Machine learning device, vehicle testing system, machine learning method, and vehicle testing method

Abstract

The present invention simplifies the reproduction of loads such as road gradient resistance when traveling on an actual road, and comprises: a training data acquisition unit 51 that acquires training data comprising load data including the amount of fluctuation in travel resistance due to vehicle behavior and road gradient resistance of the travel route, and vehicle travel data including the vehicle speed, amount of accelerator pedal operation, and amount of brake pedal operation when the vehicle traveled on the travel route; and a machine learning unit 52 that machine learns correlations between the load data and the vehicle travel data, and generates a simulated gradient prediction model that shows the correlations between the load data and the vehicle travel data.

Description

Machine learning device, vehicle testing system, machine learning method, and vehicle testing method

[0001] The present invention relates to a machine learning device, a vehicle testing system, a machine learning method, and a vehicle testing method for easily reproducing loads such as road gradient loads during actual road driving.

[0002] Conventionally, road load simulations are performed in which a vehicle is simulated to run using a dynamometer installed in a test room based on vehicle running data obtained from actual road running. In this road load simulation, in order to reproduce the running resistance in actual road running, the target running resistance that the dynamometer applies to the vehicle is calculated by dividing the rolling resistance (A+B×V) and the air resistance (C×V ２ ), inertial resistance (I×α), road gradient resistance (M×g×sinθ), etc. In addition to the above-mentioned running resistance, inertial resistance, and road gradient resistance, running resistance resulting from vehicle behavior (acceleration / deceleration, etc.) also occurs during actual road travel.

[0003] Specifically, the target running resistance (RL [N]) is expressed by the following equation (1): RL=A+B×V+C×V ２ + I × α + M × g × sin θ Ｒｏａｄ +F Ｖａｒｉａｂｌｅ Where, V: vehicle speed [m / s], A, B, C: running resistance constants θ Ｒｏａｄ : Road gradient (road gradient in the direction of travel) [deg] M: Test vehicle weight I: Equivalent inertial mass [kg] α: Acceleration [m / s ２ ] g: Gravitational acceleration F Ｖａｒｉａｂｌｅ : Running resistance fluctuation resulting from vehicle behavior (acceleration / deceleration, steering operation, etc.) Note that the values ​​of terms A, B, and C vary depending on the vehicle type, road surface conditions, tire type, etc., and differ for each individual vehicle. Term C is the air resistance coefficient, and terms A and B are coefficients that indicate rolling resistance.

[0004] Here, in the road simulation, the rolling resistance (A + B × V) and air resistance (C × V) ２ ), inertial resistance (I × α), as well as road gradient resistance (M × g × sinθ Ｒｏａｄ) and the running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ However, in the past, it was important to reproduce the running resistance fluctuation (F Ｖａｒｉａｂｌｅ ) is not considered, and the road gradient (θ Ｒｏａｄ ) (and the measurement accuracy of the road gradient is poor), and a gradient command value (torque command value) based on the road gradient is input to the dynamometer.

[0005] On the other hand, road gradient resistance (M × g × sinθ Ｒｏａｄ ) and the running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ ), a method for determining a gradient profile from the result of torque matching has been proposed, as disclosed in Patent Document 1. In this method, the road gradient (θ Ｒｏａｄ ), plus the running resistance fluctuation (F Ｖａｒｉａｂｌｅ ) and the gradient-transformed value (hereinafter referred to as the simulated gradient (θ Ｅｍｕｌａｔｉｏｎ ) can be calculated. As a result, the simulated gradient (θ Ｅｍｕｌａｔｉｏｎ ) can be input to the dynamometer. Patent Document 2 discloses a method for calculating road gradient with high accuracy by combining a gyro sensor, an acceleration sensor, GPS information, etc. Other methods include calculation from values ​​measured during driving using a gyro sensor, wheel torque sensor, or external air pressure measurement, or calculation from altitude information on commercially available topographical maps or high-precision three-dimensional maps.

[0006] However, the method disclosed in Patent Document 1 requires torque matching for each actual road run, which is time-consuming. Furthermore, the method disclosed in Patent Document 2 requires time and effort to install various sensors such as gyro sensors and wheel torque sensors. Furthermore, it is difficult to detect minute changes in road gradient when measuring external air pressure, and it is affected by pressure fluctuation factors other than altitude, such as when passing through a tunnel. Furthermore, the simulated gradient (θ Ｅｍｕｌａｔｉｏｎ In addition, it is difficult to obtain data covering all roads at a high level of granularity at low cost using commercially available topographic maps. Ｅｍｕｌａｔｉｏｎ ) cannot be obtained.

[0007] Special Publication No. 2021-508799 Publication Patent No. 5273599

[0008] Therefore, the present invention has been made in consideration of the above-mentioned problems, and its main object is to simplify the reproduction of loads such as road gradient loads during actual road driving and fluctuations in running resistance resulting from vehicle behavior.

[0009] That is, the machine learning device of the present invention is a machine learning device that generates a simulated gradient prediction model to be used in vehicle testing, and is characterized by comprising: a learning data acquisition unit that acquires learning data consisting of load data including road gradient resistance of a driving route and running resistance fluctuations resulting from vehicle behavior, and vehicle driving data including vehicle speed, accelerator operation amount, and brake operation amount when the vehicle drives the driving route; and a machine learning unit that performs machine learning on the correlation between the load data and the vehicle driving data, and generates a simulated gradient prediction model that indicates the correlation between the load data and the vehicle driving data.

[0010] This machine learning device can calculate the road gradient resistance (M×g×sinθ) of the travel route. Ｒｏａｄ ) and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ ) and vehicle travel data including the vehicle speed, accelerator operation amount, and brake operation amount when the vehicle travels along the travel route. Therefore, by inputting various vehicle travel data into the simulated gradient prediction model, it is possible to generate a simulated gradient prediction model that shows the correlation between the load data including Ｒｏａｄ ) and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ As a result, the road gradient resistance (M×g×sinθ) during actual road travel can be calculated. Ｒｏａｄ ) and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ ) and other loads can be easily reproduced.

[0011] Here, it is desirable that the load data include a load calculated from a simulated gradient that takes into account a gradient-converted value of a running resistance variation due to vehicle behavior to the road gradient of the travel route.

[0012] In a specific embodiment of the present invention, the learning data acquisition unit acquires learning data consisting of load data used in torque matching or road simulation and vehicle driving data used at that time, and the machine learning unit performs machine learning on the correlation between the acquired load data and the vehicle driving data, and generates a simulated gradient prediction model that indicates the correlation between the load data and the vehicle driving data.

[0013] Different vehicles have different pedal characteristics. Furthermore, even within the same vehicle, changing the pedal settings (e.g., pedal settings for eco mode and sport mode) can result in different pedal characteristics for each mode. In this case, to apply the generated simulated gradient prediction model to vehicles with different pedal characteristics, the machine learning unit desirably uses a pedal coefficient indicating the pedal characteristics of the vehicle in the simulated gradient prediction model to generate a simulated gradient prediction model corresponding to the pedal coefficient. With this configuration, when the generated simulated gradient prediction model is used for a different vehicle, a simulated gradient prediction model corresponding to the different vehicle can be easily generated. Furthermore, when the pedal settings are changed within the same vehicle, a simulated gradient prediction model corresponding to the changed pedal settings can be easily generated.

[0014] It is desirable that the learning data acquisition unit acquires, as the learning data, a road model including a correct or known road gradient of at least a part of the travel route, and the machine learning unit performs machine learning on a correlation between the simulated gradient excluding the road gradient and vehicle travel data, and generates a vehicle behavior load model indicating the correlation between the vehicle behavior load data obtained by excluding the road gradient from the load data and the vehicle travel data. With this configuration, it is desirable to calculate the road gradient (θ Ｒｏａｄ ) is known, road gradient resistance (M × g × sinθ Ｒｏａｄ ) and the running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ ) can be separated.

[0015] It is desirable that the learning data acquisition unit acquires first learning data consisting of the load data, the vehicle driving data, and the road model on which the first vehicle has driven for a first vehicle, and second learning data consisting of the load data, the vehicle driving data, and the road model on which the second vehicle has driven for a second vehicle, and that the machine learning unit generates the vehicle behavior load model of the first vehicle from the first learning data of the first vehicle, and generates the vehicle behavior load model of the second vehicle from the second learning data of the second vehicle.

[0016] The vehicle testing system according to the present invention is a vehicle testing system having a vehicle testing device such as a dynamometer that applies a load to a vehicle or a part of a vehicle as a test object during running, and is configured to measure the road gradient resistance (M×g×sinθ) of a running route. Ｒｏａｄ ) and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ and a model storage unit that stores a simulated gradient prediction model that indicates a correlation between load data including a vehicle speed, an accelerator operation amount, and a brake operation amount when the vehicle travels along the travel route, and vehicle travel data including the vehicle speed, an accelerator operation amount, and a brake operation amount when the vehicle travels along the travel route, and a load data calculation unit that calculates load data to be input to the vehicle testing device from the vehicle travel data and the simulated gradient prediction model.

[0017] It is desirable that the model storage unit stores vehicle behavior load data indicating the running resistance variation resulting from the vehicle behavior and a vehicle behavior load model indicating a correlation with the vehicle driving data, and that the load data calculation unit calculates the vehicle behavior load data from the vehicle driving data and the vehicle behavior load model.

[0018] The second vehicle is tested using data from a first vehicle, and it is desirable that the load data calculation unit calculates load data to be input into a vehicle testing device that tests the second vehicle using a road model of the driving route traveled by the first vehicle and a vehicle behavior load model that shows the correlation between vehicle behavior load data that indicates the driving resistance variation resulting from the vehicle behavior of the second vehicle and the vehicle driving data.

[0019] The second vehicle is preferably a vehicle with an advanced driver assistance system (ADAS vehicle) or an autonomous driving vehicle (AD vehicle), for example. However, the second vehicle is not limited to an ADAS vehicle or an AD vehicle.

[0020] Furthermore, the machine learning method according to the present invention is preferably a machine learning method used for vehicle testing, which acquires learning data consisting of load data including road gradient resistance of a driving route and running resistance fluctuations resulting from vehicle behavior, and vehicle driving data including vehicle speed, accelerator operation amount, and brake operation amount when the vehicle travels on the driving route, and performs machine learning on the correlation between the load data and the vehicle driving data, thereby generating a simulated gradient prediction model that indicates the correlation between the load data and the vehicle driving data.

[0021] Furthermore, the vehicle testing method of the present invention is a vehicle testing method that uses a dynamometer to apply a load to a vehicle or a part of a vehicle as a test specimen during driving, and is characterized in that the load data is calculated from the vehicle driving data using a simulated gradient prediction model that shows a correlation between load data including road gradient resistance of the driving route and driving resistance fluctuations resulting from vehicle behavior, and vehicle driving data including vehicle speed, accelerator operation amount, and brake operation amount when the vehicle drives the driving route, and the calculated load data is input into the dynamometer to test the vehicle.

[0022] According to the present invention configured in this way, it is possible to easily reproduce the load such as gradient resistance of actual road travel.

[0023] FIG. 1 is a schematic diagram showing a vehicle testing system according to a first embodiment of the present invention; FIG. 2 is a schematic diagram showing machine learning and gradient prediction in the first embodiment; FIG. 3 is a schematic diagram showing a procedure of the vehicle testing system according to the first embodiment; FIG. 4 is a schematic diagram showing a procedure of the vehicle testing system according to the second embodiment; FIG. 5 is a schematic diagram showing generation of a simulated gradient prediction model in the third embodiment; FIG. 6 is a schematic diagram showing a procedure of the vehicle testing system according to the fourth embodiment; Ｅｍｕｌａｔｉｏｎ 10 is a diagram showing a procedure for calculating load data (simulated gradient (θ )) in a modified embodiment. FIG. 11 is a diagram showing a procedure for calculating load data (simulated gradient (θ )) in a modified embodiment. FIG. 12 is a diagram showing a procedure for calculating load data (simulated gradient (θ )) in a modified embodiment.Ｅｍｕｌａｔｉｏｎ ) is a diagram showing a calculation procedure of

[0024] Hereinafter, embodiments of a vehicle testing system according to the present invention will be described with reference to the drawings. Note that, for ease of understanding, the drawings shown below are schematically illustrated with appropriate omissions or exaggerations. Identical components are designated by the same reference numerals, and their descriptions will be omitted where appropriate.

[0025] First Embodiment As shown in FIG. 1, a vehicle testing system 100 according to a first embodiment reproduces the running load of an actual road run in a test room by road simulation control.

[0026] Specifically, the vehicle testing system 100 has a chassis dynamometer 2 for simulating a run of a vehicle under test, and a control device 3 for controlling the chassis dynamometer 2. The chassis dynamometer 2 has rollers 21 on which the vehicle tires are placed, and a dynamometer 22 connected to the rollers 21 and applying a load to the vehicle W via the rollers 21. The vehicle may not only be a completed vehicle, but also a part of a vehicle having an engine, a motor, a powertrain, or the like.

[0027] Before the vehicle test is performed using the vehicle test system 100, the following steps (1) to (3) are performed: (1) Road test, (2) Torque matching, and (3) Machine learning of a simulated gradient prediction model.

[0028] (1) Road Test This road test involves traveling on an actual road and acquiring vehicle travel data, such as vehicle speed, accelerator operation amount, brake operation amount, or steering operation amount, during travel on the actual road. The accelerator operation amount may be accelerator pedal position or throttle opening. The brake operation amount may be brake pedal position or brake pressure. The steering operation amount may be steering angle or a difference in left and right wheel speeds. Furthermore, in this road test, vehicle information acquired may include, for example, the voltage or current of the battery, motor, etc., the warm-up state of the internal combustion engine (e.g., coolant temperature or engine oil temperature), the state of charge or deterioration of the battery, or the temperature inside or around the battery, motor, etc.; environmental information acquired may include the temperature, humidity, or atmospheric pressure around the vehicle; and location information acquired may include latitude, longitude, or altitude. Additionally, the amount of emission of target components contained in exhaust gas emitted from the vehicle's internal combustion engine may be measured.

[0029] (2) Torque Matching This torque matching involves running a test vehicle W on the chassis dynamometer 2 to obtain the dynamo load (torque matching load (RepForce)) that the chassis dynamometer 2 applies to the vehicle W.

[0030] Specifically, torque matching involves controlling the driving robot 4 mounted on the vehicle W to drive the vehicle W to the accelerator operation amount, brake operation amount, and / or shift position required for the road test, while controlling the speed of the chassis dynamometer 2 so that the vehicle speed of the vehicle W matches the speed required for the road test. The dynamo load obtained from the dynamometer through these controls is acquired. The dynamo load thus acquired (torque matching load (RepForce)) is used for machine learning, as described below. Furthermore, torque matching is used to calculate the road gradient (θ Ｒｏａｄ ), plus the running resistance fluctuation (F Ｖａｒｉａｂｌｅ ) and the gradient-transformed value (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ ) can be obtained.

[0031] θ Ｅｍｕｌａｔｉｏｎ The derivation method is expressed as the following formula: RepForce = A + BV + CV２ + Iα + Mg sin θ Ｒｏａｄ +F Ｖａｒｉａｂｌｅ Mg sinθ Ｅｍｕｌａｔｉｏｎ = Mg sin θ Ｒｏａｄ +F Ｖａｒｉａｂｌｅ θ Ｅｍｕｌａｔｉｏｎ = sin －１ ((RepForce-(A+BV+CV ２ + Iα)) / Mg)

[0032] (3) Machine Learning of Simulated Slope Prediction Model This machine learning is performed by a machine learning device 5 shown in Fig. 1. The machine learning device 5 has a learning data acquisition unit 51 that acquires learning data for generating a simulated slope prediction model, and a machine learning unit 52 that generates a simulated slope prediction model using the learning data.

[0033] As shown in "During Learning" in FIG. 2, the learning data consists of load data indicating the dynamo load obtained by the torque matching described above, and vehicle running data used in the torque matching.

[0034] The load data used for the learning data is, for example, the road gradient (θ Ｒｏａｄ ), plus the running resistance fluctuation (F Ｖａｒｉａｂｌｅ ) is converted into a gradient and added (here, added) to the value (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) and "Gradient resistance (M × g × sinθ Ｒｏａｄ ) and "running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ ) converted into gradient resistance.

[0035] In addition, the vehicle driving data used for learning data is the vehicle speed, accelerator operation amount, brake operation amount, steering operation amount, etc. when driving along the driving route, or calculated values ​​such as square values ​​or differential values ​​obtained from these.

[0036] The machine learning unit 52 performs machine learning to learn the correlation between the load data obtained by the torque matching and the vehicle running data used in the torque matching, and calculates the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ)) and vehicle travel data is used to generate a simulated gradient prediction model (machine learning model). As shown in "Prediction" in Figure 2, by inputting vehicle travel data into this simulated gradient prediction model, a simulated gradient (θ Ｅｍｕｌａｔｉｏｎ The machine learning algorithm in the machine learning unit 52 is a machine learning algorithm (for example, a recurrent neural network (RNN)) that inputs past information and uses it to predict the present and future.

[0037] <Test Procedure of First Embodiment> The test procedure of the first embodiment is shown in Fig. 3. A first road test (Road Test no. 1) is conducted on a vehicle as a test specimen to obtain vehicle driving data that will serve as learning data.

[0038] Here, it is desirable that the vehicle driving data used as learning data include a wide driving range. Therefore, the longer the number of trips or the driving time, the wider the driving range, and the learning accuracy or prediction accuracy of machine learning tends to improve. For example, a driving test that covers a wide range of calculated values ​​such as vehicle speed, acceleration / deceleration, steering operation amount, road gradient, or their squared values ​​or differential values, and that covers a wide range of combinations of these is desirable. In detail, for example, a 30-minute driving test may include a combination of urban areas, mountain roads, and expressways. Furthermore, it is also conceivable to obtain vehicle driving data by conducting driving tests with multiple drivers on each of multiple driving routes.

[0039] Then, torque matching is performed using the vehicle running data obtained in this first road running test, and the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ ) to generate a

[0040] This load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ)), the dynamometer control unit 33 performs road load simulation (RLS) control of the dynamometer 22, thereby conducting a simulated running test of the vehicle W on the chassis dynamometer 2. Note that while it is considered that the vehicle is driven using the driving robot 4 in the road load simulation (RLS) control, it may also be driven by a person.

[0041] On the other hand, the vehicle running data obtained in the first road running test and the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) are used as learning data to learn the correlation between them through machine learning, and the vehicle driving data and load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) is generated. This simulated gradient prediction model is stored in the model storage unit 31 of the control device 3 (see FIG. 1). Note that the simulated gradient prediction model may be stored in an information processing device separate from the control device 3.

[0042] When the second to Nth road driving tests (Road Test no. 2 to N) are carried out, the vehicle driving data obtained in those road driving tests is input to the control device 3 without torque matching. As a result, the load data calculation unit 32 of the control device 3 inputs the vehicle driving data to the simulated gradient prediction model to generate load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) is calculated. Ｅｍｕｌａｔｉｏｎ )), the dynamometer control unit 33 performs road load simulation (RLS) control of the dynamometer 22, thereby conducting a simulated running test of the vehicle W on the chassis dynamometer 2. Note that while it is considered that the vehicle is driven using the driving robot 4 in the road load simulation (RLS) control, it may also be driven by a person.

[0043] <Effects of First Embodiment> According to the vehicle testing system 100 of this embodiment configured as described above, the gradient resistance (M×g×sinθ) of the travel route can be calculated. Ｒｏａｄ ) and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ Load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ)) and vehicle travel data including the vehicle speed, accelerator operation amount, and brake operation amount when the vehicle travels along the travel route. Therefore, by inputting various vehicle travel data into the simulated gradient prediction model, it is possible to generate a simulated gradient prediction model that shows a correlation between the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ As a result, the gradient resistance (M×g×sinθ) of the actual road can be obtained. Ｒｏａｄ ) and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ ) and other loads can be easily reproduced.

[0044] Second Embodiment Next, a vehicle testing method and a machine learning method according to a second embodiment of the present invention will be described.

[0045] In the vehicle testing method and machine learning method of the second embodiment, the machine learning method of the machine learning device 5 is different from that of the first embodiment. As shown in FIG. 4 , a simulated gradient prediction model is generated without performing torque matching using vehicle driving data obtained from a road driving test.

[0046] Specifically, the learning data acquisition unit 51 of the machine learning device 5 acquires learning data consisting of load data (input gradient θ) used in an arbitrary road load simulation (RLS) and vehicle driving data obtained in a road driving test. Note that the vehicle driving data is generated from actual driving data obtained by a real vehicle driving on an actual driving route, or generated from virtual driving data obtained by a virtual vehicle driving on a virtual driving route.

[0047] The machine learning unit 52 then performs machine learning to learn the correlation between the load data (input gradient θ) used in the road simulation and the vehicle travel data used in the road simulation, and generates a gradient prediction model (machine learning model) that shows the correlation between the load data (input gradient θ) and the vehicle travel data. By inputting the vehicle travel data into this gradient prediction model, a simulated gradient (θ) of the travel route from which the vehicle travel data was obtained is calculated. Ｅｍｕｌａｔｉｏｎ ) can be calculated.

[0048] <Effects of the Second Embodiment> According to the second embodiment, a simulated gradient prediction model is generated without performing torque matching using vehicle travel data obtained from a road travel test, thereby making it possible to predict the gradient resistance (M×g×sinθ) of actual road travel. Ｒｏａｄ ) and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ ) and other loads can be easily reproduced.

[0049] Third Embodiment Next, a vehicle testing method and a machine learning method according to a third embodiment of the present invention will be described.

[0050] In the vehicle testing method and machine learning method of the third embodiment, the configuration of the machine learning unit 52 of the machine learning device 5 differs from that of the above-described embodiments.

[0051] Specifically, the machine learning unit 52 generates a simulated gradient prediction model for each pedal characteristic using the simulated gradient prediction model generated in the first or second embodiment and different pedal coefficients. The pedal coefficient is a coefficient that indicates the pedal characteristic and can be expressed as a value obtained by dividing the driving force by the pedal depression amount. Here, the calculation of this value may take into account vehicle speed or acceleration.

[0052] For example, as shown in FIG. 5, the machine learning unit 52 may use the simulated gradient prediction model f generated in the first vehicle for different vehicles (first vehicle and second vehicle). Ａ (x) and the pedal coefficient K of the first vehicle Ａ and the pedal coefficient K of the second vehicle Ｂ The ratio (K Ｂ / K Ａ ) and the simulated gradient prediction model f of the second vehicle Ｂ (x) to f Ｂ (x) = f Ａ (x) × (K Ｂ / K Ａ ) is calculated.

[0053] Furthermore, even for the same vehicle, when the pedal setting (pedal setting for eco mode, pedal setting for sport mode, etc.) is changed, the machine learning unit 52 may change the generated simulated gradient prediction model f Ａ (x) and the pedal coefficient K of the pedal setting when the simulated gradient prediction model is generated.Ａ and the pedal coefficient K of the changed pedal setting Ｃ The ratio (K Ｃ / K Ａ ) and the simulated gradient prediction model f of the changed pedal setting Ｃ (x) to f Ｃ (x) = f Ａ (x) × (K Ｃ / K Ａ ) is calculated.

[0054] Advantages of the Third Embodiment In addition to the advantages of the first and second embodiments, the vehicle testing system 100 of this embodiment configured as described above can easily generate a simulated gradient prediction model for a different vehicle when the generated simulated gradient prediction model is used for the different vehicle. Furthermore, when the pedal setting is changed in the same vehicle, a simulated gradient prediction model for the changed pedal setting can easily be generated.

[0055] Fourth Embodiment Next, a vehicle testing method and a machine learning method according to a fourth embodiment of the present invention will be described.

[0056] In the vehicle testing method and machine learning method of the fourth embodiment, the configuration of the machine learning unit 52 of the machine learning device 5 differs from that of the above-described embodiments.

[0057] Specifically, as shown in FIG. 6 , the learning data acquisition unit 51 of the machine learning device 5 acquires the road gradient (θ Ｒｏａｄ The road model includes vehicle information (for example, vehicle speed or longitude and latitude position information) and road gradient (θ Ｒｏａｄ ) The gradient output model may be road map data.

[0058] Then, the machine learning unit 52 performs machine learning on the correlation between the road model, the simulated gradient excluding the road gradient, and the vehicle travel data, and calculates the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) to the road gradient (θ Ｒｏａｄ A vehicle behavior load model is generated that indicates the correlation between the vehicle behavior load data excluding the vehicle behavior load data and the vehicle driving data.

[0059] FIG. 6 illustrates an example of a procedure for generating a vehicle behavior load model and a road model for each road running test.

[0060] In the road driving test, the road gradient (θ Ｒｏａｄ The road model is then acquired by defining the section where the road gradient (θ ) is known as Road Test No. 1. Then, the road model is acquired together with the vehicle travel data and the load data obtained by torque matching (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) is obtained as training data.

[0061] Then, the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) to the road gradient (θ Ｒｏａｄ Vehicle behavior load data (sin －１ (F Ｖａｒｉａｂｌｅ / Mg)) and vehicle driving data are machine-learned to determine the correlation between the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) to the road gradient (θ Ｒｏａｄ Vehicle behavior load data (sin －１ (F Ｖａｒｉａｂｌｅ A vehicle behavior load model is generated that indicates the correlation between the vehicle speed and the vehicle driving data.

[0062] When the second to Nth road driving tests (Road Test no. 2 to N) were conducted, the gradient resistance (M×g×sinθ) of the driving route in each road driving test was calculated by subtracting the “vehicle behavior load data obtained from the vehicle driving data of each road driving test and the above-mentioned vehicle behavior load model” from the “load data obtained from the vehicle driving data of each road driving test and the above-mentioned simulated gradient prediction model.” Ｒｏａｄ ) can be obtained. The road gradient (θ Ｒｏａｄ ) and latitude and longitude information, a road model for each road test can be generated.

[0063] Load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ 7, the calculation of the running resistance variation (F) due to the vehicle behavior will be described. The load data calculation unit 32 of the control device 3 inputs vehicle running data (vehicle speed, accelerator operation amount, brake operation amount, etc.) into the vehicle behavior load model, and calculates the running resistance variation (F ＶａｒｉａｂｌｅThe load data calculation unit 32 also inputs vehicle travel data (vehicle speed, longitude, latitude, etc.) into a gradient output model included in the road model to calculate the road gradient (θ Ｒｏａｄ ) and calculate the road gradient (θ Ｒｏａｄ ) is input into a model that calculates the load due to the road gradient to calculate the gradient load (M × g × sinθ Ｒｏａｄ Furthermore, the load data calculation unit 32 calculates the running resistance fluctuation (F Ｖａｒｉａｂｌｅ ) and gradient load (M × g × sinθ Ｒｏａｄ ) are added together. The load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ Then, the dynamometer control unit 33 calculates the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) is used to perform road load simulation (RLS) control of the dynamometer 22, thereby conducting a simulated running test of the vehicle W on the chassis dynamometer 2. Note that while it is considered that the vehicle is driven using a driving robot 4 in the road load simulation (RLS) control, it may also be driven by a person.

[0064] <Effects of Fourth Embodiment> According to the fourth embodiment, in addition to the effects of the above-described embodiments, the road gradient resistance (M×g×sinθ Ｒｏａｄ ) load model and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ As a result, road gradient resistance (M×g×sinθ) can be calculated from any vehicle travel data. Ｒｏａｄ ) load and the running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ ) can be obtained separately, and the resulting road model can be used for testing and simulations with other vehicles.

[0065] Fifth Embodiment Next, a vehicle testing method and a machine learning method according to a fifth embodiment of the present invention will be described.

[0066] The vehicle testing method and machine learning method of the fifth embodiment generate a vehicle behavior load model of a first vehicle and / or a road model of a route traveled by the first vehicle, and generate a vehicle behavior load model of a second vehicle and / or a road model of a route traveled by the second vehicle.

[0067] Specifically, the learning data acquisition unit 51 of the machine learning device 5 acquires load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )), first learning data consisting of vehicle travel data and a road model, and load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ ) and second learning data consisting of vehicle travel data and road models.

[0068] Then, the machine learning unit 52 generates a vehicle behavior load model of the first vehicle and / or a road model of the route traveled by the first vehicle from the first learning data of the first vehicle, and generates a vehicle behavior load model of the second vehicle and / or a road model of the route traveled by the second vehicle from the second learning data of the second vehicle.

[0069] <Vehicle Evaluation in Fifth Embodiment> The machine learning device 5 can build a road model database, a vehicle behavior load model database, or a driver model database from past test results of various vehicles (first vehicles), as shown in Figures 8 and 9. The driver model is a model that indicates the driving characteristics of a driver, and may be obtained from data of past vehicle tests (vehicle speed, engine state, accelerator and pedal usage, brake pedal usage, steering angle, etc.), may be obtained from time-series vehicle speed and shift data, may be obtained by a predetermined calculation, or a simple vehicle speed profile may be used as is.

[0070] When the vehicle to be developed (second vehicle) is an AD vehicle (see FIG. 8), an evaluation is performed by combining a vehicle behavior load model obtained from the results of road driving tests of the vehicle to be developed, a driving method using AD control (AD system) for the vehicle to be developed, and a road model in the database. This makes it possible to evaluate the AD vehicle in various scenarios (driving profiles) such as various driving routes. For example, it is possible to evaluate whether the AD system operates correctly or whether the dynamic behavior of the vehicle is as expected.

[0071] When the vehicle to be developed (second vehicle) is an ADAS vehicle (see FIG. 9 ), an evaluation is performed by combining a vehicle behavior load model obtained from the results of road driving tests of the vehicle to be developed with a road model and a driver model stored in the database. This allows the ADAS vehicle to be evaluated in various scenarios (driving profiles), such as various driving routes. For example, it is possible to evaluate whether the ADAS system operates correctly or whether the dynamic behavior of the vehicle is as expected. The vehicle may be driven using a driving robot 4 during this evaluation, but it may also be driven by a person.

[0072] It is anticipated that the above-mentioned vehicle testing of AD / ADAS vehicles will be conducted when a new AD / ADAS system is developed, when the same AD / ADAS system is applied to another vehicle, or when an AD / ADAS system is added as an add-on to an existing vehicle.

[0073] <Effects of Fifth Embodiment> With the vehicle testing method of the fifth embodiment, various scenarios can be generated using a gradient load model, a driver model, and the like of the first vehicle obtained from past vehicle tests. As a result, a large number of experience-based evaluation scenarios can be prepared. In particular, by executing these many scenarios on a bench test device or in a simulation, it is possible to perform high-quality evaluations (with experience-based scenarios and reproducible test methods) in a short amount of time compared to tests on public roads or test courses.

[0074] Other Embodiments For example, the machine learning unit 52 may calculate the road gradient (θ Ｒｏａｄ) and load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) and vehicle driving data, and the correlation is machine-learned to calculate the gradient resistance (M × g × sinθ Ｒｏａｄ ) and vehicle driving data, and a gradient load model showing the correlation between the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) to the road gradient (θ Ｒｏａｄ ) may be excluded from the vehicle behavior load model, which indicates the correlation between the vehicle behavior load data and the vehicle driving data.

[0075] In this case, as shown in FIG. 10, the load data calculation unit 32 of the control device 3 inputs vehicle running data (vehicle speed, accelerator operation amount, brake operation amount, etc.) into the vehicle behavior load model, and calculates the running resistance fluctuation (F Ｖａｒｉａｂｌｅ The load data calculation unit 32 inputs vehicle travel data (vehicle speed, accelerator operation amount, brake operation amount, etc.) into the gradient load model to calculate a gradient load (M×g×sinθ Ｒｏａｄ Furthermore, the load data calculation unit 32 calculates the running resistance fluctuation (F Ｖａｒｉａｂｌｅ ) and gradient load (M × g × sinθ Ｒｏａｄ ) is added up. This added value is the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ Then, the dynamometer control unit 33 calculates the load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) is used to perform road load simulation (RLS) control of the dynamometer 22, whereby a simulated running test of the vehicle W is performed on the chassis dynamometer 2.

[0076] Furthermore, although the vehicle testing system in each of the above embodiments uses a chassis dynamometer that applies a load to the vehicle, it may also use a dynamometer that applies a load to a part of the vehicle (for example, a powertrain, an engine, etc.). In other words, the machine learning model obtained in each of the above embodiments can be used in a test using a part of the vehicle as a test specimen or in a full simulation test.

[0077] Furthermore, when the test subject is an ADAS vehicle or an AD vehicle, the accelerator operation amount, brake operation amount, etc. included in the vehicle driving data may not be physical operation amounts, but may be replaced by vehicle signals.

[0078] Furthermore, the learning data may include the road surface conditions of the travel route (e.g., road surface friction coefficient, puddles, unevenness, etc.), which allows for the generation of a gradient prediction model that takes into account the road surface conditions of the travel route.

[0079] Furthermore, the learning data may include vehicle information such as vehicle weight, temperatures of various parts of the vehicle, battery charge status, etc. This makes it possible to generate a simulated gradient prediction model, a vehicle behavior load model, or a road gradient load model that takes into account differences in vehicle conditions.

[0080] The learning data may also include environmental information such as ambient temperature, humidity, altitude, etc. This allows a simulated slope prediction model to be generated that takes environmental information into account.

[0081] The load data used for learning data is the road gradient (θ Ｒｏａｄ ), plus the running resistance fluctuation (F Ｖａｒｉａｂｌｅ ) and the gradient-transformed value (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )), but the overall running load (rolling resistance (A + B × V), air resistance (C × V ２ ), inertial resistance (I × α), gradient resistance (M × g × sinθ Ｒｏａｄ ) and running resistance fluctuation due to vehicle behavior (F Ｖａｒｉａｂｌｅ )) can also be used.

[0082] In each of the above embodiments, the calculated load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) was input to the dynamometer 22 to perform a simulated running test of the vehicle as a test specimen, but the calculated load data (simulated gradient (θ Ｅｍｕｌａｔｉｏｎ )) to perform a simulation on a computer (test in a virtual environment).

[0083] In addition, in each of the above embodiments, load data obtained by torque matching was used, but it is also possible to use loads measured during actual road driving, such as loads calculated from the torque applied to the axles, for machine learning.

[0084] In addition, various modifications and combinations of the embodiments may be made as long as they do not go against the spirit of the present invention.

[0085] According to the present invention, it is possible to easily reproduce loads such as road gradient resistance during actual road travel.

[0086] 100: Vehicle testing system 2: Chassis dynamometer 3: Control device 31: Model storage unit 32: Load data calculation unit 33: Dynamometer control unit 4: Driving robot 5: Machine learning device 51: Learning data acquisition unit 52: Machine learning unit

Claims

1. A machine learning device for generating a simulated gradient prediction model for use in vehicle testing, comprising: a learning data acquisition unit that acquires learning data including load data including a road gradient resistance of a travel route and a running resistance fluctuation due to a vehicle behavior, and vehicle travel data including a vehicle speed, an accelerator operation amount, and a brake operation amount when the vehicle travels on the travel route; a machine learning unit that performs machine learning on a correlation between the load data and the vehicle travel data, and generates a simulated gradient prediction model that indicates the correlation between the load data and the vehicle travel data.

2. The machine learning device according to claim 1 , wherein the load data includes a load calculated from a simulated gradient of the road gradient of the travel route, the simulated gradient being calculated by adding a gradient-converted value of a running resistance variation caused by vehicle behavior.

3. The learning data acquisition unit acquires learning data including load data used in a road simulation and vehicle travel data used at that time, 3. The machine learning device according to claim 1, wherein the machine learning unit performs machine learning on a correlation between the acquired load data and the vehicle driving data, and generates a simulated gradient prediction model indicating the correlation between the load data and the vehicle driving data.

4. The machine learning device according to claim 1 , wherein the machine learning unit uses a pedal coefficient indicating a pedal characteristic of a vehicle in the simulated gradient prediction model to generate a simulated gradient prediction model corresponding to the pedal coefficient.

5. the learning data acquisition unit acquires, as the learning data, a road model including a correct or known road gradient of at least a part of the travel route; 3. The machine learning device according to claim 2, wherein the machine learning unit performs machine learning on a correlation between the simulated gradient excluding the road gradient and vehicle driving data, and generates a vehicle behavior load model indicating a correlation between the vehicle behavior load data excluding the road gradient from the load data and the vehicle driving data.

6. the learning data acquisition unit acquires first learning data consisting of the load data, the vehicle travel data, and the road model on which the first vehicle has traveled, for a first vehicle, and second learning data consisting of the load data, the vehicle travel data, and the road model on which the second vehicle has traveled, for a second vehicle; The machine learning device according to claim 5 , wherein the machine learning unit generates the vehicle behavior load model of the first vehicle from first learning data of the first vehicle, and generates the vehicle behavior load model of the second vehicle from second learning data of the second vehicle.

7. A vehicle testing system having a dynamometer that applies a load during running to a test specimen that is a vehicle or a part thereof, a model storage unit for storing a simulated gradient prediction model indicating a correlation between load data including a road gradient resistance of a travel route and a running resistance variation due to a vehicle behavior, and vehicle running data including a vehicle speed, an accelerator operation amount, and a brake operation amount when the vehicle runs on the travel route; a load data calculation unit that calculates load data to be input to the dynamometer from the vehicle driving data and the simulated gradient prediction model.

8. the model storage unit stores vehicle behavior load data indicating a running resistance variation caused by the vehicle behavior and a vehicle behavior load model indicating a correlation between the vehicle behavior load data and the vehicle running data, The vehicle testing system according to claim 7 , wherein the load data calculation unit calculates the vehicle behavior load data from the vehicle running data and the vehicle behavior load model.

9. A test of a second vehicle is performed using data of the first vehicle; 9. The vehicle testing system according to claim 8, wherein the load data calculation unit calculates load data to be input to a vehicle testing device that tests the second vehicle, using a road model of a driving route traveled by the first vehicle and a vehicle behavior load model that indicates a correlation between vehicle behavior load data indicating a driving resistance variation resulting from a vehicle behavior of the second vehicle and the vehicle driving data.

10. 10. The vehicle testing system of claim 9, wherein the second vehicle is a vehicle having an advanced driver assistance system or an autonomous vehicle.

11. 1. A machine learning method for use in vehicle testing, comprising: Acquire learning data including load data including a road gradient resistance of a travel route and a travel resistance variation due to a vehicle behavior, and vehicle travel data including a vehicle speed, an accelerator operation amount, and a brake operation amount when the vehicle travels on the travel route; A machine learning method for performing machine learning on a correlation between the load data and the vehicle driving data, and generating a simulated gradient prediction model indicating the correlation between the load data and the vehicle driving data.

12. A vehicle testing method using a dynamometer to apply a load during running to a test specimen that is a vehicle or a part thereof, comprising: calculating the load data from the vehicle travel data using a simulated gradient prediction model that indicates a correlation between load data including a road gradient resistance of a travel route and a travel resistance variation due to a vehicle behavior, and vehicle travel data including a vehicle speed, an accelerator operation amount, and a brake operation amount when the vehicle travels on the travel route; The vehicle testing method includes inputting the calculated load data into the dynamometer to test the vehicle.