Data device, learning data generation device, data conversion program product, learning data generation program product, and control system

Abstract

The invention provides a data device, a learning data generation device, a data conversion program product, a learning data generation program product, and a control system, which facilitate the construction of a machine learning model used for controlling a mechanical device. The data device includes a reference conversion unit and a re-conversion unit. The reference conversion unit inputs a reference operation parameter output from a reference machine device control model constructed by machine learning of an operation on a reference machine device, converts the reference operation parameter into an operation parameter independent of each machine device, and outputs the operation parameter. The re-conversion unit re-converts the operation parameter to a control operation parameter, which is an operation parameter of the same type as the reference operation parameter, on the basis of the mechanical characteristics of the mechanical device to be controlled.

Description

Technical Field

[0001] This disclosure relates to data devices for controlling mechanical devices, learning data generation devices, data transformation program products, learning data generation program products, and control systems. Background Technology

[0002] Patent document 1 describes a neural network system for autonomous driving of self-driving vehicles.

[0003] In recent years, data-driven machine control methods based on machine learning have attracted attention as a solution for controlling appropriate mechanical devices under complex conditions involving various factors. Traditional rule-based and model-based methods construct algorithms based on prior knowledge and mathematical models. In contrast, machine learning constructs algorithms based on data, hence the term "data-driven." Data-driven planners do not require prior knowledge or mathematical models; on the other hand, they require a large amount of learning data to enable the machine learning model to function. Data-driven planners take the surrounding information and state information of the mechanical device as input and output the operational parameters of the mechanical device.

[0004] Existing technical documents Patent documents Patent Document 1: Japanese Patent Publication No. 2019-533810 Summary of the Invention

[0005] The problem that the invention aims to solve The inventors' detailed research revealed the following problems. Model-based planners clearly divide tasks into identification, judgment, and operation roles, and the parameters within each role, as well as the variables exchanged between roles, are interpretable and clear. Therefore, parameter adjustments are easily made in response to design changes. On the other hand, data-driven planners lack clear divisions into identification, judgment, and operation roles and are composed of large-scale neural network models. Therefore, the parameters and variables within data-driven planners are difficult to interpret. Consequently, parameter adjustments are difficult in response to design changes. Therefore, in data-driven planners, whenever the specifications of the mechanical device are changed, the learning data needs to be changed and relearning is required. Furthermore, when applying to mechanical devices with different models than those for which learning data was acquired, the different characteristics of each mechanical device may result in undesirable behavior. In such cases, relearning must be performed from scratch for each mechanical device.

[0006] The purpose of this disclosure is to make it easier to construct machine learning models for controlling mechanical devices.

[0007] Solution for solving the problem like Figure 9As shown, one aspect of this disclosure is a data device that includes a reference conversion unit and a re-conversion unit, which outputs control operation parameters that have been re-converted by the re-conversion unit to the controlled mechanical device.

[0008] The reference transformation unit is configured to: input at least one reference operating parameter, which is the operating parameter output by a machine learning model (i.e., a reference mechanical device control model) constructed by machine learning for the operation of a mechanical device that serves as a reference, and use a physical model (i.e., a mechanical device model) based on the mechanical characteristics of the reference mechanical device to transform at least one reference operating parameter into at least one action parameter that is independent of each mechanical device, and output the transformed at least one action parameter.

[0009] The re-transformation unit is configured to transform at least one action parameter output by the reference transformation unit into at least one control operation parameter, which is the same type as at least one reference operation parameter, based on the mechanical characteristics of the mechanical device that is the object of control.

[0010] The data device of this disclosure, configured in this way, outputs control operating parameters to the controlled mechanical device by transforming the reference operating parameters output by the reference mechanical device control model. Therefore, the control device of this disclosure does not require relearning and constructing a new mechanical device control model to control the controlled mechanical device, which outputs operating parameters through machine learning of the operations performed on the controlled mechanical device. Thus, the data device of this disclosure can easily construct a mechanical device control model for controlling multiple controlled mechanical devices of different models without requiring the construction of a mechanical device control model for each of them.

[0011] like Figure 10 As shown, another aspect of this disclosure is a learning data generation apparatus for generating learning data for machine learning, which includes a collection and transformation unit and a generation unit.

[0012] The collection and transformation unit is configured to input at least one collection operation parameter, which is an operation parameter for operating the mechanical device for data collection, i.e., the collection target mechanical device, and use a physical model based on the mechanical characteristics of the collection target mechanical device, i.e., the mechanical device model, to transform the at least one collection operation parameter into at least one action parameter independent of each mechanical device, and output the transformed at least one action parameter.

[0013] The generation unit is configured to, based on the mechanical characteristics of a reference mechanical device, transform at least one action parameter output by the collection and transformation unit into at least one reference operating parameter, which is the same type of operating parameter as at least one collection operating parameter, thereby generating learning data.

[0014] The learning data generation apparatus of this disclosure, configured in this way, can use multiple collection object mechanical devices of different models to generate learning data for building a benchmark mechanical device control model by machine learning on the operation of the benchmark mechanical device, thus enabling easy construction of a mechanical device control model for controlling the mechanical device.

[0015] Another aspect of this disclosure is a data transformation program that enables a computer to function as both a reference transformation unit and a re-transformation unit, outputting control operation parameters, which are then transformed by the re-transformation unit, to the controlled mechanical device.

[0016] A computer controlled by the data transformation program of this disclosure can constitute part of the data device of this disclosure and can achieve the same effect as the data device of this disclosure.

[0017] Another aspect of this disclosure is a learning data generation program that enables a computer of a learning data generation device for generating learning data for machine learning to function as a collection and transformation unit and a generation unit.

[0018] A computer controlled by the learning data generation program of this disclosure can form part of the learning data generation apparatus of this disclosure and can achieve the same effect as the learning data generation apparatus of this disclosure.

[0019] Another aspect of this disclosure is a control system comprising a reference transformation unit and a re-transformation unit, which controls the controlled mechanical device based on control operation parameters that have been re-transformed by the re-transformation unit.

[0020] The control system disclosed herein is a system equipped with the data device of this disclosure and can achieve the same effect as the data device of this disclosure. Attached Figure Description

[0021] Figure 1 It is a block diagram showing the structure of a vehicle control system.

[0022] Figure 2 It is a functional block diagram representing the functional configuration of the vehicle control device.

[0023] Figure 3 This is a block diagram showing the structure of the compensator.

[0024] Figure 4 It is a graph showing the trajectory of a vehicle when it turns left at an intersection.

[0025] Figure 5 It is a graph showing the change in steering input over time when a vehicle turns left at an intersection.

[0026] Figure 6 It is a graph showing the change in yaw rate over time when a vehicle is making a left turn at an intersection.

[0027] Figure 7 This is a block diagram showing the structure of the learning data generation device.

[0028] Figure 8 This is a functional block diagram representing the functional structure of the learning data generation device.

[0029] Figure 9 This is a block diagram illustrating the configuration of the data apparatus of this disclosure.

[0030] Figure 10 This is a block diagram illustrating the configuration of the learning data generation apparatus of this disclosure.

[0031] Explanation of reference numerals in the attached figures 2…Vehicle control device 11…Data-driven planner 13…Basic Vehicle Model 14…compensator 100… Learning Data Generation Device 101…Collect vehicle models 102… compensator Detailed Implementation

[0032] [First Implementation Method] The following is related to the appendix. Figure 1 The first embodiment of this disclosure will be described below.

[0033] The vehicle control system 1 of this embodiment is installed in a vehicle capable of autonomous driving. Autonomous driving refers to automatically performing driving operations in place of the vehicle's occupants. The vehicle control system 1 is capable of, for example, Level 3 or higher autonomous driving. The levels of autonomous driving are defined by the Society of Automobile Manufacturers (SAAM).

[0034] In addition to having autonomous driving capabilities, vehicles equipped with vehicle control system 1 can also have manual driving capabilities. The vehicle can also be a hybrid vehicle with both an engine and an electric motor as its driving force. The vehicle is not limited to vehicles with autonomous driving capabilities or hybrid vehicles; it can also be a vehicle with only an engine or only an electric motor as its driving force. Hereinafter, the vehicle equipped with vehicle control system 1 will be referred to as the control vehicle.

[0035] like Figure 1 As shown, the vehicle control system 1 includes a vehicle control device 2 and an actuator 3.

[0036] Vehicle control device 2 is an electronic control device centered around a microcomputer including CPU 2a, ROM 2b, RAM 2c, and GPU 2d. The various functions of the microcomputer are implemented by the CPU 2a executing programs stored on a non-volatile physical recording medium. In this example, ROM 2b is equivalent to the non-volatile physical recording medium storing the program. Furthermore, by executing this program, methods corresponding to the program are executed. In addition, some or all of the functions executed by CPU 2a can be constructed in hardware using one or more ICs. Furthermore, the number of microcomputers constituting vehicle control device 2 can be one or more.

[0037] The vehicle control unit 2 takes as input sensor data generated by one or more sensors (not shown) that detect the state of the surrounding environment of the controlled vehicle and control the state of the vehicle, and outputs a target control quantity. Sensor data may include camera images from an onboard camera, vehicle operation quantities, etc. Alternatively, it may be sensor information used to understand the surrounding environment of the controlled vehicle and control the state of the vehicle, such as LiDAR or radar. Furthermore, examples of control quantities include acceleration / deceleration ~a(t) and front wheel steering angle ~δ(t), but it may also include acceleration, accelerator opening, brake opening, brake hydraulic pressure, steering wheel angle, etc.

[0038] Actuator 3 is based on acceleration and deceleration ～ a(t) and front wheel steering angle ～ δ(t) enables the accelerator, brakes, and steering mechanism of the vehicle to operate.

[0039] like Figure 2 As shown, the vehicle control device 2 includes a data-driven planner 11 and a model follower controller 12, which are functional blocks implemented by the CPU 2a executing a program stored in the ROM 2b.

[0040] The data-driven planner 11 has a learning model generated by machine learning using multiple vehicle surrounding image data of the base vehicle and multiple vehicle operation data (e.g., steering gear operation data, accelerator operation data, brake operation data) for operating the base vehicle. The base vehicle is a vehicle for which the aforementioned vehicle surrounding image data and vehicle operation data are collected for the learning of the data-driven planner 11.

[0041] This learning model, for example, takes image data captured by an onboard camera as input data and outputs vehicle operation data.

[0042] The model follower controller 12 has a basic vehicle model 13 and a compensator 14.

[0043] The methods for implementing these elements constituting the vehicle control device 2 are not limited to software; some or all of these elements may also be implemented using one or more hardware components. For example, when the aforementioned functions are implemented by electronic circuits as hardware, these electronic circuits may also be implemented by digital circuits containing a large number of logic circuits, or analog circuits, or a combination thereof.

[0044] The basic vehicle model 13 is a model that takes acceleration / deceleration a(t) and front wheel steering angle δ(t) as inputs and vehicle speed V(t) and yaw rate γ(t) as outputs. It is constructed from a vehicle model with the same specifications as the vehicle that acquired the learning data (i.e., the basic vehicle). The vehicle model in this embodiment is described, for example, by a vehicle mathematical model such as the dynamic two-wheel model shown in equations (1) to (4).

[0045] [Number 1] The definitions of the variables and parameters in equations (1) to (4) are shown in Table 1.

[0046] [Table 1] Here, the vehicle speed V(t) is updated by equation (5), and is used not only as the output of the basic vehicle model 13, but also in equations (1) to (4).

[0047] [Number 2] It can be assumed that the basic vehicle model 13 interprets the output of the data-driven planner 11 through an internally defined vehicle model. In this embodiment, the output of the planner, prompted by physical quantities of actuator actions, is interpreted as physical quantities of vehicle behavior such as vehicle speed and yaw rate.

[0048] The compensator 14 uses the vehicle speed V(t) and yaw rate γ(t) output from the base vehicle model 13 as target values, and determines acceleration and deceleration in a way that actually applies the data-driven planner 11 to the vehicle following the target values. ～ a(t) and front wheel steering angle ～ The target value of δ(t) follows the control system. Furthermore, the relationship between acceleration / deceleration and vehicle speed is independent of vehicle specifications; therefore, acceleration / deceleration... ～ a(t) becomes the same value as the output of the data-driven planner 11 (i.e., the acceleration / deceleration a(t)). Therefore, the front wheel steering angle will be described below. ～ Calculation of δ(t).

[0049] The target value following control system, like the basic vehicle model 13, is implemented through a control system based on the vehicle mathematical model.

[0050] For example, this can be achieved through adaptive control based on a dynamic two-wheel model.

[0051] The compensator 14 uses the nonlinear system shown in equations (6) and (7).

[0052] [Number 3] Here, x(t) is ∈R n ,u(t)∈R m y(t)∈R l G(x)∈R n H(x)∈R n×m , C∈R l×n G(x) and H(x) are smooth nonlinear functions of x(t).

[0053] like Figure 3 As shown, the compensator 14 includes a subtractor 21, a feedback linearization controller 22, an arithmetic unit 23, and a multiplier 24.

[0054] Subtractor 21 outputs the target value of y(t), i.e., y r The difference is obtained by subtracting y(t) from the output of multiplier 24.

[0055] The feedback linearization controller 22 calculates u(t) based on v(t) output by the subtractor 21 and outputs u(t).

[0056] The arithmetic unit 23 calculates x(t) based on u(t) output by the feedback linearization controller 22 according to equation (6) and outputs x(t).

[0057] The multiplier 24 multiplies the x(t) output by the arithmetic unit 23 by a preset constant C and outputs the resulting multiplier as y(t).

[0058] In this embodiment, the longitudinal motion of the vehicle is described by a point mass model, and the lateral motion and rotational motion around the center of gravity are described by a dynamic two-wheel model. Furthermore, x(t), u(t), y(t), G(x), H(x), and C in equations (6) and (7) are defined by equations (8), (9), (10), (11), (12), and (13), respectively. Here, F(t) represents the total driving force. Additionally, the vehicle parameters for each model use the specifications of the vehicle (i.e., the control vehicle) applied by the planner.

[0059] [Number 4] [Number 5] [Number 6] [Number 7] Calculate acceleration and deceleration using equation (14) ～ a(t).

[0060] [Number 8] Figure 4 It represents the result of a vehicle simulation, which shows the trajectory of the base vehicle and the control vehicle when turning left at an intersection.

[0061] Figure 4 Curve L1 represents the driving trajectory of the base vehicle. Curve L2 represents the driving trajectory of the controlled vehicle when the vehicle control device 2 has a data-driven planner 11 and a model follower controller 12. Curve L3 represents the driving trajectory of the controlled vehicle when the vehicle control device 2 has a data-driven planner 11 but no model follower controller 12.

[0062] like Figure 4 As shown, if only the data-driven planner 11 is applied to the control vehicle which is different from the base vehicle, the driving trajectory of the control vehicle deviates significantly from the driving trajectory of the base vehicle.

[0063] On the other hand, if a data-driven planner 11 and a model-following controller 12 are applied to the controlled vehicle, the driving trajectory of the controlled vehicle becomes a trajectory that is consistent with and similar to the driving trajectory of the base vehicle at the target position G.

[0064] Figure 5 It represents the results of vehicle simulations showing the time-varying steering amount of the base vehicle and the control vehicle when making a left turn at an intersection.

[0065] Figure 5 Line L11 represents the time variation of the steering amount of the base vehicle. Line L12 represents the time variation of the steering amount of the controlled vehicle when the vehicle control unit 2 has a data-driven planner 11 and a model-following controller 12. Line L13 represents the time variation of the steering amount of the controlled vehicle when the vehicle control unit 2 has a data-driven planner 11 but no model-following controller 12.

[0066] like Figure 5 As shown, when the data-driven planner 11 and the model follower controller 12 are applied to the controlled vehicle, the steering amount of the controlled vehicle is corrected relative to the steering amount of the base vehicle.

[0067] Figure 6It represents the results of vehicle simulations showing the time-varying yaw rate of the base vehicle and the control vehicle when making a left turn at an intersection.

[0068] Figure 6 Line L21 represents the time variation of the yaw rate of the base vehicle. Line L22 represents the time variation of the yaw rate of the controlled vehicle when the vehicle control unit 2 is equipped with the data-driven planner 11 and the model follower controller 12. Line L23 represents the target yaw rate output by the data-driven planner 11.

[0069] like Figure 6 As shown, since the steering input is corrected through model following control, the yaw rate of the controlled vehicle is approximately the same as the target yaw rate.

[0070] The vehicle control device 2 thus constructed includes a basic vehicle model 13 and a compensator 14.

[0071] The basic vehicle model 13 is configured such that the acceleration / deceleration a(t) and the front wheel steering angle δ(t) are output by the data-driven planner 11, which is constructed by machine learning through machine learning of the operation of the basic vehicle. The physical model based on the mechanical characteristics of the basic vehicle, namely the dynamic two-wheel model, is used to transform the acceleration / deceleration a(t) and the front wheel steering angle δ(t) into vehicle speed V(t) and yaw rate γ(t) independent of each vehicle, and outputs the transformed vehicle speed V(t) and yaw rate γ(t).

[0072] The compensator 14 is configured to, based on the mechanical characteristics of the controlled vehicle, transform the vehicle speed V(t) and yaw rate γ(t) output from the base vehicle model 13 into the same operating parameters as the acceleration / deceleration a(t) and the front wheel steering angle δ(t), namely acceleration / deceleration. ～ a(t) and front wheel steering angle ～ δ(t).

[0073] Vehicle control unit 2 will then change the acceleration and deceleration. ～ a(t) and front wheel steering angle ～ δ(t) is output to control the vehicle. That is, the vehicle control unit 2 outputs the re-evaluated acceleration and deceleration. ～ a(t) and front wheel steering angle ～ δ(t) controls the vehicle.

[0074] The base vehicle model 13 takes acceleration / deceleration a(t) and front wheel steering angle δ(t) as inputs and vehicle speed V(t) and yaw rate γ(t) as outputs, and is constructed from a vehicle model with the same specifications as the base vehicle.

[0075] The compensator 14 uses the action represented by the yaw rate γ(t) as the target value and determines the front wheel steering angle by controlling the vehicle to follow the target value. ～ The target value of δ(t) is followed by control.

[0076] Such a vehicle control device 2 will generate acceleration and deceleration by transforming the acceleration / deceleration a(t) and the front wheel steering angle δ(t) output by the data-driven planner 11. ～ a(t) and front wheel steering angle ～ δ(t) is output to the controlled vehicle. Therefore, the vehicle control unit 2 does not need to reconstruct a new data-driven planner 11 through relearning to control the vehicle; this data-driven planner 11 outputs acceleration and deceleration by performing machine learning on the operations performed on the controlled vehicle. ～ a(t) and front wheel steering angle ～ δ(t). Therefore, the vehicle control device 2 can easily construct a data-driven planner 11 for controlling the vehicles without having to construct a data-driven planner 11 for each of the multiple controlled vehicles with different models.

[0077] In the embodiments described above, the vehicle control device 2 is equivalent to a data device and a control system, the basic vehicle model 13 is equivalent to a reference transformation unit, the compensator 14 is equivalent to a re-transformation unit, and the data-driven planner 11 is equivalent to a reference mechanical device control model.

[0078] In addition, the base vehicle is equivalent to the reference mechanical device, the acceleration and deceleration a(t) and the front wheel steering angle δ(t) are equivalent to the reference operating parameters, the dynamic two-wheel model is equivalent to the mechanical device model, and the vehicle speed V(t) and yaw rate γ(t) are equivalent to the motion parameters.

[0079] In addition, controlling a vehicle is equivalent to controlling a mechanical device, such as acceleration and deceleration. ～ a(t) and front wheel steering angle ～ δ(t) is equivalent to the control operation parameter.

[0080] [Second Implementation] The following is related to the appendix. Figure 1 The second embodiment of this disclosure will now be described.

[0081] The learning data generation apparatus 100 in this embodiment is an apparatus for generating learning data for machine learning of the data-driven planner 11, such as... Figure 7 As shown, it is centered around a microcomputer equipped with CPU100a, ROM100b, RAM100c and GPU100d.

[0082] The various functions of the microcomputer are implemented by the CPU 100a executing programs stored in a non-volatile physical recording medium. In this example, the ROM 100b is equivalent to the non-volatile physical recording medium storing the program. Furthermore, by executing the program, the corresponding method is executed. In addition, some or all of the functions executed by the CPU 100a can be configured in hardware using one or more ICs. Furthermore, the number of microcomputers constituting the learning data generation device 100 can be one or more.

[0083] like Figure 8 As shown, by observing the driving of the data collection vehicle, multiple control data OD1 (e.g., steering gear operation data, accelerator operation data, brake operation data) are acquired for controlling the data collection vehicle.

[0084] The learning data generation device 100 includes a vehicle model 101 and a compensator 102 for the base vehicle, which are implemented as functional blocks by the CPU 100a executing a program stored in the ROM 100b.

[0085] The vehicle model 101, similar to the base vehicle model 13, is a dynamic two-wheeled model that takes acceleration / deceleration and front wheel steering angle as inputs and vehicle speed and yaw rate as outputs. It is constructed from a vehicle model with the same specifications as the vehicle that acquired the learning data (i.e., the data collection vehicle). Furthermore, the acceleration / deceleration and front wheel steering angle input to the vehicle model 101 are calculated using control data OD1.

[0086] The compensator 102 for the base vehicle has the same configuration as the compensator 14. It is a target value following control system that uses the output of the vehicle model 101, namely the vehicle speed V and the yaw rate γ, as target values ​​to make the base vehicle follow the target values ​​and determines the target value of acceleration, deceleration and front wheel steering angle.

[0087] That is, the compensator 102 for the base vehicle uses the nonlinear system shown in equations (6) and (7). Then, the longitudinal motion of the vehicle is described by a mass model, and the lateral motion and rotational motion around the center of gravity are described by a dynamic two-wheel model, and x(t), u(t), y(t) and G(x), H(x), C in equations (6) and (7) are defined by equations (8), (9), (10), (11), (12), and (13), respectively. Here, the vehicle parameters of each model use the specifications of the base vehicle.

[0088] The compensator 102 for the base vehicle outputs multiple control data OD2 based on the acceleration and deceleration input from the collected vehicle model 101 and the front wheel steering angle.

[0089] The learning data generation device 100 thus configured includes a vehicle model 101 for collecting data and a compensator 102 for the base vehicle.

[0090] The vehicle model 101 is configured to take as input multiple control data OD1 used to operate the data collection vehicle, and use a dynamic two-wheel model based on the mechanical characteristics of the data collection vehicle to transform the control data OD1 into vehicle speed and yaw rate independent of each vehicle, and output the transformed vehicle speed and yaw rate.

[0091] The compensator 102 for the base vehicle is configured to transform the vehicle speed and yaw rate output by the vehicle model 101 into multiple control data OD2, which are the same type of operational data as multiple control data OD1, based on the mechanical characteristics of the base vehicle, thereby generating learning data.

[0092] Such a learning data generation device 100 can use multiple data collection vehicles with different models to generate learning data for a data-driven planner 11 built by machine learning on the operation of the base vehicle, thus making it easy to build a data-driven planner 11 for controlling the vehicle.

[0093] In the embodiments described above, the vehicle model 101 is equivalent to the collection transformation unit, the compensator 102 facing the base vehicle is equivalent to the generation unit, the data collection vehicle is equivalent to the collection object mechanical device, the control data OD1 is equivalent to the collection operation parameters, and the control data OD2 is equivalent to the reference operation parameters.

[0094] The above describes one embodiment of the present disclosure, but the present disclosure is not limited to the above embodiment and can be implemented in various modifications.

[0095] [Variation Example 1] In the above embodiments, the mechanical device that becomes the controlled object is shown to be a vehicle, but it is not limited to vehicles. For example, it can also be a robot, an airplane, an artificial satellite, a ship, etc.

[0096] [Variation Example 2] In the above embodiment, a method in which the base vehicle model 13 and the compensator 14 are mounted on the control vehicle is shown. However, the base vehicle model 13 and the compensator 14 can also be mounted on a device located outside the control vehicle (e.g., a server capable of data communication with the vehicle).

[0097] The vehicle control device 2 and method described in this disclosure can also be implemented using a dedicated computer, which is provided by a processor and memory programmed to perform one or more functions embodied in a computer program. Alternatively, the vehicle control device 2 and method described in this disclosure can also be implemented using a dedicated computer provided by a processor composed of one or more dedicated hardware logic circuits. Alternatively, the vehicle control device 2 and method described in this disclosure can also be implemented using one or more dedicated computers composed of a processor and memory programmed to perform one or more functions and a processor composed of one or more hardware logic circuits. In addition, the computer program can also be stored as instructions executed by the computer on a computer-readable non-volatile tangible recording medium. In the method of implementing the functions of the various parts included in the vehicle control device 2, it is not necessary to include software, and all its functions can also be implemented using one or more hardware components.

[0098] Multiple functions of a single component in the above embodiments can be achieved through multiple components, or a single function of a single component can be achieved through multiple components. Alternatively, multiple functions of multiple components can be achieved through a single component, or a single function achieved by multiple components can be achieved through a single component. Furthermore, a portion of the configuration in the above embodiments can be omitted. Additionally, at least a portion of the configuration in the above embodiments can be added to or substituted relative to the configurations of other above embodiments.

[0099] In addition to the vehicle control device 2 described above, this disclosure can also be implemented in various ways, such as a system incorporating the vehicle control device 2, a program for enabling a computer to function as the vehicle control device 2, a non-volatile physical recording medium such as a semiconductor memory storing the program, and a control method.

[0100] In addition to the learning data generation apparatus 100 described above, this disclosure can also be implemented in various ways, such as a system incorporating the learning data generation apparatus 100, a program for enabling a computer to function as the learning data generation apparatus 100, a non-volatile physical recording medium such as a semiconductor memory storing the program, and a learning data generation method.

Claims

1. A data device, characterized in that, have: The reference transformation unit is configured to: input at least one reference operating parameter, which is an operating parameter output by a reference mechanical device control model constructed through machine learning for the operation of a reference mechanical device (i.e., a reference mechanical device) that serves as a reference; and use a physical model based on the mechanical characteristics of the reference mechanical device (i.e., a mechanical device model) to transform the at least one reference operating parameter into at least one action parameter independent of each mechanical device, and output the transformed at least one action parameter; and The re-transformation unit is configured to, based on the mechanical characteristics of the controlled mechanical device (i.e., the controlled mechanical device), transform at least one of the action parameters output by the reference transformation unit into at least one control operation parameter of the same type as the at least one reference operation parameter. The control operation parameters, after being transformed by the re-transformation unit, are output to the controlled mechanical device.

2. The data device according to claim 1, characterized in that, The reference transformation unit takes at least one of the reference operation parameters as input and at least one of the action parameters as output, and is constructed from the mechanical device model having the same specifications as the reference mechanical device.

3. The data device according to claim 1 or 2, characterized in that, The re-transformation unit takes the action represented by the action parameter as the target value and determines the target value of the control operation parameter for follow-up control in a manner that causes the controlled mechanical device to follow the target value.

4. A learning data generation apparatus for generating learning data for machine learning, characterized in that, have: The data collection and transformation unit is configured to: input at least one data collection operation parameter, which is an operation parameter for operating a mechanical device for data collection, i.e., the target mechanical device; and use a physical model based on the mechanical characteristics of the target mechanical device, i.e., a mechanical device model, to transform the at least one data collection operation parameter into at least one action parameter independent of each mechanical device, and output the transformed at least one action parameter; and The generation unit is configured to: based on the mechanical characteristics of a reference mechanical device, at least one of the action parameters output by the collection and transformation unit is transformed into at least one reference operation parameter, which is the same as the at least one collection operation parameter, thereby generating the learning data.

5. A data transformation program product, characterized in that, This data transformation program product enables the computer to function as both a reference transformation unit and a re-transformation unit. The reference transformation unit is configured to: input at least one reference operating parameter, which is an operating parameter output by a machine learning model (i.e., a reference mechanical device control model) constructed through machine learning for the operation of a reference mechanical device; and use a physical model (i.e., a mechanical device model) based on the mechanical characteristics of the reference mechanical device to transform the at least one reference operating parameter into at least one action parameter independent of each mechanical device, and output the transformed at least one action parameter. The re-transformation unit is configured to, based on the mechanical characteristics of the controlled mechanical device (i.e., the controlled mechanical device), transform at least one of the action parameters output by the reference transformation unit into at least one control operation parameter of the same type as the at least one reference operation parameter. The data transformation program product is used to output the control operation parameters, which have been transformed by the re-transformation unit, to the controlled mechanical device.

6. A learning data generation program product, characterized in that, This learning data generation program product enables the computer of the learning data generation device, which generates learning data for machine learning, to function as both a collection and transformation unit and a generation unit. The data collection and transformation unit is configured to: input at least one data collection operation parameter, which is an operation parameter for operating the mechanical device for data collection, i.e., the data collection target mechanical device; and use a physical model based on the mechanical characteristics of the data collection target mechanical device, i.e., a mechanical device model, to transform the at least one data collection operation parameter into at least one action parameter independent of each mechanical device, and output the transformed at least one action parameter. The generation unit is configured to: based on the mechanical characteristics of a reference mechanical device, at least one of the action parameters output by the collection and transformation unit is transformed into at least one reference operation parameter, which is the same as at least one collection operation parameter, thereby generating the learning data.

7. A control system, characterized in that, have: The reference transformation unit is configured to: input at least one reference operating parameter, which is an operating parameter output by a reference mechanical device control model constructed through machine learning for the operation of a reference mechanical device (i.e., a reference mechanical device) that serves as a reference; and use a physical model based on the mechanical characteristics of the reference mechanical device (i.e., a mechanical device model) to transform the at least one reference operating parameter into at least one action parameter independent of each mechanical device, and output the transformed at least one action parameter; and The re-transformation unit is configured to, based on the mechanical characteristics of the controlled mechanical device (i.e., the controlled mechanical device), transform at least one of the action parameters output by the reference transformation unit into at least one control operation parameter of the same type as the at least one reference operation parameter. The mechanical device to be controlled is controlled based on the control operation parameters after being transformed by the transformation unit.

Images