JP7359094B2 - Driving type classification device, driving type classification program and driving type classification method - Google Patents

Description

The present disclosure relates to a driving type classification device that classifies the driving type of a driver of a vehicle.

Patent Document 1 discloses a driving method that determines the driving characteristics of a driver based on vehicle information indicating the operating state of the vehicle and the operating state of the driver, and shows the tendency of the driver's driving behavior based on the results of determination of a plurality of driving characteristics. A driving support device that diagnoses the type is described.

JP 2017-191133 Publication

In the technology described in Patent Document 1, for example, sudden acceleration, sudden start, and sudden steering occur frequently, and the margin of distance between the vehicle and the vehicle in front is small, and there is no margin in the start timing when following the vehicle in front. In this case, the driving type is classified as "Quick" because the driving type is "a type of driving in which you drive while thinking that you don't care if the fuel consumption is bad and want to arrive early." In this way, in the technique described in Patent Document 1, driving types are classified based on a combination of determination results of a plurality of driving characteristics.

However, the driving type classification using the technique described in Patent Document 1 has a problem in that the driving type classification accuracy is low.
The present disclosure aims to improve the classification accuracy of driving types.

One aspect of the present disclosure is a driving type classification device (3) including a driving scene selection section (S120, S130), a learning data classification section (S140), and a driving type definition section (S170).

The driving scene selection unit selects at least one driving scene from among a plurality of preset driving scenes as a selected driving scene based on preset selection conditions indicating that a plurality of pieces of driving behavior data can be classified according to labels. Configured to select. The driving behavior data is data that represents at least one of the behavior of the vehicle and the driving operation of the driver for each vehicle, and is set to one of a plurality of labels.

The learning data classification unit is configured to classify each of the plurality of driving behavior data into a plurality of clusters based on the driving behavior data in a selected driving scene selected by the driving scene selection unit. Ru.

The driving type definition unit is configured to define a driving type for each of the plurality of clusters classified by the learning data classification unit.
The driving type classification device of the present disclosure configured in this manner can classify driving behavior data into a plurality of clusters by limiting the driving behavior data to driving scenes in which the classification accuracy is high, and define a driving type for each cluster. Therefore, the classification accuracy of driving types can be improved.

FIG. 2 is a block diagram showing the configuration of a driving type classification system. FIG. 2 is a block diagram showing the configuration of an on-vehicle device. FIG. 2 is a block diagram showing the configuration of a driving type classification device. It is a flowchart which shows driving type learning. FIG. 2 is a diagram showing an overview of symbolization. It is an explanatory diagram showing a calculation process of topic distribution. FIG. 3 is a diagram showing the structure of vinyl sentence information. FIG. 3 is a diagram showing a method of composing a topic sentence. FIG. 3 is a diagram showing input and output of driving scene classification for a plurality of driving behavior data. FIG. 3 is a diagram showing an image of driving scene classification rule definition. FIG. 3 is a diagram illustrating driving scene classification for driving behavior data. It is a flowchart showing driving scene/characteristic selection. FIG. 2 is a diagram illustrating learning and application of a random forest classifier. FIG. 3 is a diagram showing driving scene selection. It is a flowchart which shows integrated feature calculation. It is a figure explaining integrated feature calculation. It is a flowchart which shows learning data classification. FIG. 3 is a diagram showing a visualized feature distribution according to the first embodiment. FIG. 3 is a diagram illustrating manual fixation of clusters. It is a figure showing an operation type definition table. It is a flow chart which shows driving type estimation/FB of a 1st embodiment. It is a figure explaining the calculation method of similarity. It is a figure which shows the radar chart of similarity. FIG. 3 is a diagram showing driving improvement steps. It is a figure which shows the driving improvement step display. It is a flow chart which shows driving type estimation/FB of a 2nd embodiment.

[First embodiment]
A first embodiment of the present disclosure will be described below with reference to the drawings.
[1. overall structure]
The driving type classification system 1 of this embodiment includes a plurality of on-vehicle devices 2 and a driving type classification device 3, as shown in FIG.

The on-vehicle device 2 is mounted on a vehicle and has a function of performing data communication with the driving type classification device 3 via the wide area wireless communication network NW.
The driving type classification device 3 has a function of classifying driving types, which will be described later, based on data acquired from the in-vehicle device 2 via the wide area wireless communication network NW.

The in-vehicle device 2 includes a control section 11, a CAN communication section 12, a storage section 13, and a communication section 14, as shown in FIG. CAN is an abbreviation for Controller Area Network. CAN is a registered trademark.

The control unit 11 is an electronic control device mainly composed of a microcomputer including a CPU 21, a ROM 22, a RAM 23, and the like. Various functions of the microcomputer are realized by the CPU 21 executing programs stored in a non-transient physical recording medium. In this example, the ROM 22 corresponds to a non-transitional physical recording medium that stores a program. Furthermore, by executing this program, a method corresponding to the program is executed. Note that part or all of the functions executed by the CPU 21 may be configured in hardware using one or more ICs. Further, the number of microcomputers configuring the control section 11 may be one or more.

The CAN communication unit 12 is connected to a plurality of ECUs via communication lines so as to be capable of data communication, and transmits and receives data according to a CAN communication protocol. Specifically, the plurality of ECUs 111, 112, 113, 114, ... connected to the CAN communication section 12 include a motor ECU 111 that performs motor control, a brake ECU 112 that performs brake control, a steering ECU 113 that performs steering control, and a suspension control. This is the suspension ECU 114 etc. that performs this. In FIG. 2, only ECUs 111, 112, 113, and 114 are shown as ECUs connected to the CAN communication unit 12.

The storage unit 13 is a storage device for storing various data.
The communication unit 14 performs data communication with the driving type classification device 3 via the wide area wireless communication network NW.

As shown in FIG. 3, the driving type classification device 3 includes a control section 31, a communication section 32, a storage section 33, a display section 34, and an operation input section 35.
The control unit 31 is an electronic control device mainly composed of a microcomputer including a CPU 41, a ROM 42, a RAM 43, and the like. Various functions of the microcomputer are realized by the CPU 41 executing programs stored in a non-transient physical recording medium. In this example, the ROM 42 corresponds to a non-transitional physical recording medium that stores a program. Furthermore, by executing this program, a method corresponding to the program is executed. Note that part or all of the functions executed by the CPU 41 may be configured in hardware using one or more ICs. Further, the number of microcomputers configuring the control section 31 may be one or more.

The communication unit 32 performs data communication with the plurality of in-vehicle devices 2 via the wide area wireless communication network NW.
The storage unit 33 is a storage device for storing various data.

The display unit 34 includes a display device (not shown) and displays various images on the display screen of the display device.
The operation input unit 35 outputs input operation information for specifying input operations performed by the user via a keyboard and mouse (not shown).

[2. Driving type learning]
Next, the procedure of driving type learning executed by the CPU 41 of the control unit 31 will be explained. Driving type learning is executed by activating a driving type learning program stored in the storage unit 33 by an input operation by a driving type designer. Note that the driving type learning program may be installed in the driving type classification device 3 in advance, or may be installed via a recording medium or a network. Examples of the recording medium include optical disks, magnetic disks, and semiconductor memories.

When driving type learning is executed, the CPU 41 first executes learning pre-processing in S100, as shown in FIG. 4.
[2-1. Pre-learning processing]
The driving type classification device 3 performs symbolization learning, topic sentence generation, symbolization application, and driving scene classification as pre-learning processing in S100.

[2-1-1. Symbolization learning]
Next, encoding learning will be explained.
As shown in FIG. 5, the driving type classification device 3 segments data showing vehicle behavior and driving operations in time series, and sets each segmented data as a symbol segment. Further, the driving type classification device 3 calculates a topic distribution for each of the plurality of symbol segments using a topic model disclosed in, for example, Japanese Patent Application Publication No. 2014-235605.

The driving type classification device 3 repeatedly collects driving behavior data including vehicle position information, driving operation data, and vehicle behavior data from a plurality of in-vehicle devices 2 . The driving operation data is data related to the driving operation performed by the driver when driving the vehicle in which the in-vehicle device 2 is mounted. The vehicle behavior data is data related to vehicle behavior that appears as a result of a driver's driving operation of a vehicle in which the on-vehicle device 2 is mounted.

The driving type classification device 3 generates difference data by calculating the difference in the time direction of each of the driving operation data and the vehicle behavior data, and converts the multidimensional data consisting of the driving operation data, the vehicle behavior data, and the difference data into driving behavior data. Generate as. Note that, as the driving operation data, for example, the accelerator opening, brake pressure, steering angle, direction indicator status, transmission shift position, etc. can be used. Furthermore, for example, vehicle speed, yaw rate, etc. can be used as the vehicle behavior data.

The driving type classification device 3 statistically analyzes driving behavior data using a model ranging from the driver's environmental awareness to operation, and extracts the switching points of driving scenes that the driver feels. A time series of behavior data is segmented into a plurality of subsequences, each of which represents some symbol segment.

Regarding the segmentation of driving behavior data, there are non-patent documents published separately by the applicant, T. Taniguchi et al, "Semiotic Prediction of Driving Behavior using Unsupervised Double Articulation Analyzer" IEEE Intelligent Vehicles Symposium, 2012, and K. .Takenaka et al, "Contextual Scene Segmentation of Driving Behavior based on Double Articulation Analyzer" IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012, etc., etc., have detailed explanations, so we will omit the explanation here.

The driving type classification device 3 generates a driving behavior histogram for each symbol segment using driving behavior data belonging to the corresponding subsequence.
The driving type classification device 3 expresses the generated driving behavior histogram by a weighted sum of a plurality of representative driving behavior histograms (i.e., driving topics) learned in advance. This weighted vector is called a topic distribution, and each weight is called a topic ratio, and the calculated topic distribution is stored in the storage unit 33.

The driving behavior histogram is a distribution of the range of each signal of the driving behavior data belonging to the subsequence corresponding to the symbol segment of interest. For example, as shown in Fig. 6, a histogram is generated for each signal such as accelerator opening, brake pressure, steering angle, vehicle speed, and their respective difference data, and the generated histograms are connected to form the driving behavior histogram. It is.

The driving topics consist of representative histograms Topic1 to TopicK obtained by clustering K (for example, 100) generated driving behavior histograms using LDA. LDA is an abbreviation for Latent Dirichlet Allocation.

The topic ratio is the content rate of each driving topic in the driving behavior histogram, and a K-dimensional vector having K topic ratios as elements is called a topic distribution.
For more information on these techniques, see, for example, D. Blei et al, "Latent Dirichlet Allocation" Journal of Machine Learning Research, 2003, and T. Griffiths & M. Steyvers, "Finding Scientific Topics," Proceedings of the National Academy of Sciences, 2004, etc., so the explanation will be omitted here. Further, the learning method for driving topics is a known method described in, for example, Japanese Patent Application Laid-Open No. 2014-235605, and therefore a description thereof will be omitted.

[2-1-2. Topic sentence generation]
Next, topic sentence generation will be explained.
The driving type classification device 3 stores the vin sentence information in the storage unit 33.

As shown in FIG. 7, the instant information includes vehicle speed, vehicle speed difference, steering angle, steering angle difference, accelerator opening, accelerator opening difference, brake pressure, and brake pressure difference. , sentences are set according to the size of the value.

For example, when the vehicle speed is between 0 and 21 [km/h], the sentence "Drive slowly" is set. When the vehicle speed is between 72 and 102 [km/h], the sentence "Drive at high speed" is set.

When the vehicle speed difference value is between −0.5 and 0 [km/h/0.1 s], a vintage sentence of “slightly deceleration” is set. When the vehicle speed difference value is between +0.5 and [km/h/0.1 s], a vintage sentence of "rapidly acceleration" is set.

When the steering angle is between -30 and -10[deg], the sentence "a little turning the wheel to the right" is set. When the steering angle is between +15 and +45 [deg], the sentence "turning the wheel to the left" is set.

When the difference value of the steering angle is between −8 and 0 [deg/0.1s], the sentence “operating steering to the left” is set. When the difference value of the steering angle is between +8 and [deg/0.1s], the sentence "quickly operating steering to the right" is set.

When the accelerator opening is between 11% and 16%, the sentence "a little pressing on the gas pedal" is set. When the accelerator opening is between 31% and 100%, the sentence "strongly pressing down the gas pedal" is set.

When the difference value of the accelerator opening is between −1 and 0 [%/0.1 s], the sentence “slightly letting up on the gas pedal” is set. When the difference value of the accelerator opening is between +2 and [%/0.1 s], the sentence "quickly operating the gas pedal" is set.

When the brake pressure is between 0.11 and 0.5 [Mpa], the sentence "a little pressing on the brake" is set. When the brake pressure is between 1.0 and 2.0 [Mpa], the sentence "pressing down the brake" is set.

When the brake pressure difference value is between −0.12 and 0 [Mpa/0.1 s], the sentence “slightly letting up on the brake” is set. When the differential value of the brake pressure is between +0.11 and [Mpa/0.1s], the sentence "quickly operating the brake" is set.

The driving type classification device 3 then assigns the Vin sentence that takes the maximum value to each of the 100 driving topics.
For example, in "Topic 16", among the vehicle speed, the difference value of the vehicle speed, the steering angle, the difference value of the steering angle, the accelerator opening, the difference value of the accelerator opening, the brake pressure, and the difference value of the brake pressure, +7.0 of the brake pressure It is maximum in the bin of ~9.5 [Mpa]. In this case, as shown in FIG. 8, a Vincent sentence (ie, "strongly pressing down the brake") corresponding to the brake pressure of +2.0 to [Mpa] is assigned to "Topic16."

[2-1-3. Driving scene classification]
Next, driving scene classification will be explained.
As shown in FIG. 9, the driving type classification device 3 classifies the plurality of driving behavior data acquired from the in-vehicle device 2 into "start" or "not", "deceleration" or "not", "steady running" or "normal driving" for each driving scene. Classified into driving scenes such as Not.

The driving type classification device 3 generates in advance a driving scene classifier corresponding to each driving scene, and classifies the driving scene into each driving scene using each driving scene classifier. The driving type classification device 3 generates, for example, an acceleration scene classifier, a deceleration scene classifier, a steady driving scene classifier, and the like.

The procedure for generating a ``rule-based model based on topic distribution'' as a driving scene classifier using the symbolization results will be explained below.
The definition method of driving scene classification rules will be explained using a driving scene of "pause" as an example. As shown in Fig. 10, when the driving scene of "temporarily stopping" is expressed using vehicle speed, steering angle, accelerator opening, brake pressure, and the difference value between these, "temporarily stopping" means "vehicle speed is It corresponds to "low speed" and "pressing the brakes hard."

The designer then assigns a driving topic that applies to the expression of each driving scene based on the topic sentence assigned to each driving topic.
For example, the designer assigns "Topic 11" and "Topic 16" to "Vehicle speed is low" and assigns "Topic 5" and "Topic 9" to "The brake is pressed hard."

The driving type classification device 3 stores driving scene classification rule information used in driving scene classification in the storage unit 33. The driving type classification device 3 then generates each driving scene classifier based on the driving scene classification rule information. The driving type classification device 3 then determines whether the driving scene classification rules are satisfied for each symbol segment.

For example, as shown in FIG. 11, topic distributions TD1, TD2, TD3, TD4, TD5, and TD6 are calculated for each symbol segment CK1, CK2, CK3, CK4, CK5, and CK6 for driving behavior data through symbolization learning. Suppose there is.

In the topic distribution TD1, it is assumed that the proportion of "low speed" is 0.1, the proportion of "strong acceleration" is 0.2, and the proportion of "strong braking" is 0.2.
In the topic distribution TD2, it is assumed that the proportion of "low speed" is 0.08, the proportion of "strong acceleration" is 0.4, and the proportion of "strong braking" is 0.05.

In topic distribution TD3, it is assumed that the proportion of "low speed" is 0.6, the proportion of "strong acceleration" is 0.01, and the proportion of "strong braking" is 0.2.
In topic distribution TD4, it is assumed that the proportion of "low speed" is 0.4, the proportion of "strong acceleration" is 0.02, and the proportion of "strong braking" is 0.3.

In the topic distribution TD5, it is assumed that the proportion of "low speed" is 0.05, the proportion of "strong acceleration" is 0.6, and the proportion of "strong braking" is 0.05.
In topic distribution TD6, it is assumed that the proportion of "low speed" is 0.3, the proportion of "strong acceleration" is 0.01, and the proportion of "strong braking" is 0.4.

Assume that the "pause" scene classification rule is "low speed topic ratio ≧0.3 and strong brake topic ratio ≧0.2".
In this case, the driving type classification device 3 determines that the symbol segments CK3, CK4, and CK6 satisfy the driving scene classification rules, and sets the symbol segment CK3, CK4, and CK6 to "pause"=True, and the symbol segment Set "temporary stop"=False for CK1, CK2, and CK5.

Further, the driving scene classification method is not limited to this. For example, a driving scene classification model may be generated by supervised learning using a topic distribution and a driving scene teacher signal, and driving scenes may be classified.
[2-2. Label settings]
When the process of S100 is completed, the CPU 41 performs label setting in S110, as shown in FIG. Specifically, the CPU 41 sets a label for each of the plurality of driving behavior data. Specifically, in the initial state, the CPU 41 sets a label for each of the plurality of driving behavior data based on label information input from the operation input unit 35. After the driving type is defined in subsequent processing, the CPU 41 uses the defined driving type as a label. Note that some labels may be selected and used depending on the purpose. For example, two labels, "Professional" and "Amateur", are used to evaluate driving operation skills, and two labels, "Good Electricity Cost" and "Bad Electricity Cost" are used to evaluate eco-driving skills for electric cars. You may use it. This makes it possible to classify driving types with a certain degree of directionality in accordance with the designer's purpose.

[2-3. Driving scene/feature selection]
When the process of S110 is completed, the CPU 41 executes driving scene/characteristic selection in S120. The procedure for selecting driving scenes and features will be explained below.

When driving scene/feature selection is executed, as shown in FIG. 12, the CPU 41 first selects driving scenes that have not yet been selected as processing targets from S220 to S250 from among the plurality of driving scenes preset in S210. Select a scene as the target driving scene. Driving scenes include, for example, "starting", "accelerating", "steady running", "decelerating", "pause", "creeping", "right turn", "right curve", "left turn", "left curve", " Examples include "Go straight ahead."

In S220, the CPU 41 extracts the topic distribution of symbol segments corresponding to the target driving scene for each of the plurality of driving behavior data for which the labels were set in S110.
For example, when the target driving scene is "acceleration", the CPU 41 extracts the topic distribution of symbol segments classified as "acceleration" for each of the plurality of driving behavior data.

In S230, the CPU 41 performs learning using the label set in S110. Specifically, the CPU 41 constructs a driving scene-based label classifier that classifies the labels set in S110 in the target driving scene. In this embodiment, the CPU 41 constructs a classifier using Random Forest (hereinafter referred to as RF), which is a well-known machine learning algorithm.

As shown in FIG. 13, the CPU 41 extracts, for example, topic distributions TD1, TD4, and TD5 in the driving scene of "acceleration" from the driving behavior data DD1 to which the label of "professional" is set. For example, the CPU 41 extracts topic distributions TD2, TD3, and TD6 in the driving scene of "acceleration" from the driving behavior data DD2 to which the label of "amateur" is set.

The CPU 41 uses the extracted topic distribution to train a classifier that classifies "professionals" and "amateurs" as shown in feature space FS1. Classification boundaries BD1, BD2, and BD3 in the feature space FS1 are boundaries for classifying "professional" and "amateur".

When the process of S230 ends, as shown in FIG. 12, in S240, the CPU 41 selects a topic with a high degree of contribution in the RF learning of S230, and sets it as the selected topic. In addition, in the target driving scene, the frequency distribution of each feature value for each topic ratio is compared between labels (for example, between "professional" and "amateur"), and the top N topics with a large difference are selected based on the degree of contribution. It may be selected as a high-rank feature quantity.

For example, when the target driving scene is "deceleration", the CPU 41 selects topic numbers "16", "31", "36", "47", "48", "57", "74", "87". ”, “89”, and “95” are selected as selected topics. Further, when the target driving scene is a "left curve", the CPU 41 selects topic numbers "2", "7", "28", "31", "34", "37", "70", "75". ”, “81”, and “95” are selected as selected topics.

The CPU 41 calculates the AUC of the ROC curve in S250. AUC is an abbreviation for Area Under the Curve, and hereinafter the AUC of the ROC curve will be simply described as AUC. ROC is an abbreviation for Receiver Operating Characteristic. Specifically, the CPU 41 calculates the AUC of the classifier constructed in S230 using evaluation data described below.

As shown in FIG. 13, the CPU 41 uses, for example, the driving behavior data DD3 that is not used for learning the label classifier as evaluation data, and extracts topic distributions TD11, TD12, and TD13 in the "acceleration" scene from this evaluation data. Then, the CPU 41 causes the label classifier to classify the topic into "professional" or "amateur" based on the extracted topic distributions TD11, TD12, and TD13. The CPU 41 calculates the AUC in the "acceleration" scene based on the classification result of the label classifier.

When the process in S250 is completed, the CPU 41 determines in S260 whether all driving scenes have been selected as target driving scenes, as shown in FIG. 12. Here, if all driving scenes are not selected as target driving scenes, the CPU 41 moves to S210. On the other hand, if all the driving scenes are selected as target driving scenes, the CPU 41 determines in S270 that the AUC of each driving scene calculated in S250 exceeds the preset scene selection judgment value TH1. The driving scene in which the vehicle is running is selected as the selected driving scene. Further, in S270, the CPU 41 stores the pair of the selected driving scene selected and the selected topic selected in S240 in the RAM 43, and ends the driving scene/feature selection.

For example, as shown in FIG. 14, since the AUC of the "deceleration" scene and the AUC of the "left curve" scene exceed the scene selection judgment value TH1, "deceleration" and "left curve" are selected as the selected driving scenes. Ru.

[2-4. Integrated feature calculation]
When the driving scene/feature selection in S120 is completed, the CPU 41 executes integrated feature calculation in S130, as shown in FIG. Here, the procedure for calculating integrated features will be explained.

When the integrated feature calculation is executed, as shown in FIG. 15, first in S310, the CPU 41 determines which of the one or more selected driving scenes selected in S270 has not yet been selected as the processing target in the following S320 to S350. A selected driving scene that has not been selected is selected as a target driving scene.

In S320, the CPU 41 extracts the topic distribution of symbol segments corresponding to the target driving scene for each of the plurality of driving behavior data for which the labels were set in S110.
For example, when the target driving scene is "deceleration", the CPU 41 extracts the topic distribution of symbol segments classified as "deceleration" from the driver's driving behavior data for each of the plurality of drivers. For example, as shown in FIG. 16, the CPU 41 extracts symbol segments CK1, CK4, and CK6 classified as "deceleration" from the driving behavior data DD4, and extracts symbol segments CK1, CK4, and CK6 classified as "deceleration" from the driving behavior data DD5. Extract segments CK12 and CK15. Further, when the target driving scene is a "left curve", the CPU 41 extracts symbol segments CK2, CK3, and CK5 classified as "left curve" from the driving behavior data DD4, and extracts symbol segments CK2, CK3, and CK5 classified as "left curve" from the driving behavior data DD5. Symbol segments CK11, CK13, and CK14 classified as "left curve" are extracted.

When the process of S320 is completed, the CPU 41 extracts the selected topic corresponding to the target driving scene in S330, as shown in FIG. 15.
In S340, the CPU 41 averages the topic ratios corresponding to the selected topics extracted in S330.

For example, as shown in FIG. 16, when the target driving scene is "deceleration", the CPU 41 selects topic number "16" from the topic distribution included in N symbol segments classified as "deceleration". , "31", "36", "47", "48", "57", "74", "87", "89", "95", and extract the topic proportions for each topic number. Calculate the average of the percentages.

In addition, when the target driving scene is a "left curve", the CPU 41 selects topic numbers "2", "7", Extract the topic percentages corresponding to "28", "31", "34", "37", "70", "75", "81", and "95" and calculate the average topic percentage for each topic number do.

When the process of S340 is completed, the CPU 41 calculates the average of similar topics in S350, as shown in FIG. Specifically, the CPU 41 first groups similar selected topics together at a predetermined threshold based on the degree of similarity in topic distribution between the selected topics. An inter-distribution distance measure called Earth Mover's Distance (EMD) is used for the similarity of topic distributions. Then, the CPU 41 calculates the average topic ratio of similar selected topics (hereinafter referred to as integrated topic ratio).

For example, as shown in FIG. 16, when the target driving scene is "deceleration", the CPU 41 sets the topic numbers "16", "47", "36", "57", and the topic numbers "31", " 74'', ``87'', ``89'', ``95'' and topic number ``48'' are grouped together as similar selected topics, respectively, and integrated topic ratios IT1, IT2, and IT3 are calculated.

Further, when the target driving scene is a "left curve", the CPU 41 inputs topic numbers "2", "81", topic numbers "34", topic numbers "7", topic numbers "31", " 37'', ``95'', topic number ``75'', topic number ``28'', and topic number ``70'' are grouped together as similar selected topics, respectively, and the integrated topic ratios IT11, IT12, IT13, IT14, Calculate IT15, IT16, and IT17.

When the process of S350 is completed, as shown in FIG. 15, the CPU 41 determines in S360 whether all selected driving scenes have been selected as target driving scenes. Here, if all selected driving scenes are not selected as target driving scenes, the CPU 41 moves to S310. On the other hand, when all selected driving scenes are selected as target driving scenes, the CPU 41 calculates an integrated feature by combining all integrated topic ratios calculated in S350 in S370, and ends the integrated feature calculation. Thereby, integrated features are calculated for each driver.

For example, as shown in FIG. 16, the CPU 41 calculates the integrated feature IF1 by combining the integrated topic ratios IT1 to IT3 and the integrated topic ratios IT11 to IT17.

[2-5. Learning data classification]
When the integrated feature calculation in S130 is completed, the CPU 41 executes learning data classification in S140, as shown in FIG. Here, the procedure for classifying learning data will be explained.

When the learning data classification is executed, as shown in FIG. 17, first in S410, the CPU 41 clusters the integrated features of each driver obtained in S130 by machine learning, and classifies them into a plurality of clusters. In S410, the CPU 41 uses, for example, the X-means method to perform clustering including optimization of the number of clusters.

In S420, the CPU 41 performs principal component analysis (hereinafter referred to as PCA) on the integrated feature data of each driver.
When the process of S420 is completed, the CPU 41 selects three integrated feature axes in order of highest first principal component score in S430, and as shown in FIG. A visualized feature distribution in which the integrated features of each driver are arranged in a coordinate system is generated, and the generated visualized feature distribution is displayed on the display screen of the display unit 34.

When the process of S430 is completed, the CPU 41 performs cluster adjustment in S440, as shown in FIG. Specifically, the CPU 41 performs manual separation, manual combination, and manual fixation of clusters based on input operation information input from the operation input unit 35. For example, as shown in FIG. 19, when there is an aggregated cluster CL2 inside the cluster CL1, when input operation information instructing manual fixation of the cluster CL2 is input from the operation input unit 35, the CPU 41 , manually fix cluster CL2. Further, when input operation information instructing the end of cluster adjustment is input from the operation input unit 35, the CPU 41 ends the process of S440. Note that if the label set in S110 is similar to the driving type assumed by the designer (hereinafter referred to as the assumed type), the designer can manually separate the clusters, referring to the set label. Manual joining and manual fixing may also be performed.

When the process of S440 ends, the CPU 41 ends the learning data classification, as shown in FIG. 17.
[2-6. Convergence judgment]
When the learning data classification is completed, the CPU 41 executes convergence determination in S160, as shown in FIG. Specifically, the CPU 41 first calculates the intra-class/inter-class variance ratio (hereinafter referred to as degree of separation) classified in the learning data classification in S140. Then, the CPU 41 determines that "convergence has been achieved" when the calculated degree of separation is greater than or equal to a preset convergence determination threshold. The convergence determination threshold is set to have a negative correlation with the number of clusters.

If it is not determined that the process has converged, the CPU 41 moves to S110. On the other hand, if it is determined that the process has converged, the CPU 41 moves to S170.
[2-7. Operation type definition]
When the process moves to S170, the CPU 41 executes operation type definition that defines the operation type of the cluster. Specifically, the CPU 41 defines a cluster driving type for each cluster based on the driving type definition table stored in the storage unit 33 and the positional relationship between the integrated feature axis and the cluster.

As shown in FIG. 20, the driving type definition table is a table in which a correspondence relationship between a feature number for identifying the integrated feature, a driving scene, a driving behavior, and a driving type element is set for each integrated feature. . A driving type element is defined for each integrated feature based on the meaning of the integrated feature (ie, the frequency of driving behavior during a driving scene).

Then, the CPU 41 defines the driving type of each cluster, taking into account the barycentric coordinates of the cluster and the meaning of each integrated feature axis (that is, driving type element). For example, if a certain cluster is close to the integrated feature axis of feature number 1 and the integrated feature axis of feature number 2, both the driving type element "braking impatiently" and the driving type element "slowly going straight" are considered. Then, the driving type "timid" is defined.

[3. Driving type estimation/FB]
Next, the procedure of driving type estimation/FB executed by the CPU 41 of the control unit 31 will be explained. Driving type estimation/FB is executed by activating a driving type estimation program stored in the storage unit 33 by an input operation by the user. Note that the driving type estimation program may be installed in the driving type classification device 3 in advance, or may be installed via a recording medium or a network.

When the driving type estimation/FB is executed, the CPU 41 first performs estimation preprocessing in S510, as shown in FIG. 21. Specifically, the CPU 41 performs symbolization and driving scene classification on the driving behavior data stored in the storage unit 33 in the same manner as the above-described pre-learning process.

In S520, the CPU 41 calculates integrated features for each of the plurality of drivers based on the driver's driving behavior data in the same manner as the integrated feature calculation in S130.
In S530, the CPU 41 calculates the degree of similarity with each cluster for each of the plurality of drivers using the learned multi-class classifier.

For example, as shown in FIG. 22, the CPU 41 first calculates the distance d1 between the centroids CG11, CG12, CG13, CG14, CG15 of the clusters CL11, CL12, CL13, CL14, CL15 in the integrated feature space and the integrated feature IF10 of a certain driver. , d2, d3, d4, and d5. Note that integrated feature 1 indicates the frequency of driving behavior a during driving scene A. Integrated feature 2 indicates the frequency of driving behavior b during driving scene B. Integrated feature 3 indicates the frequency of driving behavior c during driving scene C.

Then, the CPU 41 sets the reciprocals of the calculated distances d1, d2, d3, d4, and d5 as similarities with the clusters CL11, CL12, CL13, CL14, and CL15, respectively.
When the process of S530 is completed, as shown in FIG. 21, in S540, the CPU 41 sets improvement steps and recommended driving behavior to bring the driver closer to the target driving type set in advance for each driver.

For example, as shown in FIG. 24, the integrated feature IF10 of a certain driver is closest to the cluster CL11 defined as the driving type "Hot", and second closest to the cluster CL12 defined as the driving type "Cool", It is assumed that the cluster CL13 is the third closest to the cluster CL13, which is defined as the driving type "Eco". It is assumed that the target driving type of this driver is "Eco". In this case, the CPU 41 sets the transition from cluster CL11 to cluster CL12 as STEP1, and sets the transition from cluster CL12 to cluster CL13 as STEP2.

Also, since the transition from cluster CL11 to cluster CL12 mainly follows integrated feature 2, the CPU 41 sets "Try to adopt driving behavior b in driving scene B" as the recommended driving behavior in STEP 1. do. In addition, since the transition from cluster CL12 to cluster CL13 mainly follows integrated feature 3, the CPU 41 sets "Try to adopt driving behavior c in driving scene C" as the recommended driving behavior in STEP 2. do.

When the process of S540 is completed, as shown in FIG. 21, the CPU 41 transmits feedback information for each of the plurality of drivers to the in-vehicle device 2 of each driver's vehicle in S550, and ends the driving type estimation/FB. . The feedback information includes information indicating the similarity calculated in S530, information indicating the improvement steps set in S540, and information indicating the recommended driving behavior set in S540.

When the in-vehicle device 2 receives the feedback information, it transfers the feedback information to the navigation device mounted on the vehicle. The navigation device that has received the feedback information displays the feedback information on the display screen of the navigation device. This allows the driver to check the feedback information.

The information indicating the degree of similarity is, for example, a radar chart indicating the reciprocals of the distances d1, d2, d3, d4, and d5, as shown in FIG.
The information indicating the improvement step is, for example, as shown in FIG. 25, a radar chart indicating the improvement step from "Hot" to "Eco" and the Eco score. The information indicating recommended driving behavior is "Step 1: Try to take driving behavior b during driving scene B" and "Step 2: Try to take driving behavior c during driving scene C."

In the radar chart shown in FIG. 25, a pentagon HG1 indicates the current degree of similarity, a pentagon HG2 indicates the degree of similarity for transitioning to "Cool," and a pentagon HG3 indicates the degree of similarity for transitioning to "Eco." It is configured. Further, this radar chart includes an arrow SP11 indicating STEP 1 for transitioning from "Hot" to "Cool", and an arrow SP12 indicating STEP 2 for transitioning from "Cool" to "Eco".

[4. effect]
The driving type classification device 3 configured in this manner selects at least one driving scene from among a plurality of preset driving scenes based on preset selection conditions indicating that a plurality of driving behavior data can be classified according to labels. A driving scene is selected as a selected driving scene. The driving behavior data represents at least one of the vehicle behavior and the driver's driving operation for each vehicle, and is data to which one of a plurality of labels is set.

The driving type classification device 3 classifies each of the plurality of driving behavior data into a plurality of clusters based on the driving behavior data in the selected driving scene.

The driving type classification device 3 defines a driving type for each of the plurality of classified clusters.
In this way, the driving type classification device 3 can classify driving behavior data into a plurality of clusters by limiting the driving behavior data to driving scenes where the classification accuracy is high, and can define a driving type for each cluster. classification accuracy can be improved.

Further, the driving type classification device 3 selects at least one topic selected for each selected driving scene as a selected topic. The driving type classification device 3 then classifies the plurality of driving behavior data into a plurality of clusters in the selected feature space constituted by the selected selected topics. In this way, the driving type classification device 3 limits driving behavior data to driving scenes for which the classification accuracy is high, and furthermore, within the limited driving scenes, driving behavior data is limited to topics for which the classification accuracy of driving behavior data is high. The classification accuracy of driving types can be further improved.

Further, the driving type classification device 3 extracts a topic distribution in a driving scene from a plurality of driving behavior data for each of a plurality of driving scenes, and uses a driving type classification device for classifying a plurality of driving behavior data according to labels based on the topic distribution. Train a scene-specific label classifier. Further, the driving type classification device 3 calculates AUC representing the performance of the label classification model for each of the plurality of driving scenes, based on the label classification model obtained by learning the label classifier for each driving scene. Then, the driving type classification device 3 selects a driving scene in which the performance of the label classification model is high based on the AUC as a selected driving scene. That is, the above selection conditions are set based on the AUC to indicate a driving scene in which the performance of the label classification model is high.

Further, the driving type classification device 3 selects a plurality of selected topics from among the selected plurality of selected topics, and generates a visualization feature indicating the distribution of driving behavior data in a visualized feature space constituted by the selected plurality of selected topics. Generate and display distributions. Then, the driving type classification device 3 adjusts the plurality of clusters in the visualized feature distribution based on input operation information indicating an instruction from the user using the driving type classification device 3. As a result, the driving type classification device 3 allows the designer to visually check the visualized feature distribution and manually adjust the clusters while taking the driving characteristics into consideration, allowing the designer to appropriately classify the clusters according to the purpose. can.

The driving type classification device 3 also performs PCA on a plurality of driving behavior data in the integrated feature space, selects selected topics with high first principal component scores, and generates a visualized feature distribution. Thereby, the driving type classification device 3 can quantify the selection criteria for selecting the selected topic in the visualized feature space, and can select the selected topic simply and rationally.

Further, the driving type classification device 3 selects a driving scene based on the AUC of the label classifier. Thereby, the driving type classification device 3 can appropriately select a driving scene with high classification accuracy of driving behavior data along the label.

The driving type classification device 3 also compares the frequency distribution of each topic ratio between labels, and selects a feature quantity with a large difference. Thereby, the driving type classification device 3 can appropriately select a feature quantity with high classification accuracy of driving behavior data along the label.

Further, the driving type classification device 3 repeatedly selects a selected driving scene, selects a selected topic, and classifies learning data using the selected selected driving scene and selected topic. Thereby, the driving type classification device 3 can converge the classification results by performing learning data classification multiple times, and can classify clusters with high accuracy.

Further, when the degree of separation is equal to or greater than the convergence determination threshold, the driving type classification device 3 determines that the classification result has converged, and ends the iterative process. Thereby, the driving type classification device 3 can automate the selection of selected driving scenes and selected topics, and the repeated processing of cluster classification, and can reduce the workload of the designer.

Furthermore, multiple pieces of driving behavior data are automatically segmented by symbolization. Thereby, the driving type classification device 3 can eliminate the need for manual segmentation work.
Further, the driving type classification device 3 selects a plurality of selected topics for each selected driving scene. Then, the driving type classification device 3 extracts the selected topic ratio from the plurality of driving behavior data for each of the plurality of selected driving scenes, and calculates the average of the selected topic ratios as the average topic ratio. Further, the driving type classification device 3 integrates a plurality of selected topics that are similar to each other, and calculates the average of the average topic ratios of the integrated selected topics as an integrated topic ratio. The driving type classification device 3 then combines the plurality of integrated topic ratios calculated for each of the plurality of selected driving scenes to form an integrated feature. Thereby, the driving type classification device 3 can perform highly accurate driving type classification by limiting driving scenes. Further, the driving type classification device 3 can facilitate the semantic interpretation of the selected topics by reducing the dimensionality by integrating selected topics that are similar to each other.

Further, the driving type classification device 3 defines the driving type of the cluster for each cluster based on the position of the cluster in the integrated feature space and a driving type definition table that sets the meaning of the integrated feature axis as a driving type element. . Thereby, the driving type classification device 3 can easily define driving types.

Furthermore, the driving type classification device 3 uses the diagnosis result of diagnosing the driving type of the driver and assigns the driver to the target driving type based on the positional relationship between the classified clusters and the integrated features of the driver and the target driving type of the driver. Notify the driver of improvement steps to improve the situation. Thereby, the driving type classification device 3 can make step-by-step proposals to bring the driver closer to the target driving type, making it easier for the driver to make improvements.

Further, the driving type classification device 3 notifies the driver of recommended driving behavior to bring the driver closer to the target driving type, limited to the selected driving scene. As a result, the driving type classification device 3 can limit and propose driving scenes that bring the driver closer to the target driving type, making it easier for the driver to make improvements.

In the embodiment described above, S120 and S130 correspond to processing as a driving scene selection section, S140 corresponds to processing as a learning data classification section, and S170 corresponds to processing as a driving type definition section.

Further, S240 and S130 correspond to processing as a feature selection unit, a topic corresponds to a feature amount, and a selected topic corresponds to a selected feature amount.
Further, S230 corresponds to processing as a label classifier learning unit, S250 corresponds to processing as an evaluation index calculation unit, topic distribution corresponds to driving scene data, and the label classifier by driving scene corresponds to the label classifier. AUC corresponds to an evaluation index.

Further, S420 and S430 correspond to processing by the learning data display section, S440 corresponds to processing by the cluster adjustment section, and input operation information corresponds to instruction information.
Further, S160 corresponds to processing as a repetitive execution determination unit, S340 corresponds to processing as an average topic ratio calculation unit, S350 corresponds to an integrated topic ratio calculation unit, and S130 and S370 correspond to processing as an integrated feature calculation unit. S510 to S550 correspond to the processing of the driver FB unit.

[Second embodiment]
A second embodiment of the present disclosure will be described below with reference to the drawings. Note that in the second embodiment, different parts from the first embodiment will be explained. Common configurations are given the same reference numerals.

The driving type classification system 1 of the second embodiment differs from the first embodiment in that driving type estimation/FB has been changed.
The driving type estimation/FB of the second embodiment differs from the first embodiment in that, as shown in FIG. 26, the process of S570 is executed instead of S550.

That is, when the process of S540 is completed, the CPU 41 transmits feedback information for each of the plurality of drivers to the motor ECU 111 of each driver's vehicle in S570, and ends the driving type estimation/FB.

When motor ECU 111 receives the feedback information, it intervenes in motor control in stages according to improvement steps. For example, the motor ECU 111 executes motor control in accordance with the first step until the improvement step is canceled or the vehicle runs ten times, and thereafter executes motor control in accordance with the second step until the improvement step is canceled. Note that when intervening in motor control in stages according to improvement steps, the motor ECU 111 may notify the driver of the control details in advance. This allows the driver to recognize that control is being performed to bring the vehicle closer to the target driving type.

The driving type classification device 3 configured in this manner analyzes the driver's driving type and the diagnosis result of the driver's driving type based on the positional relationship between a plurality of classified clusters and the driver's integrated characteristics and the driver's target driving type. The motor ECU 111 is notified of improvement steps to bring the motor closer to the target driving type. As a result, the driving type classification device 3 can cause the motor ECU 111 to perform motor control (hereinafter referred to as improvement control) in stages in order to bring the driver closer to the target driving type. The resulting discomfort can be suppressed.

Further, the driving type classification device 3 notifies the motor ECU 111 of recommended driving behavior for bringing the driver closer to the target driving type, limited to the selected driving scene. Thereby, the driving type classification device 3 can perform the improvement control only in the selected driving scene, and can further suppress the sense of discomfort caused to the driver by performing the improvement control.

In the embodiment described above, S510 to S540 and S570 correspond to the processing of the vehicle control FB section, and the motor ECU 111 corresponds to the electronic control unit that controls the running of the vehicle.

Although one embodiment of the present disclosure has been described above, the present disclosure is not limited to the above embodiment, and can be implemented with various modifications.
[Modification 1]
In the above embodiment, a mode is shown in which three integrated feature axes are selected in order of the highest first principal component score in S430. However, three integrated feature axes having high first principal component scores and low second principal component scores may be selected.

[Modification 2]
In the embodiment described above, in S430, three integrated feature axes are selected in descending order of the first principal component score to generate the visualized feature distribution. However, we performed PCA and set the first principal component with the highest degree of contribution, the second principal component with the second highest degree of contribution, and the third principal component with the third highest degree of contribution into three axes. The distribution of integrated features of the driver in the dimensionally compressed space of the three-dimensional orthogonal coordinate system and the positions of the plurality of integrated feature axes in the dimensionally compressed space may be displayed. When input operation information instructing the three integrated feature axes is input via the operation input unit 35, a visualized feature distribution composed of the three integrated feature axes indicated by the input operation information is generated and displayed. You can also do this. Thereby, the designer can select the integrated feature axis of the visualized feature distribution by considering the positional relationship between the integrated feature distribution and the integrated feature axis.

[Modification 3]
In the embodiment described above, in S430, three integrated feature axes are selected in descending order of the first principal component score to generate the visualized feature distribution. However, the integrated feature axis may be arbitrarily selected according to the purpose of the designer.

For example, if the set label is similar to the assumed type, an integrated feature axis along the assumed type may be selected preferentially. Furthermore, if there is an assumed type, an integrated feature axis along the assumed type may be selected preferentially. For example, if there is an assumed type of "impatient", (1) "strong braking during deceleration scenes", (2) "strong acceleration during acceleration scenes", and (3) "high speed driving during steady driving scenes". It is preferable to select an integrated feature axis corresponding to the .

[Modification 4]
In the above embodiment, the label is set in S110. However, in the initial state, all symbol segments and all topics may be used for clustering without setting labels, and the cluster labels obtained by this clustering may be used. Thereby, the driving type classification device 3 can easily select driving scenes and feature amounts without preparing labels.

[Modification 5]
In the above embodiment, the convergence determination is automatically performed in S150. However, the designer may visually check the clusters of the visualized feature distribution displayed on the display screen of the display unit 34 and subjectively determine convergence depending on the purpose.

[Modification 6]
In the above embodiment, the feedback information is displayed on the display screen of the navigation device. However, the feedback information may be transmitted to the driver's smartphone and displayed on the display screen of the smartphone.

[Modification 7]
In the above embodiment, the driving type classification device 3 performs the driving type estimation/FB, but the in-vehicle device 2 may perform the driving type estimation/FB. In this case, it is necessary to store in the vehicle-mounted device 2 the exercise type model generated by the driving type classification device 3 through the driving type learning process. When the driving type classification device 3 updates the exercise type model through the driving type learning process, it is preferable that the exercise type model stored in the vehicle-mounted device 2 is also updated by OTA or the like. OTA is an abbreviation for Over The Air.

[Modification 8]
In the embodiment described above, PCA is performed on the integrated feature data of each driver, a visualized feature distribution is generated, and clusters are manually separated, manually combined, and manually fixed. However, the integrated features of each driver arranged in an integrated feature space made up of a plurality of integrated feature axes may be classified into a plurality of clusters using unsupervised learning. This eliminates the need for manual work by the designer for cluster classification, making it possible to simplify cluster classification.

[Modification 9]
In the embodiment described above, PCA is performed on the integrated feature data of each driver, a visualized feature distribution is generated, and clusters are manually separated, manually combined, and manually fixed. However, instead of always manually adjusting the clusters, the number of times the clusters are manually adjusted may be limited. Further, the manual adjustment timing of clusters may be set arbitrarily (for example, only at the beginning, in the middle, only at the end, etc.).

The control unit 31 and its method described in the present disclosure are implemented by a dedicated computer provided by configuring a processor and memory programmed to perform one or more functions embodied by a computer program. May be realized. Alternatively, the controller 31 and its techniques described in this disclosure may be implemented by a dedicated computer provided by a processor configured with one or more dedicated hardware logic circuits. Alternatively, the control unit 31 and its method described in the present disclosure are a combination of a processor and memory programmed to execute one or more functions and a processor configured by one or more hardware logic circuits. It may be realized by one or more dedicated computers configured with. The computer program may also be stored as instructions executed by a computer on a computer-readable non-transitory tangible storage medium. The method of realizing the functions of each part included in the control unit 31 does not necessarily need to include software, and all the functions may be realized using one or more pieces of hardware.

A plurality of functions of one component in the above embodiment may be realized by a plurality of components, and a function of one component may be realized by a plurality of components. Further, a plurality of functions possessed by a plurality of constituent elements may be realized by one constituent element, or one function realized by a plurality of constituent elements may be realized by one constituent element. Further, a part of the configuration of the above embodiment may be omitted. Further, at least a part of the configuration of the above embodiment may be added to or replaced with the configuration of other embodiments.

In addition to the above-mentioned driving type classification device 3, there is also a system including the driving type classification device 3 as a component, a program for making a computer function as the driving type classification device 3, a non-transitional system such as a semiconductor memory in which this program is recorded. The present disclosure can also be implemented in various forms, such as a tangible recording medium and a driving type classification method.

3... Operation type classification device, 31... Control unit

Claims (28)

A plurality of driving behavior data in which one label among a plurality of labels representing at least one of a driving operation of a driver driving the vehicle and a behavior of the vehicle appearing as a result of the driving operation is set for each vehicle. a driving scene configured to select at least one driving scene as a selected driving scene from among a plurality of preset driving scenes based on a preset selection condition indicating that the driving scene can be classified according to the label; A selection unit (S120, S130),
For each of the plurality of driving behavior data corresponding to the plurality of drivers , based on the driving behavior data in the selected driving scene selected by the driving scene selection section, the plurality of driving behavior data are clustered into a plurality of driving behavior data. a learning data classification unit (S140) configured to classify into clusters;
a driving type definition unit (S170) configured to define a driving type for each of the plurality of clusters classified by the learning data classification unit;
A driving type classification device (3) comprising: A plurality of driving behavior data in which one of the plurality of labels representing at least one of the driving operation of the driver driving the vehicle and the behavior of the vehicle appearing as a result of the driving operation is set for each vehicle. Driving scene selection configured to select at least one driving scene as a selected driving scene from among a plurality of preset driving scenes based on a preset selection condition indicating that the driving scene can be classified according to the label. part (S120, S130),
For each of the plurality of driving behavior data corresponding to the plurality of drivers, based on the driving behavior data in the selected driving scene selected by the driving scene selection section, the plurality of driving behavior data are clustered into a plurality of driving behavior data. a learning data classification unit (S140) configured to classify into clusters;
a driving type definition unit (S170) configured to define a driving type for each of the plurality of clusters classified by the learning data classification unit;
Equipped with
The driving scene selection unit includes a feature selection unit (S240, S130) configured to select at least one feature amount selected corresponding to the at least one selected driving scene as a selection feature amount. Prepare,
The learning data classification unit is a driving type classification device that classifies a plurality of the driving behavior data into a plurality of clusters in a selected feature space constituted by the selected feature quantities selected by the feature selection unit. The driving type classification device according to claim 2,
The driving scene selection section includes:
For each of the plurality of driving scenes, the driving behavior data in the driving scene is extracted from the plurality of driving behavior data as driving scene data, and based on the driving scene data, the plurality of driving behavior data are extracted according to the label. a label classifier learning unit (S230) configured to learn a label classifier for classification;
The configuration is configured to calculate an evaluation index for evaluating the performance of the label classification model by the label classifier learning unit for each of the plurality of driving scenes based on the label classification model obtained by the label classifier learning unit. an evaluation index calculation unit (S250),
The driving type classification device is configured such that the selection condition is set to indicate that the driving scene has a high performance of the label classification model based on the evaluation index calculated by the evaluation index calculation unit. The driving type classification device according to claim 2 or 3,
The learning data classification unit is
Selecting a plurality of selected feature quantities from among the plurality of selected feature quantities selected by the feature selection unit, and showing a distribution of the driving behavior data in a feature space constituted by the selected plurality of selected feature quantities. a learning data display unit (S420, S430) configured to generate and display a visualized feature distribution;
a cluster adjustment unit (S440) configured to adjust the plurality of clusters in the visualized feature distribution based on instruction information indicating an instruction from a user using the driving type classification device; and a driving type classification comprising: Device. The driving type classification device according to claim 4,
The learning data classification unit is
By performing principal component analysis on the plurality of driving behavior data in the selected feature space, a plurality of principal component scores corresponding to each of the plurality of selected feature quantities are calculated, and a plurality of principal component scores are calculated based on the plurality of principal component scores. A driving type classification device that generates and displays the visualized feature distribution made up of the three selected feature amounts based on the above. The driving type classification device according to claim 4,
The learning data classification unit is
A principal component analysis process is executed on the plurality of driving behavior data to calculate a first principal component, a second principal component, and a third principal component, and the first principal component, the second principal component, and the third principal component are calculated. Displaying the distribution of the driving behavior data in a dimensionally compressed space composed of principal components and the positions in the dimensionally compressed space of a plurality of feature axes that constitute the selected feature space;
The learning data classification unit is
A driving type classification device that generates and displays the visualized feature distribution made up of the three feature axes based on feature axis instructions input from the outside. The driving type classification device according to any one of claims 4 to 6,
The adjustment of the clusters by the cluster adjustment unit includes manual separation, manual combination, and manual fixation of the clusters. The driving type classification device according to claim 2,
The learning data classification unit is a driving type classification device that classifies the plurality of driving behavior data into the plurality of clusters by performing unsupervised learning in the selected feature space. A plurality of driving behavior data in which one of the plurality of labels representing at least one of the driving operation of the driver driving the vehicle and the behavior of the vehicle appearing as a result of the driving operation is set for each vehicle. Driving scene selection configured to select at least one driving scene as a selected driving scene from among a plurality of preset driving scenes based on a preset selection condition indicating that the driving scene can be classified according to the label. part (S120, S130),
For each of the plurality of driving behavior data corresponding to the plurality of drivers, based on the driving behavior data in the selected driving scene selected by the driving scene selection section, the plurality of driving behavior data are clustered into a plurality of driving behavior data. a learning data classification unit (S140) configured to classify into clusters;
a driving type definition unit (S170) configured to define a driving type for each of the plurality of clusters classified by the learning data classification unit;
Equipped with
The driving scene selection unit is a driving type classification device that selects the selected driving scene based on the classification accuracy of the label in the plurality of driving behavior data. The driving type classification device according to claim 2,
The feature selection unit is a driving type classification device that selects the selected feature amount based on the classification accuracy of the label in the plurality of driving behavior data. The driving type classification device according to claim 1 or 2,
The plurality of labels are selected according to the purpose of the driving type classification device. The driving type classification device according to claim 2,
A driving type classification device that uses cluster labels obtained by clustering using all the driving scenes and all the feature amounts in a plurality of driving behavior data as a plurality of the labels without setting the labels. The driving type classification device according to claim 2,
using at least one of the selection of the selected driving scene by the driving scene selection section, the selection of the selected feature amount by the feature selection section, and at least one of the selected driving scene and the selected feature amount. A driving type classification device comprising a repeated execution determination section (S160) configured to repeatedly perform the classification by the learning data classification section. The driving type classification device according to claim 13,
The iterative execution determination unit automatically determines whether or not the classification results by the learning data classification unit have converged, and when it is determined that the classification results have converged, the repetitive execution determination unit automatically determines whether or not the classification results by the learning data classification unit have converged. and a driving type classification device that terminates the repetition of processing by the learning data classification section. The driving type classification device according to any one of claims 1 to 14,
A driving type classification device in which the plurality of driving behavior data are data obtained by classifying the driving scenes by symbolization. The driving type classification device according to claim 2,
The ratio of the driving topics obtained when the driving behavior data is expressed by a mixture of the driving topics using a plurality of driving topics prepared in advance that represent characteristic patterns of the driving behavior data is defined as a topic ratio. ,
For each of the plurality of selected driving scenes, the topic ratio corresponding to the plurality of driving topics selected by the feature selection unit is extracted from the plurality of driving behavior data, and the topic ratio is determined for each driving topic. an average topic ratio calculation unit (S340) configured to calculate the average of as an average topic ratio;
an integrated topic ratio calculation unit (S350 )and,
an integrated feature calculation unit (S130, S370) configured to calculate an integrated feature by combining the plurality of integrated topic ratios calculated in each of the plurality of selected driving scenes,
The feature selection unit is a driving type classification device in which the integrated topic ratio constituting the integrated feature is the selected feature amount. The driving type classification device according to claim 2,
The driving type definition unit is a driving type classification device that defines the driving type of the cluster based on the position of the cluster in the integrated feature space for each cluster. The driving type classification device according to claim 17,
The driving type definition unit is configured to define, for each cluster, a position in the integrated feature space of the cluster, and a driving type definition table that sets the meaning of the selected feature amount as a driving type element for each selected feature amount. and a driving type classification device that defines the driving type of the cluster. The driving type classification device according to any one of claims 1 to 18,
The driving behavior of the driver of the vehicle is based on the positional relationship between the plurality of clusters classified by the learning data classification unit and the integrated features of the plurality of driving behavior data, and the target driving type of the driver of the vehicle. A driving type classification device comprising a driver FB unit (S510 to S550) configured to notify the driver of a diagnosis result of the type and an improvement step for bringing the driver closer to the target driving type. The driving type classification device according to claim 19,
The driver FB unit is a driving type classification device that notifies the driver of recommended driving behavior for bringing the driver closer to the target driving type, limited to the selected driving scene. The driving type classification device according to any one of claims 1 to 20,
The driving behavior of the driver of the vehicle is based on the positional relationship between the plurality of clusters classified by the learning data classification unit and the integrated features of the plurality of driving behavior data, and the target driving type of the driver of the vehicle. A vehicle control FB unit (S510 to S540) configured to notify an electronic control device that controls the driving of the vehicle of the diagnosis result of the type and improvement steps for bringing the driver closer to the target driving type. , S570). The driving type classification device according to claim 21,
The vehicle control FB unit is a driving type classification device that notifies the driver of a recommended driving behavior to bring the driver closer to the target driving type, limited to the selected driving scene. computer,
A plurality of driving behavior data in which one of the plurality of labels representing at least one of the driving operation of the driver driving the vehicle and the behavior of the vehicle appearing as a result of the driving operation is set for each vehicle. Driving scene selection configured to select at least one driving scene as a selected driving scene from among a plurality of preset driving scenes based on a preset selection condition indicating that the driving scene can be classified according to the label. part (S120, S130),
For each of the plurality of driving behavior data corresponding to the plurality of drivers, based on the driving behavior data in the selected driving scene selected by the driving scene selection section, the plurality of driving behavior data are clustered into a plurality of driving behavior data. a learning data classification unit (S140) configured to classify into clusters, and
a driving type definition unit (S170) configured to define a driving type for each of the plurality of clusters classified by the learning data classification unit;
Driving type classification program to function as a. computer,
A plurality of driving behavior data in which one label among a plurality of labels representing at least one of a driving operation of a driver driving the vehicle and a behavior of the vehicle appearing as a result of the driving operation is set for each vehicle. Driving scene selection configured to select at least one driving scene as a selected driving scene from among a plurality of preset driving scenes based on a preset selection condition indicating that the driving scene can be classified according to the label. part (S120, S130),
For each of the plurality of driving behavior data corresponding to the plurality of drivers, based on the driving behavior data in the selected driving scene selected by the driving scene selection section, the plurality of driving behavior data are clustered into a plurality of driving behavior data. a learning data classification unit (S140) configured to classify into clusters, and
a driving type definition unit (S170) configured to define a driving type for each of the plurality of clusters classified by the learning data classification unit;
function as
The driving scene selection unit includes a feature selection unit (S240, S130) configured to select at least one feature amount selected corresponding to the at least one selected driving scene as a selection feature amount. Prepare,
The learning data classification unit is a driving type classification program that classifies the plurality of driving behavior data into the plurality of clusters in a selected feature space configured with the selected feature quantities selected by the feature selection unit. computer,
A plurality of driving behavior data in which one of the plurality of labels representing at least one of the driving operation of the driver driving the vehicle and the behavior of the vehicle appearing as a result of the driving operation is set for each vehicle. Driving scene selection configured to select at least one driving scene as a selected driving scene from among a plurality of preset driving scenes based on a preset selection condition indicating that the driving scene can be classified according to the label. part (S120, S130),
For each of the plurality of driving behavior data corresponding to the plurality of drivers, based on the driving behavior data in the selected driving scene selected by the driving scene selection section, the plurality of driving behavior data are clustered into a plurality of driving behavior data. a learning data classification unit (S140) configured to classify into clusters, and
a driving type definition unit (S170) configured to define a driving type for each of the plurality of clusters classified by the learning data classification unit;
function as
The driving scene selection unit is a driving type classification program that selects the selected driving scene based on the classification accuracy of the label in the plurality of driving behavior data. A driving type classification method performed by a driving type classification device, comprising:
The driving type classification device sets, for each vehicle, one of a plurality of labels representing at least one of a driving operation of a driver driving the vehicle and a behavior of the vehicle that appears as a result of the driving operation. selecting at least one driving scene as a selected driving scene from among a plurality of preset driving scenes based on preset selection conditions indicating that the plurality of driving behavior data can be classified according to the label;
The driving type classification device classifies a plurality of driving behavior data by clustering, based on the driving behavior data in the selected driving scene, for each of the plurality of driving behavior data corresponding to the plurality of drivers. classified into clusters of
The driving type classification device defines a driving type for each of the plurality of classified clusters. A driving type classification method performed by a driving type classification device, comprising:
The driving type classification device sets, for each vehicle, one of a plurality of labels representing at least one of a driving operation by a driver driving the vehicle and a behavior of the vehicle that appears as a result of the driving operation. selecting at least one driving scene from among a plurality of preset driving scenes as a selected driving scene, based on preset selection conditions indicating that the plurality of driving behavior data can be classified according to the label;
The driving type classification device classifies a plurality of driving behavior data by clustering, based on the driving behavior data in the selected driving scene, for each of the plurality of driving behavior data corresponding to the plurality of drivers. classified into clusters of
The driving type classification device defines a driving type for each of the plurality of classified clusters,
The driving type classification device selects at least one feature amount selected corresponding to the at least one selected driving scene as a selected feature amount,
The driving type classification device is a driving type classification method for classifying a plurality of the driving behavior data into a plurality of clusters in a selected feature space constituted by the selected selected feature amounts. A driving type classification method performed by a driving type classification device, comprising:
The driving type classification device sets, for each vehicle, one of a plurality of labels representing at least one of a driving operation by a driver driving the vehicle and a behavior of the vehicle that appears as a result of the driving operation. selecting at least one driving scene from among a plurality of preset driving scenes as a selected driving scene, based on preset selection conditions indicating that the plurality of driving behavior data can be classified according to the label;
The driving type classification device classifies the plurality of driving behavior data by clustering, based on the driving behavior data in the selected driving scene, for each of the plurality of driving behavior data corresponding to the plurality of drivers. classified into clusters of
The driving type classification device defines a driving type for each of the plurality of classified clusters,
The driving type classification device selects the selected driving scene based on the classification accuracy of the label in the plurality of driving behavior data.