividing the crack area by the area of the survey object, and is a value indicating the degree of cracks, which is one of the road surface characteristics. The flatness is a standard deviation value from an average value of a difference in level between a profile of a vertical section to be measured and a paved road surface assumed to be flat, and is a value representing the degree of flatness which is one of road surface characteristics. MCI (Maintenace Control Index), which is called a road surface maintenance Index, is an Index for quantitatively evaluating the service performance of a road surface from 3 road surface characteristic values of crack rate, rut depth, and flatness, and is also a comprehensive Index for performing maintenance judgment as one of road surface characteristics. More specifically, the MCI is divided into 10 highest points, decreases with the deterioration of the road surface, and is obtained from a matrix table of a crack rate and a rutting depth prepared for each flatness value. The pits are generated by the enlargement of the damaged surface due to the deterioration of the road surface under the action of cracks generated due to the deterioration over time or water and traffic load permeating from road surface joints and the like, which significantly reduces the safety performance of passing vehicles, and the presence or absence of these pits indicates the degree of danger of the road surface characteristics when the vehicles pass. In order to grasp the road surface characteristics in this way, it is necessary to refer to various indexes and quantified values that are useful for determining whether or not road repair is necessary.

When an index or a value indicating the road surface characteristic at the road position is found in advance, the database 4 stores data associated with each index or value indicating the road position and the road surface characteristic. Specifically, for example, as shown in fig. 4, coordinates for specifying a road position are registered in the database 4 in association with all the items of the road surface characteristics.

The input data calculation unit 3 processes the vibration information and the raw data of the vehicle information received via the communication unit 2. The input data calculation unit 3 is realized by the CPU20 executing a program and executing the processing of the input data calculation unit 3. The input data calculation unit 3 extracts and processes the original data of the vibration information of the sprung acceleration, the unsprung acceleration, and the stroke displacement of the 1 minute component stored in the file, which is the vibration information of the 3 second component of the predetermined time, and then obtains the sample data and the new data input to the determination unit 5. The sample data is associated with position information obtained from the raw data, is teacher data provided by the determination unit 5 for machine learning of each item of road surface characteristics, is data obtained by the input data calculation unit 3 from vibration information obtained when the vehicle V travels on a road surface of which road surface characteristics are determined in advance, and is provided to the determination unit 5 in a state in which a label is attached to a road surface characteristic item. In each item of the road surface characteristics, the quantitative values of the road surface for obtaining the IRI, the rutting depth, the crack rate, the flatness and the MCI data are used as labels, whether a pit slot exists in the road surface for obtaining the pit slot data is used as a label to be associated with the data, and teacher data are generated. As described above, the additional label is a label that associates data with each item of road surface characteristics of the actual road surface from which the data is obtained.

On the other hand, the new data is associated with the position information, is data to which no tag is added to the road surface feature item, and is data which is input to the determination unit 5 by the determination unit 5 after the machine learns the road surface feature. That is, the new data is data obtained by processing the raw data collected by the vibration information collection device 10 by the input data calculation unit 3, but is data that is input to the determination unit 5 for determining each item of road surface characteristics without being associated with each item information of road surface characteristics.

In the determination unit 5, since the sample data is used as the teacher data and the machine learning is performed to determine each item of the road surface feature, the new data needs to be determined by the same type of data obtained by the same processing as the processing performed on the original data to obtain the sample data. Therefore, as will be described later in detail, in the present embodiment, the sample data and the new data are values such as 4 sprung acceleration, maximum value, minimum value, average value, median value, standard deviation, variance, POA value in a frequency band of 0.2Hz to 3Hz, POA value in a frequency band of 3Hz to 8Hz, POA value in a frequency band of 8Hz to 20Hz and POA value in a frequency band of 15Hz to 30Hz, a value obtained by dividing the unsprung acceleration by the speed of the vehicle V, and a value obtained by dividing the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration, which are obtained by subjecting the original data to the same processing. In addition, the sample data and the new data may include original data cut at a predetermined time without performing a process of obtaining a maximum value or a POA value of the original data.

In this way, the determination unit 5 performs machine learning on the road surface features in advance based on the sample data with the tag, and outputs the determination result for each item of the road surface features with respect to the input of new data.

The input data calculation unit 3 performs an overlap process for overlapping the original data cut out when the previous sample data or new data is obtained and a part of the original data cut out when the next sample data or new data to be input to the determination unit 5 is obtained, and sequentially obtains new data. Specifically, as shown in fig. 5, the overlap time is set to 2 seconds, and the input data calculation unit 3 sequentially generates sample data or new data from the raw data of 3 seconds, which is a predetermined time, while shifting the time every 1 second, and obtains 57 data sets from the raw data of 1 minute of vibration information. In the above example, the predetermined time taken by the input data calculation unit 3 from the raw data of the vibration information is set to 3 seconds, but may be set arbitrarily.

The input data calculation unit 3 processes the raw data of 4 sprung accelerations, 4 unsprung accelerations, and 4 stroke displacements detected by the sensor unit 12, which are included in the data set, obtains the maximum value, the minimum value, the average value, the median value, the standard deviation, and the variance for each of the 4 sprung accelerations, the unsprung accelerations, and the stroke displacements, and includes these values as feature quantities in the sample data and the new data.

Vibrations during vehicle travel cause vehicle occupants to feel different levels of discomfort in each frequency band depending on the intensity. For example, vibrations in the frequency band from 0.2Hz to 3Hz cause the rider to feel a flicking, vibrations in the frequency band from 3Hz to 8Hz cause the rider to feel unevenness, vibrations in the frequency band from 8Hz to 20Hz cause the rider to feel a dendendenuded pit, and vibrations in the frequency band from 15Hz to 30Hz cause the rider to feel an abnormal sound. Further, the rider feels different degrees of discomfort to the vibration having a vibration frequency of about 16Hz and the vibration having a vibration frequency of about 5 Hz. In this way, the rider feels different degrees of discomfort to vibrations in different frequency bands.

On a relatively flat road surface, when vibration is large in a frequency band from 0.2Hz to 3Hz and the rider feels a light drift, it is considered that the road surface is deteriorated, and when vibration is large in a frequency band from 3Hz to 8Hz and the rider feels an unevenness, it is considered that cracks may occur on the road surface. In this way, the uncomfortable feeling felt by the rider due to the vibration intensity of the vibration frequency band is caused by the road surface characteristics.

Therefore, the input data calculation unit 3 includes data extracted from the original data of 4 sprung accelerations, unsprung accelerations, and stroke displacements, which are 3 second components of a predetermined time, in the sample data and the new data using a predetermined frequency band. More specifically, the input data calculation unit 3 obtains the sum of power values of vibrations in a predetermined frequency band for 4 sprung acceleration, unsprung acceleration, and stroke displacement, and includes the sum in sample data and new data. That is, the input data calculation unit 3 calculates the power spectrum of the vibration by performing an operation of Fast Fourier Transform (FFT) analysis processing on each of the raw data of 4 sprung accelerations, 4 unsprung accelerations, and 4 stroke displacements by the CPU20, and calculates the sum of the power values (intensities) of the vibrations in the predetermined frequency band. The value obtained in this way by the input data calculation unit 3 is included as a feature amount in the sample data and the new data, and is input to the determination unit 5. The predetermined frequency band is a frequency band that gives a sense of discomfort to the occupant while the vehicle is traveling. It is considered that the vibration of the frequency band giving the driver a sense of discomfort while the vehicle is running is caused by the deterioration of the road surface characteristics, and the value obtained by the input data calculation unit 3 in this way becomes an index that contributes to the understanding of the road surface characteristics.

In the present embodiment, the input data calculation unit 3 performs FFT analysis processing on raw data of 4 sprung accelerations, 4 unsprung accelerations, and 4 stroke displacements to obtain a total local vibration amount (POA) value that is the sum of power values of a plurality of frequency bands. Specifically, the input data calculation unit 3 calculates the POA value in the frequency band from 0.2Hz to 3Hz, the POA value in the frequency band from 3Hz to 8Hz, the POA value in the frequency band from 8Hz to 20Hz, and the POA value in the frequency band from 15Hz to 30Hz, respectively, and inputs these POA values as feature quantities to the determination unit 5. Further, although the POA value is a sum of power values of vibrations in a frequency band, it is a value that can clearly indicate the vibration intensity in the frequency band, and is advantageous for the determination of the road surface characteristics by the determination unit 5 described later in that the degree of correlation with the road surface characteristics is highly expressed, a value indicating the characteristics of the frequency band may be obtained as a characteristic amount from data of the frequency band such as an average value of vibration amplitudes included in a predetermined frequency band, and used as sample data and new data.

Further, in the present embodiment, unlike the above-described values, the input data calculation unit 3 processes the raw data of 4 sprung accelerations and unsprung accelerations that are 3 second components of a predetermined time, calculates a value obtained by dividing the unsprung acceleration of the vehicle V by the velocity of the vehicle V obtained from the vehicle information and a value obtained by dividing the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration by calculation, and includes these two values in the sample data and the new data.

The road surface characteristics may not be accurately evaluated because the unsprung acceleration increases as the speed of the vehicle V increases. On the other hand, the influence of the speed of the vehicle V when traveling on the road surface can be alleviated by dividing the unsprung acceleration by the speed of the vehicle V, and in the present embodiment, the input data calculation unit 3 uses this as a feature amount and as sample data and new data. In order to mitigate the influence of the speed of the vehicle V when traveling on the road surface, a value obtained by dividing the maximum value of the absolute value of the unsprung acceleration by the speed of the vehicle V may be used, or a value obtained by dividing the absolute value of the unsprung acceleration by the speed of the vehicle V may be used by multiplying the value by a coefficient or the like.

Further, when the vehicle V passes through a damaged portion such as a pit, the unsprung portion of the vehicle V vibrates greatly, and therefore the maximum value of the unsprung acceleration increases, but the maximum value does not increase as an effective value of a certain period of time. In contrast, when the vehicle V continues to run on a rough road surface, both the maximum value and the effective value of the unsprung acceleration increase. Thus, since the value obtained by dividing the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration becomes an index indicating whether the road surface is entirely rough or locally rough, the input data calculation unit 3 in the present embodiment uses this as a feature amount and as sample data and new data. As an index indicating whether the road surface is entirely rough or partially rough, in addition to a value obtained by dividing the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration, a value obtained by multiplying the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration, a value obtained by dividing the effective value of the unsprung acceleration by the maximum value of the unsprung acceleration, a value obtained by dividing the minimum value of the unsprung acceleration by the effective value of the unsprung acceleration, or the like can be used.

As described above, the sample data and the new data in the road surface characteristic determination device 1 according to the present embodiment are values obtained by dividing the unsprung acceleration by the effective value of the unsprung acceleration, and values obtained by dividing the unsprung acceleration by the velocity of the vehicle V, for 4 values, such as the maximum value, the minimum value, the average value, the median value, the standard deviation, the variance, the POA value in the frequency band of 0.2Hz to 3Hz, the POA value in the frequency band of 3Hz to 8Hz, the POA value in the frequency band of 8Hz to 20Hz, and the POA value in the frequency band of 15Hz to 30 Hz.

Further, when the operator of the road surface feature determination device 1 inputs the label of each item of road surface features with respect to the data obtained when traveling on the road surface of each item of road surface features which are known, the input data calculation unit 3 associates the label with the data obtained by the above processing, generates sample data serving as teacher data, and supplies the sample data to the determination unit 5. For example, in each of the items of road surface characteristics of the road surface from which the data is obtained, the input data calculation unit 3 associates sample data with a sample data when the sample data of the road surface having an IRI of 3m/km, a rut depth of 10.0mm, a crack rate of 10%, a flatness of 3mm, an MCI of 5.0, and no pit is present, and when the IRI is 3m/km, the rut depth of 10mm, the crack rate of 10%, the flatness of 3mm, the MCI of 5.0, and no pit is present, these sample data are used as a label.

The determination unit 5 includes a plurality of item learning units 5a, 5b, 5c, 5d, 5e, and 5f that determine the road surface characteristics for each of the items described above. The respective item learning units 5a, 5b, 5c, 5d, 5e, and 5f of the determination unit 5 are considered to be artificial intelligence for performing machine learning in accordance with a learning model, and are realized by the CPU20 executing a program and executing the processing of the respective item learning units 5a, 5b, 5c, 5d, 5e, and 5f. The item learning unit 5a determines only IRI in the road surface feature items. Specifically, the item learning unit 5a performs machine learning by using the sample data as teacher data to determine IRI by machine learning, and outputs a value of IRI for input of new data after the machine learning. The item learning unit 5b determines only the depth of the rut in the road surface feature item. Specifically, the project learning unit 5b performs machine learning by using the sample data as teacher data to determine the rut depth through the machine learning, and outputs a value of the rut depth for input of new data after the machine learning. The item learning unit 5c determines only the crack rate in the road surface feature item. Specifically, the project learning unit 5c performs machine learning by using the sample data as teacher data to determine the crack rate through the machine learning, and outputs a value of the crack rate for inputting new data after the machine learning. The item learning unit 5d determines only the flatness in the road surface feature items. Specifically, the item learning unit 5d performs machine learning by using the sample data as teacher data to determine flatness through the machine learning, and outputs a flatness value for inputting new data after the machine learning. The item learning unit 5e determines only the MCI in the road surface feature items. Specifically, the project learning unit 5e performs machine learning by using the sample data as teacher data to determine MCI through the machine learning, and outputs a value of MCI for inputting new data after the machine learning. The item learning unit 5f determines only whether or not there is a pit in the road surface feature item. Specifically, the project learning unit 5f performs machine learning to determine whether or not there is a pit by machine learning using the sample data as teacher data, and determines whether or not there is a pit for inputting new data after the machine learning.

As described above, in order to perform machine learning on the road surface feature items corresponding to the respective item learning units 5a, 5b, 5c, 5d, 5e, and 5f, the input data arithmetic unit 3 inputs the sample data as teacher data to the respective item learning units 5a, 5b, 5c, 5d, 5e, and 5f, and causes the respective item learning units 5a, 5b, 5c, 5d, 5e, and 5f to perform machine learning.

The model for machine learning is a machine learning model for machine learning using teacher data, and the road surface feature determination device 1 according to the present embodiment is a recognizer such as a support vector machine, but may be a machine learning model using a learning algorithm such as back propagation or ID3 (Iterative Dichotomiser 3).

The item learning units 5a, 5b, 5c, 5d, 5e, and 5f receive input of sample data associated with the label for each item of the road surface feature generated by the input data calculation unit 3, and perform machine learning on the corresponding item. The number of sample data is the number of items that can be sufficiently machine-learned by the item learning units 5a, 5b, 5c, 5d, 5e, and 5f. When the machine learning based on the sample data is completed and a machine learning completed model used for determining the road surface feature is generated, the item learning units 5a, 5b, 5c, 5d, 5e, and 5f determine and output the corresponding items with respect to the input of new data that is the unlearned data not associated with the tag. The item learning units 5a, 5b, 5c, 5d, 5e, and 5f generate learned models by adjusting parameters for weighting respective values of sample data so that respective items of road characteristics can be estimated with high accuracy by machine learning. As described above, the learned model is an inference program in which the road surface characteristic determination device 1 incorporates learned parameters generated by learning each item representing the road surface characteristic by inputting sample data as teacher data. Here, the inference program is a program capable of outputting a determination result of each item of the road surface characteristics with respect to the input of the new data by applying the programmed learned parameters.

The identifier is an identifier for determining a hyperplane for classifying the label from the sample data, and identifying, using the hyperplane as a learned model, to which of two categories into which the unvarnished data belongs. Therefore, the item learning units 5a, 5b, 5c, 5d, and 5e have a plurality of discriminators therein in order to obtain the values of IRI, rut depth, crack rate, flatness, and MCI. For example, when the IRI is judged to be not more than 5m/km at a scale of 1m/km, the item learning unit 5a may judge the value of the IRI by making the identifier into a plurality of layers, for example, a recognizer judging whether it is less than 2.5m/km or not less than 2.5m/km, a recognizer judging whether it is less than 1.75m/km or not less than 1.75m/km, a recognizer judging whether it is less than 1m/km or not less than 1m/km, a recognizer judging whether it is less than 2m/km or not less than 2m/km, a recognizer judging whether it is less than 3m/km or not less than 3m/km, or not less than 4 m/km. In each of the discriminators included in the item learning unit 5a, the numerical value that divides the boundary between two categories is exemplary, and design changes can be made.

In this way, the project learning units 5a, 5b, 5c, 5d, and 5e set and classify values of IRI, rut depth, crack rate, flatness, and MCI, make a plurality of discriminators into a plurality of layers, perform machine learning by inputting sample data, and only need to obtain values of IRI, rut depth, crack rate, flatness, and MCI for inputting new data that is data that has not been learned after completion of learning. If a classification of fine values is desired, the recognizer may be added. Further, since the item learning unit 5f determines whether or not there is a pit, it is not necessary to make the identifier multilayered.

Instead of or in addition to the above items, the determination unit 5 may determine the presence or absence of a rut depth, the presence or absence of a crack, and the presence or absence of a road surface irregularity, and may determine whether or not road repair is required, and an item arbitrarily determined by the operator of the road surface characteristic determination device 1. The determination unit 5 may be configured to perform machine learning on a plurality of items, but may be provided with an item learning unit for each set item.

As described above, the determination unit 5 performs machine learning on each item of road surface features using the labeled sample data generated by the input data calculation unit 3 as teacher data, and generates a learned model (road surface feature learning step). When the machine learning is completed, the determination unit 5 determines to which label the new data belongs by comparing the learned model with respect to the input of the new data that is generated by the input data calculation unit 3 and is not learned, and outputs the new data. That is, if IRI is used, the item learning unit 5a in the determination unit 5 determines the value of IRI for the input of new data and outputs it, using the learned model after machine learning. Similarly, the project learning units 5b, 5c, 5d, and 5e output the corresponding rut depth, crack rate, flatness, and MCI values, respectively. The item learning unit 5f outputs whether or not there is a pit.

The road surface characteristic determination device 1 according to the first embodiment is configured as described above, and the operation thereof will be described below with reference to fig. 6 and 7. When receiving the file of the vibration information and the raw data of the vehicle information collected by the vibration information collection device 10, the road surface characteristic determination device 1 stores it in the storage device 21.

First, when the operator inputs an instruction to perform machine learning through the input device 22, the road surface characteristic determination device 1 processes raw data by the CPU20 executing a program for processing the raw data, and generates sample data (step S1). Specifically, the road surface characteristic determination device 1 generates 57 sample data from a file in which original data of 60-second components are stored, by the processing of step S1 executed by the CPU 20. As described above, in the present embodiment, the road surface characteristic determination device 1 generates, as the sample data of the characteristic amount, 4 pieces of the sprung acceleration, the unsprung acceleration, and the stroke displacement, each of the sample data including the maximum value, the minimum value, the average value, the median value, the standard deviation, the variance, the POA value in the frequency band of 0.2Hz to 3Hz, the POA value in the frequency band of 3Hz to 8Hz, the POA value in the frequency band of 8Hz to 20Hz, the POA value in the frequency band of 15Hz to 30Hz, the value obtained by dividing the unsprung acceleration by the speed of the vehicle V, and the value obtained by dividing the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration.

Next, the road surface characteristic determination device 1 associates the label of each item of the road surface characteristic with the sample data in accordance with the label input by the operator (step S2).

Further, the road surface characteristic determination device 1 performs machine learning so as to determine each item using sample data associated with the label as teacher data (step S3). Specifically, the CPU20 executes the machine learning process according to a predetermined learning algorithm for causing sample data associated with the labels of the items of the road surface feature to function as teacher data and as a recognizer. This process is repeatedly executed until the road surface feature determination device 1 performs machine learning and can determine each item of road surface features with high accuracy with respect to the input of the non-learned data.

As described above, when the road surface characteristic determination device 1 finishes machine learning by inputting sample data, it determines the readiness of the road surface characteristic with respect to the input of new data that is generated by the input data calculation unit 3 and has not been learned.

When an instruction to perform road surface feature determination on new data is input by the operator through the input device 22 after the machine learning described above, the road surface feature determination device 1 processes the original data of the vibration information and the vehicle information by the CPU20 executing a program for processing the original data, and generates new data (step S11). Specifically, the road surface characteristic determination device 1 generates 57 new data from the file in which the original data of the 60-second component is stored, by the processing of step S11 executed by the CPU 20. The CPU20 executes the processing of step S11, cuts out data of a predetermined time 3 second component from the original data of the vibration information, processes the cut-out data of the 3 second component to obtain new data, and inputs the new data to the determination unit 5, and executes an overlap processing for overlapping the data cut out when the previous new data is obtained and a part of the data cut out when the next input new data is obtained, and sequentially obtains the new data. As described above, in the present embodiment, the road surface characteristic determination device 1 generates, as the sample data of the characteristic amount, 4 pieces of the sprung acceleration, the unsprung acceleration, and the stroke displacement, each of the sample data including the maximum value, the minimum value, the average value, the median value, the standard deviation, the variance, the POA value in the frequency band of 0.2Hz to 3Hz, the POA value in the frequency band of 3Hz to 8Hz, the POA value in the frequency band of 8Hz to 20Hz, the POA value in the frequency band of 15Hz to 30Hz, the value obtained by dividing the unsprung acceleration by the speed of the vehicle V, and the value obtained by dividing the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration. That is, the feature amount of the new data is the same type of value as the feature amount of the sample data.

Next, the road surface feature determination device 1 determines each item of road surface features using the learned model with respect to the input of new data (step S12). Specifically, the CPU20 determines each value of IRI, rut depth, crack rate, flatness, MCI, and the presence or absence of a pit for each item of road surface characteristics using the learned model for the input of new data.

Then, the road surface characteristic determination device 1 displays the determination results obtained for the respective items of the road surface characteristics on the display device 6 (step S13). Further, when an instruction to print the determination result is received by the operation of the input device 22 of the operator, the road surface characteristic determination device 1 prints the determination results of the respective items onto a paper medium. Even if there is no instruction from the operator, the road surface characteristic determination device 1 may print the determination result on a paper medium by the printing device 7 and output the result.

The determination results of the road surface characteristics determined by the road surface characteristic determination device 1 are data that assist the operator in maintaining the road and assist the operator in determining whether or not the road needs to be repaired.

In this way, the method for generating a learned model according to the present embodiment generates a learned model by learning the road surface characteristics in advance by machine learning using, as teacher data, sample data including data extracted by at least one or more frequency bands from vibration information detected while the vehicle V is traveling. In the method for generating a learned model configured as described above, the vibration in the frequency band that is closely related to the deterioration of the road surface feature is given as the teacher data, and the determination unit performs machine learning on the road surface feature to generate the learned model. Further, the method of generating the learned model determines the road surface characteristics using new data having the same type as the sample data obtained from the vibration information, and therefore, when the road surface characteristics are determined, it is not necessary to perform high-cost analysis such as analysis of image data after scanning the road surface unevenness. Therefore, according to the method for generating a learned model of the present embodiment, the road surface characteristics can be determined with high accuracy at low cost. Further, according to the method for generating a learned model in the present embodiment, since the road surface characteristics can be determined at low cost, the road surface characteristics can be easily determined even for a general road or a living road. The data extracted using the frequency band included in the sample data and the new data may be a value indicating the characteristic of the frequency band in the frequency band data, such as an average value of the vibration amplitude of the frequency band. Further, the method for generating a learned model according to the present embodiment includes: a vibration information collection step of collecting vibration information by the vibration information collection device 10 while the vehicle is running; a road surface feature learning step of generating a learned model by learning road surface features in advance by machine learning using, as teacher data, sample data including data extracted from vibration information using at least one or more frequency bands. In the method of generating a learned model configured in this manner, the learned model can be generated simply by the two devices 1, 10.

The road surface feature determination device 1 according to the present embodiment includes a learned model obtained by learning road surface features in advance by machine learning using sample data including data extracted from vibration information by at least one or more frequency bands as teacher data, and a determination unit 5 for determining the road surface features based on the learned model by inputting new data having the same type as the sample data. In the road surface characteristic determination device 1 configured as described above, the vibration in the frequency band that is closely related to the deterioration of the road surface characteristic is given as the teacher data, and the determination unit 5 performs machine learning on the road surface characteristic, so that the determination of the road surface characteristic can be performed with high accuracy after the machine learning. Further, since the road surface feature determination device 1 determines the road surface feature using new data having the same type as the sample data obtained from the vibration information, it is not necessary to perform high-cost analysis such as analysis of image data after scanning the road surface unevenness when performing the determination processing of the road surface feature. Therefore, according to the road surface characteristic determination device 1 of the present embodiment, the determination of the road surface characteristic can be performed with high accuracy at low cost. Further, according to the road surface characteristic determination device 1 of the present embodiment, since the determination of the road surface characteristic can be performed at low cost, the determination of the road surface characteristic can be easily performed also for a general road or a living road. The data extracted using the frequency band included in the sample data and the new data may be a value indicating the characteristic of the frequency band in the frequency band data, such as an average value of the vibration amplitude of the frequency band.

Further, in order to determine the road surface characteristics, there are methods such as a method in which an inspector riding a running vehicle performs a sensory evaluation or a visual evaluation by comparing photographs related to IRI, but there is a possibility that it is difficult to accurately determine IRI by ignoring minute cracks or pits, and it takes much time and labor to perform a sensory test such as to all general roads or living roads in a management area. In contrast, in the method for generating a learned model and the road surface feature determination device 1 according to the present embodiment, visual work by an inspector is not required, so that the burden on the worker is small, and the road surface feature can be determined at low cost, and therefore, the method and the device can be flexibly applied to determination of the road surface feature of a general road or a living road.

As described above, the determination unit 5 performs machine learning using sample data as teacher data, performs learning of road surface features, determines road surface features for new data that is not learned, and generates a learned model used to determine road surface features for new data that is not learned by repeating machine learning by performing processing based on a pre-programmed learning algorithm. The learned model generated in this manner can be migrated as a program into a computer system without a learning function. Therefore, the determination unit 5 in the road surface feature determination device 1 may be configured to have a learned model generated by machine learning using sample data including data extracted from vibration information by at least one or more frequency bands as teacher data, and to perform determination of the road surface feature based on the learned model with respect to input of new data of the same type as the sample data, although it does not have a learning function.

In the present embodiment, the sample data includes, as the characteristic quantities, in-band vibration from 0.2Hz to 3Hz which causes the rider to feel a light wave, in-band vibration from 3Hz to 8Hz which causes the rider to feel unevenness, in-band vibration from 8Hz to 20Hz which causes the rider to feel a bump, in-band vibration from 15Hz to 30Hz which causes the rider to feel abnormal noise, in-band vibration near 16Hz which causes the rider to feel wobbling, and in-band vibration near 5Hz which causes the rider to feel wobbling, which are extracted from the vibration information. For example, if the road surface characteristics are determined by focusing only on vibrations in the frequency band of 0.2Hz to 3Hz, in which the driver feels the light and flutter, the determination unit 5 may perform machine learning using, as sample data, the characteristic amount obtained from the vibrations in the frequency band of 0.2Hz to 3 Hz. In this way, the frequency band that the operator wants to evaluate may be arbitrarily selected and the determination unit 5 may perform machine learning, and in this case, even with new data, the new data may be provided to the determination unit 5 as the feature amount obtained only from the in-band vibration corresponding to the sample data.

The determination unit 5 in the road surface characteristic determination device 1 according to the present embodiment includes a plurality of item learning units 5a, 5b, 5c, 5d, 5e, 5f that perform machine learning only on the road surface characteristics of the item corresponding to each of the plurality of items indicating the road surface characteristics, and the item learning units 5a, 5b, 5c, 5d, 5e, 5f determine the road surface characteristics only on the items corresponding to the input of new data. According to the road surface characteristic determination device 1 configured as described above, since the item learning units 5a, 5b, 5c, 5d, 5e, and 5f, which perform machine learning and determination exclusively for the items corresponding to the items representing the road surface characteristics at different scales, are provided, it is possible to perform determination with high accuracy for each item of the road surface characteristics.

In addition, since the acceleration and the stroke displacement obtained as the vibration information are time-axis waveforms, and when sample data and new data are generated, it is difficult to identify the pit grooves even though the overall road surface characteristics are easily determined when the time is not divided at all, it is necessary to divide the vibration information into pieces in time after division and obtain data for evaluating the road surface characteristics. The input data calculation unit 3 in the present embodiment obtains sample data and new data from the original data of the vibration information every 3 seconds based on the vibration information within 60 seconds, but if the above-described superimposition processing is not performed, the road surface characteristics can be grasped for each time slot, but information near the joint that divides the time is lacking. In contrast, the road surface feature determination device 1 according to the present embodiment includes the input data calculation unit 3 for extracting data of a predetermined amount of time from the original data of the vibration information, processing the extracted data of the predetermined amount of time to obtain new data, and inputting the new data to the determination unit 5, and the input data calculation unit 3 performs an overlap process for overlapping the data extracted when the previous new data is obtained and a part of the data extracted when the next input new data is obtained, and sequentially obtains the new data. In this way, the road surface feature determination device 1 can grasp the road surface feature for each time interval by performing the overlay processing, and can perform machine learning and determination by using sample data and new data that are not missing even in the vicinity of the time division line, and thus can determine the road surface feature with high accuracy. Further, since the predetermined time for dividing the time period is set to 3 seconds, the data of the low frequency component that affects the riding comfort of the vehicle V can be included in the sample data and the new data, and the road surface characteristics that affect the low frequency riding comfort can be accurately determined.

Further, in the road surface characteristic determination device 1 of the present embodiment, the sample data and the new data include a total local vibration magnitude value extracted by at least one or more frequency bands. According to the road surface characteristic determination device 1 configured in this way, since the value clearly indicating the vibration intensity in the frequency band, that is, the local vibration total magnitude value having a strong correlation with the road surface characteristic is included as the characteristic amount in the sample data and the new data, the determination accuracy of the road surface characteristic by the determination section 5 is improved.

Further, in the road surface characteristic determination device 1 in the present embodiment, since the vibration information includes the acceleration of the sprung member of the vehicle V, the road surface characteristic can be determined using the sprung acceleration directly related to the riding comfort of the vehicle V, and the road surface characteristic can be determined from the viewpoint of the riding comfort of the vehicle V.

Further, in the road surface characteristic determination device 1 according to the present embodiment, the sample data and the new data include a value obtained by dividing the maximum value of the unsprung acceleration of the vehicle V by the effective value of the unsprung acceleration. The unsprung acceleration is divided by the speed of the vehicle V to obtain a value that reduces the influence of the speed when the vehicle V travels on the road surface. Therefore, according to the road surface characteristic determination device 1 configured to perform machine learning and determination of the road surface characteristics by using the value obtained by dividing the unsprung acceleration by the speed of the vehicle V as sample data and new data, the road surface characteristics can be determined with high accuracy regardless of the traveling speed of the vehicle.

In the road surface characteristic determination device 1 according to the present embodiment, the sample data and the new data include a value obtained by dividing the maximum value of the unsprung acceleration of the vehicle V by the effective value of the unsprung acceleration. When the vehicle V passes through a damaged portion such as a pit, the unsprung portion of the vehicle V vibrates greatly, and therefore the maximum value of the unsprung acceleration increases, but the maximum value does not increase as an effective value for a certain period of time. In contrast, when the vehicle V continues to run on a rough road surface, both the maximum value and the effective value of the unsprung acceleration increase. Thus, a value obtained by dividing the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration is used as an index for indicating whether the road surface is entirely rough or partially rough. Therefore, according to the road surface characteristic determination device 1 configured to perform machine learning and determination of the road surface characteristic by using the value obtained by dividing the maximum value of the unsprung acceleration by the effective value of the unsprung acceleration as sample data and new data, it is possible to more accurately grasp whether the road surface is a road surface characteristic of global roughness or local roughness, and to determine the road surface characteristic with high accuracy.

Further, in the road surface characteristic determination device 1 of the present embodiment, the sample data and the new data include the maximum value, the average value, the median value, the minimum value, the variance, and the standard deviation of the sprung member of the vehicle V. According to the road surface characteristic determination device 1 configured in this way, the road surface characteristic can be machine-learned and determined using the index representing the vibration information of the sprung member.

As in the road surface characteristic determination device 1 according to the first embodiment shown in fig. 1, a determination unit 8 may be provided to determine whether or not new data is abnormal data. The determination unit 8 is realized by the CPU20 executing a processing program of the determination unit 8. The determination unit 8 determines whether or not the new data is abnormal based on the reference point having the average value of the sample data of the same tag as the coordinate and the new data determined to belong to the same tag. Specifically, when the number of data of feature amounts included in sample data is n, the determination unit 8 is configured to calculate an average value for each of the n feature amounts for all sample data to which the same label is added to the same item for machine learning. The determination unit 8 determines a distance from a reference point to a coordinate in the n-dimensional coordinate system of each sample data shown in the n-dimensional coordinate system, using the n average values obtained as a set as the reference point in the n-dimensional coordinate system having the n feature values as axes. For example, when the data of the feature amount included in the sample data is three data of the maximum value, the standard deviation, and one POA value of the acceleration of the sprung member, and the average value of the respective values is 10,0.15, and 6.6, the coordinates of the set (10,0.15,6.6) are set as the reference points in the three-axis coordinate system with the maximum value, the standard deviation, and the POA value as axes. The distance from the coordinate of the reference point (10,0.15,6.6) to the coordinate having the highest value, the standard deviation, and one POA value as a set of sample data may be obtained.

The determination unit 8 obtains an average value of distances from the reference point to the respective sample data and a standard deviation indicating a deviation of the distances from the reference point to the respective sample data.

In this way, the determination unit 8 obtains the reference point, the average value of the distances, and the standard deviation from the sample data in which the same label is given to each item of the road surface characteristics. That is, if an IRI tag of at most 5m/km is given to IRI at a scale of 1m/km as an example, the determination unit 8 obtains a reference point for all sample data for which the same tag is given to each IRI value, and obtains a reference point, an average value of distances, and a standard deviation for each of 5 tags when the IRI value is 1m/km, 2m/km, 3m/km, 4m/km, and 5 m/km.

When the judgment result of the judgment section 5 judges that the IRI of the new data is not less than 3m/km and less than 4m/km, the distance between the reference point of the sample data set of the same label and the coordinates of the new data in the n-dimensional coordinate system is obtained. The judging part 8 obtains the difference between the distance of the new data and the average value of the distances of the sample data groups of the same label, and judges that the new data is abnormal when the absolute value of the difference value exceeds 3 times of the standard deviation of the distances of the sample data groups of the same label; and when the absolute value of the difference value is less than 3 times of the standard deviation of the distance of the sample data group with the same label, judging that the new data is normal. When the sample data set and the new data are assumed to follow a normal distribution, the absolute value of the difference should enter a range within 3 times of the standard deviation of the distance of the sample data set of the same tag with a probability of 97%, and therefore, the new data deviated from the range may be special data. Therefore, the determination unit 8 performs the determination as described above to determine whether the new data is abnormal. As described above, as a reference when processing new data as abnormal data, a case where the absolute value of the difference is 3 times or more the standard deviation of the distance of the sample data group of the same tag is processed as data or more, but the reference may be set arbitrarily.

The result of the determination by the determination unit 8 is displayed on the display device 6, and is printed on a paper medium by the printing device 7 in accordance with the request of the operator. The determination result of the determination of the new data abnormality by the determination unit 8 is an index indicating that the machine learning by the item learning unit is insufficient or that a new label is required for the road surface feature item. Therefore, the operator of the road surface characteristic determination device 1 referring to the determination result of the determination unit 8 can notice the necessity of the determination unit 5 to perform machine learning again or the necessity of providing a new tag. The operator can supply the sample data newly collected again to the road surface characteristic determination device 1 as teacher data, and cause the road surface characteristic determination device 1 to perform machine learning. In this way, the road surface characteristic determination device 1 provided with the determination unit 8 can present the necessity of machine learning and the necessity of a new tag to the operator of the road surface characteristic determination device 1 again by the determination unit 5.

Second embodiment

Further, the input data calculation unit 3 may calculate the feature amounts as the sample data and the new data as described above, and may separately process the sample data and the new data for each of the traveling conditions of the vehicle V under which the feature amounts are obtained. Specifically, the input data calculation unit 3 refers to the vehicle information file, grasps the speed of the vehicle V at the time of obtaining the vibration information, compares the speed with a speed threshold set for the vehicle V, and processes the sample data and the new data obtained from the vibration information as data at the time of high-speed traveling when the speed at the time of obtaining the vibration information is equal to or higher than the threshold. The input data calculation unit 3 compares the speed with a speed threshold set for the vehicle V, and if the speed at which the vibration information is obtained is less than the threshold, processes the sample data and the new data obtained from the vibration information as data for low-speed traveling. The set speed threshold may be arbitrarily set, and may be, for example, 40km/h or the like. Further, the input data calculation unit 3 refers to the vehicle information file, grasps the driving condition of the wiper of the vehicle V when the vibration information is obtained, and processes the sample data and the new data obtained from the vibration information as data at the time of rainfall in the case where the wiper is being driven when the vibration information is obtained; on the contrary, in the case where the wiper is not driven when the vibration information is obtained, the sample data and the new data obtained from the vibration information are processed as data at the time of fine weather. In this way, the input data calculation unit 3 processes the sample data and the new data separately for each condition, using the speed at which the vehicle V travels and the weather as the travel condition. Therefore, the input data calculation unit 3 associates the information of the travel conditions with the sample data and the new data so as to process the sample data and the new data separately for the 4 travel conditions of the high speed/fine day time, the high speed/rainfall time, the low speed/fine day time, and the low speed/rainfall time.

On the other hand, the item learning unit 5a of the determination unit 5 includes 4 learning units 5a1, 5a2, 5a3, and 5a4 classified by the running conditions in correspondence with the 4 running conditions, as in the road surface characteristic determination device 1a according to the second embodiment shown in fig. 8. The learning unit 5a1 classified according to the driving conditions performs machine learning by using sample data at the time of high speed/fine day, and determines road surface characteristics for new data at the time of high speed/fine day. The learning unit 5a2 classified according to the driving conditions performs machine learning by using sample data at the time of high speed/rainfall, and determines road surface characteristics for new data at the time of high speed/rainfall. The learning unit 5a3 classified according to the driving conditions performs machine learning by using sample data at low speed/fine day, and determines road surface characteristics of new data at low speed/fine day. The learning unit 5a4 classified according to the driving conditions performs machine learning by using sample data at the time of low speed/rainfall, and determines road surface characteristics for new data at the time of low speed/rainfall. That is, the learning units 5a1, 5a2, 5a3, 5a4 classified by the driving conditions perform machine learning only with sample data associated with the corresponding driving conditions as teacher data for the value of IRI of the same item as the road surface feature, and determine only new data associated with the corresponding driving conditions. Further, as described above, 4 traveling conditions are set and 4 learning units classified according to the traveling conditions are provided, but when the traveling conditions are 2 high speeds and 2 low speeds or 2 rainy days and sunny days, the item learning unit 5a may include 2 learning units classified according to the traveling conditions corresponding thereto. When the number of the travel conditions is 5 or more, the item learning unit 5a may include 5 learning units classified according to the travel conditions in association with the travel conditions. In this way, the item learning unit 5a may include a corresponding number of learning units classified according to the driving conditions in accordance with the set number of driving conditions.

The learning units 5a1, 5a2, 5a3, and 5a4 classified according to the running conditions are discriminators having the same configuration as the item learning unit 5a in the first embodiment. Therefore, the learning units 5a1, 5a2, 5a3, 5a4 classified by the driving conditions execute the same processing in terms of machine learning and determination processing of new data, and only the supplied teacher data differs for each driving condition. Therefore, the learning unit 5a1 classified by the driving conditions generates a learned model dedicated to determination of IRI at high speed/fine days by machine learning, and determines IRI of new data at high speed/fine days. The learning unit 5a2 classified according to the driving conditions generates a learned model for determining IRI exclusively at the time of high-speed/rainfall by machine learning, and determines IRI of new data at the time of high-speed/rainfall. The learning unit 5a3 classified by the running conditions generates a learned model exclusively for determining IRI at low speed/fine day by machine learning, and determines IRI of new data at low speed/fine day. The learning unit 5a4 classified according to the driving conditions generates a learned model for determining IRI exclusively at the time of low speed/rainfall by machine learning, and determines IRI of new data at the time of low speed/rainfall.

The other item learning units 5b, 5c, 5d, 5e, and 5f of the determination unit 5 also include learning units 5b1, 5b2, 5b3, 5b4, 5c1, 5c2 5f3, and 5f4 classified according to the running conditions corresponding to the 4 running conditions, in the same manner as the item learning unit 5 a. The learning units 5b1, 5b2, 5b3, 5b4, 5c1, 5c2 5f3,5f4 classified according to the driving conditions are also identifiers having the same configuration as the item learning units 5b, 5c, 5d, 5e, 5f in the corresponding first embodiment. The learning units 5b1, 5b2, 5b3, 5b4, 5c1, 5c2 5f3,5f4 classified according to the driving conditions are similar to the learning units 5a1, 5a2, 5a3, 5a4 classified according to the driving conditions, and for the respective corresponding road surface feature items, only sample data associated with the corresponding driving conditions is machine-learned as teacher data, and only new data associated with the corresponding driving conditions is determined.

Further, the learning units 5a1, 5a2, 5a3, 5a4, 5b1, 5b2 … 5f3,5f4 classified according to the running conditions are used as discriminators in the road surface characteristic determination device 1a of the present embodiment, but may be used as machine learning models using a learning algorithm such as back propagation or ID 3.

As described above, the determination unit 5 performs machine learning on each item of road surface features for each of the driving conditions using the labeled sample data associated with the driving conditions generated by the input data calculation unit 3 as teacher data, and generates a learned model for each item and each driving condition. When the machine learning is completed, the determination unit 5 determines which label the new data belongs to by comparing the learned model with respect to the input of the new data which is generated by the input data calculation unit 3 and is not learned, and outputs the new data. For example, if IRI is at a high speed or in a rainfall, the learning unit 5a2 of the item learning unit 5a classified by the driving conditions determines the value of IRI for the input of new data using the learned model after machine learning, and outputs it. Similarly, the project learning units 5b, 5c, 5d, 5e, and 5f receive input of new data from the learning units classified by the driving conditions corresponding to the respective values of the rut depth, the crack rate, the flatness, and the MCI, and output the determination results to the display device 6. The road surface characteristic determination device 1a may print the determination result on a paper medium by the printing device 7 and output the result. The new data may be input to all of the learning units 5a1, 5a2, 5a3, 5a4, 5b1, 5b2 5f3,5f4 classified according to the running conditions, or the same new data may be determined by all of the learning units 5a1, 5a2, 5a3, 5a4, 5b1, 5b2 5f3,5f4 classified according to the running conditions. The road surface characteristic device 1a may output all the determination results to the display device 6, or may print the determination results on a paper medium by the printing device 7 and output the result. In this case, the operator of the road surface characteristic determining device 1a may grasp which driving condition the new data is obtained under, and then select an appropriate determination result from among 4 determination results for the same item output from the item learning units 5a, 5b, 5c, 5d, 5e, and 5f, respectively, to grasp the road surface characteristic.

In the road surface characteristic determination device 1a configured as described above, the item learning units 5a, 5b, 5c, 5d, 5e, and 5f include a plurality of learning units 5a1, 5a2, 5a3, 5a4, 5b1, 5b2 5f3, and 5f4 classified according to the running conditions, and the running condition learning units correspond to the respective running conditions of the vehicle V and perform machine learning of the road surface characteristics according to the corresponding respective running conditions. According to the road surface characteristic determination device 1a configured in this way, machine learning and determination conforming to the driving condition are performed for each item indicating the road surface characteristic, and therefore, even if the driving condition is different, it is possible to perform high-precision determination for each item of the road surface characteristic. As described above, the running condition is the speed of the vehicle V and the weather during running, but may be whether or not the vehicle is on a slope, a difference between turning and straight running, a difference between acceleration and deceleration, and a difference between constant speeds.

While the preferred embodiments of the present invention have been illustrated and described in detail, modifications, variations and changes may be made without departing from the scope of the claims.

Description of the symbols

1. 1a road surface characteristic determination device

3. Input data calculation unit

4. Database with a plurality of databases

5. Determination unit

5a, 5b, 5c, 5d, 5e, 5f item learning unit

5a1, 5a2, 5a3, 5a4, 5b1, 5b2, 5b3, 5b4, 5c1, 5c2, 5c3, 5c4, 5d1, 5d2, 5d3, 5d4, 5e1, 5e2, 5e3, 5e4, 5f1, 5f2, 5f3,5f4, learning parts classified according to running conditions

8. Determination unit

20 CPU

21. Storage device

22. Input device

Claims (12)

1. A method for generating a learned model is provided,

wherein the content of the first and second substances,

the method includes the steps of learning the road surface characteristics in advance by machine learning using, as teacher data, sample data including data extracted from vibration information detected while the vehicle is traveling using at least one frequency band, and generating a learned model.

2. The method for generating a learned model according to claim 1, comprising:

a vibration information collection step of collecting vibration information by a vibration information collection device while the vehicle is running;

and a road surface feature learning step of generating the learned model by learning the road surface features in advance by machine learning using, as teacher data, sample data including data extracted from the vibration information by using at least one or more frequency bands.

3. A road surface characteristic judging device is provided,

which is a road surface characteristic determination device that determines a road surface characteristic based on vibration information detected while a vehicle is running, wherein,

learned model generated with the method of generating learned model according to claim 1 or claim 2,

and a determination unit configured to determine the road surface feature based on the learned model by inputting new data having the same type as the sample data.

4. The road surface characteristic determination device according to claim 3,

the determination unit has a plurality of item learning units that correspond to each of a plurality of items indicating the road surface characteristics and perform learning of the road surface characteristics only for the corresponding item,

the item learning unit determines the road surface characteristics only for the item corresponding to the input of the new data.

5. The road surface characteristic determination device according to claim 4,

the item learning unit includes a plurality of learning units classified according to the driving conditions, and machine-learns the road surface characteristics for each driving condition of the vehicle and for each corresponding driving condition.

6. The road surface characteristic determination device according to claim 4,

the device is provided with a determination unit which determines whether or not the new data is abnormal, based on a reference point having an average value of sample data of the same tag as a coordinate and the new data determined to belong to the same tag.

7. The road surface characteristic determination device according to claim 3,

the information processing apparatus includes an input data calculation unit configured to execute an overlap process for overlapping data cut out when the previous new data is obtained and data cut out when the next new data is obtained, and sequentially obtain the new data and input the new data to the determination unit.

8. The road surface characteristic determination device according to claim 3,

the vibration information includes an acceleration of a sprung member of the vehicle.

9. The road surface characteristic determination device according to claim 3,

the sample data and the new data include a local vibration total amount value extracted using at least one or more frequency bands.

10. The road surface characteristic determination device according to claim 3,

the sample data and the new data include a value obtained by dividing the unsprung acceleration of the vehicle by the speed of the vehicle.

11. The road surface characteristic determination device according to claim 3,

the sample data and the new data include a value obtained by dividing a maximum value of the unsprung acceleration of the vehicle by an effective value of the unsprung acceleration.

12. The road surface characteristic determination device according to claim 3,

the sample and new data include a highest value, a mean value, a median value, a lowest value, a variance, and a standard deviation of sprung components of the vehicle.

Images