Machine learning device

Abstract

To provide a machine learning device that can easily predict vehicle behavior.SOLUTION: A machine learning device comprises one or a plurality of processors, and one or a plurality of memories that are connected to the processors. The processor executes a process which includes steps of inputting a first trend waveform which is a trend component of a first acceleration waveform of a vehicle to a first model and obtaining a second acceleration waveform in which an amplitude component of acceleration is added from the first model to the first trend waveform, in a predicted phase.SELECTED DRAWING: Figure 10

Description

[Technical Field]

[0001] The present invention relates to a machine learning device. [Background technology]

[0002] Conventionally, there is test equipment for testing vehicle engines. By testing an engine using test equipment, test data showing the engine's performance can be obtained. For example, there is a test equipment called an engine dynamometer (hereinafter simply referred to as EDM). An EDM generates a load on the engine and measures the power, thereby obtaining test data showing the engine's power. Specifically, a motor is directly connected to the engine's output shaft to absorb the power, and the engine torque is measured from the reaction force, thereby obtaining test data showing the engine's torque.

[0003] Patent Document 1 discloses a machine learning model (hereinafter referred to as an AI model) that inputs a command torque, which is the driving torque of a vehicle, and outputs a predicted acceleration, which is a predicted value of the vehicle's longitudinal acceleration. By having such an AI model learn using test data obtained by EDM, it is possible to predict vehicle behavior from parameters related to engine drive. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2022-185413 Summary of the Invention [Problem to be solved by the invention]

[0005] Incidentally, the shapes of the components that make up each engine may differ for each engine installed in a vehicle. When the shapes of the components differ for each engine, for example, even if the same engine speed is input, each engine will output a different torque. For this reason, parameters called ECU constants are set for each engine installed in a vehicle to adjust the torque of each engine. However, when predicting vehicle behavior using EDM and AI models, there is a problem in that the engine torque changes when the ECU constants, and therefore the engine state, change. Therefore, there is a problem in that it becomes difficult to predict vehicle behavior when the ECU constants, and therefore the engine state, change.

[0006] In view of the above problems, the present invention aims to provide a machine learning device that can easily predict vehicle behavior. [Means for solving the problem]

[0007] In order to solve the above problems, the machine learning device of the present invention comprises: one or more processors; one or more memories coupled to said processor; Equipped with The processor: In a prediction phase, a first trend waveform, which is a trend component of a first acceleration waveform of the vehicle, is input to a first model, and a second acceleration waveform is obtained from the first model by adding an acceleration amplitude component to the first trend waveform; Execute the process including. [Effects of the Invention]

[0008] According to the present invention, it is possible to easily predict vehicle behavior. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a schematic diagram showing the configuration of a machine learning system according to this embodiment. [Figure 2]FIG. 2 is a diagram for explaining an example of the learning phase in the first AI model according to this embodiment. [Figure 3] FIG. 3 is a diagram for explaining a first example of the prediction phase in the first AI model according to this embodiment. [Figure 4] FIG. 4 is a block diagram showing the configuration of the machine learning device according to this embodiment. [Figure 5] FIG. 5 is a block diagram showing an example of the functional configuration of the machine learning device according to this embodiment. [Figure 6] FIG. 6 is a diagram for explaining a second example of the prediction phase in the first AI model according to this embodiment. [Figure 7] FIG. 7 is a graph showing an example of output data when input data including a fluctuating engine speed is input to the first AI model. [Figure 8] FIG. 8 is a graph showing an example of output data when input data including a constant engine speed is input to the first AI model. [Figure 9] FIG. 9 is a diagram for explaining an example of the learning phase in the second AI model according to this embodiment. [Figure 10] FIG. 10 is a diagram for explaining an example of the prediction phase in the second AI model according to this embodiment. [Figure 11] FIG. 11 is a graph showing an example of output data when a trend waveform is input as input data to the second AI model. [Figure 12] FIG. 12 is a flowchart showing the flow of processing in the learning phase of the machine learning device according to this embodiment. [Figure 13] FIG. 13 is a flowchart showing the flow of processing in the prediction phase of the machine learning device according to this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Dimensions, materials, and other specific values ​​shown in the embodiments are merely examples for facilitating understanding of the invention and, unless otherwise specified, do not limit the present invention. In this specification and drawings, elements having substantially the same functions and configurations are designated by the same reference numerals to avoid redundant explanation, and elements not directly related to the present invention are not shown.

[0011] 1 is a schematic diagram showing the configuration of a machine learning system S according to this embodiment. The machine learning system S includes an engine 100, an EDM 200, and a machine learning device 300. The engine 100 is, for example, an engine mounted on a vehicle (not shown).

[0012] In this embodiment, an example will be described in which the vehicle is a four-wheeled private automobile. However, the vehicle is not limited to this example, and may be various types of automobiles, such as commercial vehicles such as buses, trucks, and taxis, or specialized vehicles such as police cars, fire engines, ambulances, tow trucks, snowplows, and construction vehicles, or motorcycles.

[0013] The EDM 200 generates a load on the engine 100 and measures the power, thereby obtaining test data indicating the power of the engine 100. For example, the EDM 200 can obtain test data indicating the torque of the engine 100 by inputting the engine speed. The EDM 200 transmits the obtained test data including various test results other than the engine speed to the machine learning device 300.

[0014] The machine learning device 300 outputs a predicted acceleration, which is a predicted value of the vehicle acceleration, based on the test data acquired from the EDM 200. The machine learning device 300 has two phases, broadly divided into a learning phase and a prediction phase.

[0015] The learning phase is a phase in which learning data is input to the machine learning device 300 and an AI model is constructed based on the learning data. Hereinafter, the AI ​​model constructed in the learning phase will be referred to as a trained model. Note that, as will be described in detail later, in this embodiment, multiple types of AI models are constructed. Specifically, the machine learning device 300 of this embodiment constructs a first AI model (second model) and a second AI model (first model). Details of the first AI model and the second AI model will be described later.

[0016] The learning data includes input data and correct answer data corresponding to the input data. The input data is, for example, engine data including test data. Here, the engine data is parameters related to the operation of the engine 100, and includes, for example, the engine speed, torque, and throttle opening of the engine 100. However, without being limited thereto, the engine data may also include the value of the air volume of the engine 100 and other values. Note that, for example, a numerical value obtained by testing the EDM 200 is used as the torque value of the engine 100.

[0017] FIG. 2 is a diagram illustrating an example of the learning phase of the first AI model according to this embodiment. As shown in (1) in FIG. 2, correct answer data (first correct answer data) corresponding to engine data is prepared in order to train the first AI model. As shown in (2) in FIG. 2, the learning data uses engine data, which is a parameter related to the operation of the engine 100, and acceleration data, which is correct answer data. Then, as shown in (3) in FIG. 2, the engine data and the acceleration data, which is correct answer data, are input to the first AI model, and learning is performed. In this way, a trained first AI model is constructed.

[0018] The correct answer data is acceleration data indicating acceleration values ​​obtained from an actual vehicle (hereinafter simply referred to as an actual vehicle) equipped with engine 100. The acceleration data can be obtained, for example, from an acceleration sensor mounted on the actual vehicle. Note that, hereinafter, engine data, which is a parameter related to the operation of engine 100 when the actual vehicle is run, and acceleration data obtained from the acceleration sensor when the actual vehicle is run are collectively referred to as actual vehicle data. In this embodiment, the correct answer data includes multiple types of acceleration data having multiple types of acceleration waveforms when the actual vehicle equipped with engine 100 is run multiple times.

[0019] Acceleration data, which is correct data, is linked to engine data, which is a parameter related to the operation of engine 100 when the actual vehicle is running. In this way, the acceleration data, which is correct data, is linked to the engine data and set as a set. In the learning phase, as shown in FIG. 2(3), the engine data and correct data are input to the first AI model as learning data, and the first AI model is trained, thereby constructing a trained first AI model.

[0020] The prediction phase is a phase in which input data not associated with correct answer data is input to a trained model, and output data indicating a result predicted by the trained model from the input data is output. In this embodiment, engine data not associated with correct answer data is input to a trained first AI model, and output data indicating a result of predicting vehicle acceleration as a vehicle behavior from the trained AI model. In this way, the input data input in the prediction phase is different from the training data input in the training phase.

[0021] 3 is a diagram for explaining a first example of the prediction phase in the first AI model according to this embodiment. As shown in Fig. 3, when engine data to which no correct answer data is linked is input to the trained first AI model, acceleration data including the vehicle acceleration waveform is output from the trained first AI model as output data.

[0022] Fig. 4 is a block diagram showing the configuration of a machine learning device 300 according to this embodiment. As shown in Fig. 4, the machine learning device 300 includes an I / F 310, a storage device 320, a system bus 330, one or more processors 340, and one or more memories 350. The I / F 310 is an interface for acquiring test data transmitted from the EDM 200, engine data that is a parameter related to the operation of the engine 100, and the like.

[0023] The storage device 320 is composed of RAM, flash memory, HDD, etc., and holds various information necessary for the processing of the processor 340 described below. For example, the storage device 320 holds training data including correct answer data used in the training phase. The system bus 330 electrically connects the I / F 310, storage device 320, processor 340, and memory 350, and is a transmission path for transmitting data among them.

[0024] The processor 340 includes, for example, a CPU (Central Processing Unit). The memory 350 includes, for example, a ROM (Read Only Memory) and a RAM (Random Access Memory). The ROM is a storage element that stores programs and calculation parameters used by the CPU. The RAM is a storage element that temporarily stores data such as variables and parameters used in processing executed by the CPU.

[0025] According to the machine learning device 300 of this embodiment, by inputting input data and correct answer data into the first AI model in the learning phase, machine learning is performed so that the vehicle acceleration predicted from the input data approaches the correct answer data, and a trained first AI model is constructed. As a result, by inputting engine data as input data into the trained first AI model in the prediction phase, the trained first AI model can output an appropriately predicted vehicle acceleration. As a result, it is possible to reproduce or predict the vehicle behavior of an actual vehicle equipped with engine 100 without running the actual vehicle.

[0026] Incidentally, the shapes of the components that make up each engine may differ for each engine installed in a vehicle. When the shapes of the components differ for each engine, for example, even if the same engine speed is input, each engine will output a different torque. For this reason, parameters called ECU constants are set for each engine installed in a vehicle to adjust the torque of each engine. However, when predicting vehicle behavior using EDM and AI models, there is a problem in that the engine torque changes when the ECU constants, and therefore the engine state, change. Therefore, there is a problem in that it becomes difficult to predict vehicle behavior when the ECU constants, and therefore the engine state, change.

[0027] To address this issue, the inventor discovered that while it is difficult to accurately determine vehicle behavior, which varies slightly depending on the engine, i.e., the vehicle acceleration waveform, it is possible to determine the trend in vehicle acceleration fluctuations.The inventor then devised a method for predicting ECU constants and, ultimately, vehicle behavior regardless of engine state by predicting the vehicle acceleration waveform using an AI model based on the trend in vehicle acceleration fluctuations.

[0028] In this embodiment, a first AI model shown in FIG. 2 is used to determine the fluctuation trend of the vehicle's acceleration. A second AI model, separate from the first AI model shown in FIG. 2, is constructed to predict the vehicle's acceleration waveform from the fluctuation trend of the vehicle's acceleration. The constructed second AI model is then used to predict the vehicle's acceleration waveform, i.e., vehicle behavior, from the fluctuation trend of the vehicle's acceleration. This makes it possible to predict the behavior of various vehicles equipped with various engines 100 with different ECU constants. A method for predicting a vehicle's acceleration waveform from the fluctuation trend of the vehicle's acceleration is described in detail below.

[0029] 5 is a block diagram showing an example of the functional configuration of the machine learning device 300 according to this embodiment. For example, as shown in FIG. 5, the machine learning device 300 includes a first supervised data generation unit 300a, a first learning unit 300b, a trend waveform acquisition unit 300c, a second supervised data generation unit 300d, a second learning unit 300e, and a prediction unit 300f.

[0030] The processor 340 cooperates with the programs contained in the memory 350 and executes the programs contained in the memory 350 to realize various processes including the processes described below that are performed by the first correct answer data generation unit 300a, the first learning unit 300b, the trend waveform acquisition unit 300c, the second correct answer data generation unit 300d, the second learning unit 300e, and the prediction unit 300f.

[0031] As shown by (1) in FIG. 2, the first correct answer data generation unit 300a generates first correct answer data necessary for training the first AI model. The first correct answer data generation unit 300a acquires multiple types of acceleration data having multiple types of acceleration waveforms when an actual vehicle equipped with engine 100 is driven multiple times, and sets this as correct answer data. The first correct answer data generation unit 300a links the acceleration data, which is correct answer data, to engine data, which is a parameter related to the operation of engine 100 when the actual vehicle is driven. In this way, the first correct answer data generation unit 300a generates first correct answer data that links the engine data and acceleration data, which are generated when the actual vehicle is driven, as a set.

[0032] As shown in (3) in FIG. 2, the first learning unit 300b constructs a first AI model (second model) based on the first supervised data generated by the first supervised data generating unit 300a. Specifically, the first learning unit 300b uses engine data, which is a parameter related to the operation of the engine 100, as input data, and inputs acceleration data linked to the engine data as first supervised data to the first AI model. By inputting the input data and the first supervised data to the first AI model, the first learning unit 300b performs machine learning so that the vehicle acceleration predicted from the input data approaches the first supervised data.

[0033] In this embodiment, the engine data, which is input data, is a value that varies over time, and the values ​​of the engine speed, torque, and throttle opening of engine 100 are values ​​that vary over time. Similarly, the correct answer data includes data on the acceleration value of the actual vehicle that varies over time, i.e., acceleration waveform (fourth acceleration waveform). This acceleration waveform is, for example, an acceleration waveform obtained from an acceleration sensor of the actual vehicle in which engine 100 is driven, when engine 100 is actually driven based on parameters related to the driving of engine 100. The first learning unit 300b trains the first AI model by inputting the engine data including such varying engine speed and the first correct answer data into the first AI model.

[0034] As explained using FIG. 3, simply inputting engine data into the first AI model trained by the first learning unit 300b outputs a predicted acceleration waveform of the vehicle. The inventors discovered that engine speed is one of the factors that contribute to the formation of minute amplitudes in the predicted acceleration waveform. Therefore, the inventors devised a method for generating a trend waveform, which is a trend component that excludes minute amplitude components from the predicted acceleration waveform, by setting the engine speed of the engine data input into the first AI model to a fixed value. This trend waveform represents the tendency of fluctuations in the vehicle's acceleration.

[0035] 6 is a diagram for explaining a second example of the prediction phase in the first AI model according to this embodiment. As shown in FIG. 6, when engine data in which the engine speed is set to a constant fixed value is input to the trained first AI model, a trend waveform (first trend waveform) that is the trend component of the vehicle acceleration waveform is output from the trained first AI model as output data.

[0036] As shown in Fig. 6, the trend waveform acquisition unit 300c inputs engine data in which the engine speed is set to a constant fixed value as input data to the trained first AI model trained by the first learning unit 300b. Note that, among the engine data input to the trained first AI model, parameters other than the engine speed are variable values. Specifically, the trained first AI model receives variable torque and throttle opening values ​​as input data.

[0037] As can be understood by comparing Figures 3 and 6, the output data output from the trained first AI model differs when input data including a fluctuating engine speed is input to the trained first AI model and when input data including a constant engine speed is input to the trained first AI model.

[0038] Figure 7 is a graph showing an example of output data when input data including a fluctuating engine speed is input to the first AI model. In Figure 7, the vertical axis represents acceleration and the horizontal axis represents time. As explained using Figure 3, when input data including a fluctuating engine speed is input to the trained first AI model, an acceleration waveform including minute amplitude components of acceleration, as shown in Figure 7, is output as output data. The acceleration waveform shown in Figure 7 is similar to the acceleration waveform of an actual vehicle equipped with engine 100, and roughly matches the acceleration waveform obtained from an acceleration sensor of the actual vehicle.

[0039] FIG. 8 is a graph showing an example of output data when input data including a constant engine speed is input to the first AI model. In FIG. 8, the vertical axis represents acceleration, and the horizontal axis represents time. As explained using FIG. 6, when input data including a constant engine speed is input to the trained first AI model, an acceleration waveform that does not include small amplitude components of acceleration, as shown in FIG. 8, is output as output data. In this way, when input data including a constant engine speed is input to the trained first AI model, a trend waveform that is the trend component of the acceleration waveform, as shown in FIG. 8, is obtained as output data, although it does not include small amplitude components of acceleration. Here, the trend waveform is a waveform that represents the tendency of acceleration to fluctuate over time.

[0040] In this way, the trend waveform acquisition unit 300c inputs engine data in which the engine speed is set to a constant fixed value as input data to the trained first AI model, thereby acquiring a first trend waveform as output data from the trained first AI model, as shown in Fig. 8. In other words, the trend waveform acquisition unit 300c inputs engine data in which the engine speed is constant to the trained first AI model, thereby acquiring a first trend waveform, which is the trend component of the acceleration waveform (first acceleration waveform) of the vehicle equipped with engine 100.

[0041] In this way, the trend waveform acquisition unit 300c can predict the trend waveform of the vehicle acceleration, but the trend waveform alone cannot predict minute amplitudes of the vehicle acceleration waveform, i.e., minute vehicle behavior. Therefore, in order to predict an acceleration waveform containing minute amplitude components of the vehicle from the trend waveform of the vehicle acceleration, the machine learning device 300 constructs a new second AI model.

[0042] FIG. 9 is a diagram illustrating an example of the learning phase of the second AI model according to this embodiment. As shown in (1) in FIG. 9, correct answer data (second correct answer data) is prepared to train the second AI model. Here, the second correct answer data includes, for example, acceleration data including an acceleration waveform obtained from an acceleration sensor of a real vehicle. Once the acceleration waveform, which is the acceleration data of the real vehicle, is obtained, a process is performed to remove high-frequency components from the acceleration waveform of the real vehicle, thereby obtaining a second trend waveform, which is the trend component of the acceleration waveform of the real vehicle, as shown in (2) in FIG. 9. As shown in (3) in FIG. 9, the second trend waveform obtained in (2) in FIG. 9 is linked to the acceleration waveform serving as correct answer data on which the trend waveform is based, and this set becomes the second correct answer data of the second AI model. Then, as shown in (4) in FIG. 9, the second trend waveform, which is input data, and the acceleration waveform, which is the second correct answer data, are input to the second AI model as learning data, and learning is performed. This results in the construction of a trained second AI model.

[0043] Specifically, as shown in (1) in Fig. 9, the second correct answer data generation unit 300d generates second correct answer data necessary for training the second AI model. The second correct answer data generation unit 300d acquires multiple types of acceleration data having multiple types of acceleration waveforms (third acceleration waveforms) when an actual vehicle equipped with engine 100 is driven multiple times, and sets this as correct answer data. This correct answer data may be the same data as the correct answer data used by the first correct answer data generation unit 300a, or may be different data.

[0044] As shown in (2) in Fig. 9, the second correct data generating unit 300d derives a second trend waveform, which is a trend component of the acceleration waveform of correct data, based on the acceleration waveform of correct data (third acceleration waveform). Specifically, the second correct data generating unit 300d performs a removal process to remove high-frequency components contained in the acceleration waveform of correct data, and derives a second trend waveform from which the high-frequency components have been removed from the acceleration waveform of correct data. By performing this removal process to remove high-frequency components, a trend waveform such as that shown in Fig. 8 can be obtained from the acceleration waveform shown in Fig. 7, for example.

[0045] 9(3), the second correct data generating unit 300d links the derived second trend waveform with the acceleration data, which is correct data. In this way, the second correct data generating unit 300d generates second correct data in which the second trend waveform and the acceleration data are linked as a set.

[0046] As shown by (4) in Figure 9, the second learning unit 300e constructs a second AI model based on the second trend waveform and second supervised data generated by the second supervised data generation unit 300d. Specifically, the second learning unit 300e uses the second trend waveform as input data and inputs the acceleration data linked to the second trend waveform as second supervised data to the second AI model. By inputting the second trend waveform and the second supervised data to the second AI model as learning data, the second learning unit 300e performs machine learning so that the vehicle acceleration waveform predicted from the second trend waveform approaches the second supervised data.

[0047] In this way, the second learning unit 300e uses the second trend waveform, which is the correct data, from which high-frequency components have been removed, as input data, and trains the second AI model by inputting the input data and the correct data into the second AI model.

[0048] 10 is a diagram illustrating an example of the prediction phase in the second AI model according to this embodiment. As shown in FIG. 10, when the first trend waveform is input to the trained second AI model, the trained second AI model outputs the vehicle acceleration waveform as output data.

[0049] The prediction unit 300f inputs the first trend waveform acquired by the trend waveform acquisition unit 300c as input data to the trained second AI model trained by the second training unit 300e.

[0050] FIG. 11 is a graph showing an example of output data when a trend waveform is input as input data to the second AI model. In FIG. 11, the vertical axis represents acceleration, and the horizontal axis represents time. As shown in FIG. 11, when the first trend waveform shown in FIG. 8 is input as input data to the trained second AI model, an acceleration waveform containing minute amplitude components of acceleration is output as output data. This acceleration waveform as output data is, for example, a waveform in which minute amplitude components of acceleration are added to the trend waveform shown in FIG. 8.

[0051] In this way, the prediction unit 300f inputs the first trend waveform into the trained second AI model and acquires from the trained second AI model an acceleration waveform (second acceleration waveform) in which the amplitude component of acceleration is added to the first trend waveform. In other words, the prediction unit 300f inputs the first trend waveform acquired by the trend waveform acquisition unit 300c as input data into the trained model, thereby executing a prediction phase in which the trained model outputs an acceleration of the vehicle that is appropriately predicted. This makes it possible to easily predict the behavior of various vehicles equipped with various engines 100, regardless of engine speed.

[0052] Furthermore, as described in Figure 6, the prediction unit 300f can input engine data including fluctuating engine speeds into the trained first AI model, thereby obtaining acceleration data including an acceleration waveform such as that shown in Figure 7 from the trained first AI model.

[0053] 12 is a flowchart showing the flow of processing in the learning phase of the machine learning device 300 according to this embodiment. As shown in FIG. 12, the first supervised data generation unit 300a and the second supervised data generation unit 300d determine whether or not actual vehicle data including engine data and acceleration data obtained when a real vehicle is driven has been acquired (step S100). If actual vehicle data has not been acquired (NO in step S100), the first supervised data generation unit 300a and the second supervised data generation unit 300d end the processing in the learning phase.

[0054] If the actual vehicle data is acquired (YES in step S100), the first correct answer data generation unit 300a generates the first correct answer data shown in (2) in FIG. 2 based on the engine data and acceleration data included in the actual vehicle data (step S110).

[0055] Once the first correct answer data is generated, the first learning unit 300b trains the first AI model by inputting the engine data, which is the input data, and the acceleration data, which is the correct answer data, into the first AI model, as shown in (3) in Figure 2 (step S120).

[0056] Furthermore, when actual vehicle data is acquired (YES in step S100), the second supervised data generating unit 300d performs a removal process to remove high-frequency components contained in the vehicle acceleration waveform, which is the supervised data, as shown in (2) in FIG. 9. This derives a second trend waveform from which the high-frequency components have been removed from the acceleration waveform, which is the supervised data. Furthermore, the second supervised data generating unit 300d generates second supervised data that links the derived second trend waveform and the acceleration waveform, which is the supervised data, as a set, as shown in (3) in FIG. 9 (step S130).

[0057] When the second supervised data is generated, the second learning unit 300e uses the second trend waveform, which is the supervised data and from which high-frequency components have been removed, as input data, as shown in (4) in Fig. 9, and inputs the input data and the second supervised data to the second AI model, thereby training the second AI model (step S140).

[0058] Then, the first learning unit 300b and the second learning unit 300e determine whether learning of both the first AI model and the second AI model has been completed (step S150). If learning of both the first AI model and the second AI model has not been completed (NO in step S150), the first learning unit 300b and the second learning unit 300e repeat the processes from step S110 to step S140. On the other hand, if learning of both the first AI model and the second AI model has been completed (YES in step S150), the first learning unit 300b and the second learning unit 300e end the learning phase process.

[0059] 13 is a flowchart showing the flow of processing in the prediction phase of the machine learning device 300 according to this embodiment. As shown in FIG. 13, the prediction unit 300f acquires engine data via the I / F 310 (step S200).

[0060] The prediction unit 300f determines whether the value of the engine speed included in the acquired engine data is a variable value or a fixed value (step S210).

[0061] If the engine speed is a variable value (YES in step S210), the prediction unit 300f inputs engine data including the variable engine speed into the trained first AI model, as shown in FIG. 3, and predicts the vehicle acceleration waveform (step S220).

[0062] If the engine speed is not a variable value (NO in step S210), the trend waveform acquisition unit 300c inputs engine data including the engine speed, which is a constant fixed value, into the trained first AI model, as shown in Figure 6, and acquires the first trend waveform, which is the trend component of the vehicle acceleration waveform (step S230).

[0063] After acquiring the first trend waveform, the prediction unit 300f inputs the first trend waveform into the trained second AI model, as shown in Figure 10, and acquires an acceleration waveform from the trained second AI model in which the amplitude component of acceleration has been added to the first trend waveform (step S240).

[0064] As described above, according to this embodiment, the machine learning device 300 constructs a second AI model (first model) that predicts an acceleration waveform from a trend waveform, separate from a first AI model (second model) that predicts an acceleration waveform from engine data. Then, in the prediction phase, the prediction unit 300f inputs a first trend waveform, which is the trend component of the vehicle's first acceleration waveform, into the second AI model, and acquires a second acceleration waveform from the second AI model, in which an acceleration amplitude component is added to the first trend waveform. In this way, by predicting the vehicle's acceleration waveform from a trend waveform that indicates the tendency of fluctuations in the vehicle's acceleration, it is possible to easily predict the behavior of various vehicles equipped with various engines 100 having different ECU constants.

[0065] In addition, in the learning phase, the second learning unit 300e uses the third acceleration waveform acquired from the acceleration sensor mounted on the vehicle as ground truth data and the second trend waveform obtained by removing high-frequency components from the third acceleration waveform as input to train the second AI model. This makes it easy to build the second AI model (first model) for predicting the vehicle's acceleration waveform from the trend waveform that indicates the fluctuation tendency of the vehicle's acceleration, and to generate second ground truth data for training the second AI model.

[0066] In addition, in the learning phase, the first learning unit 300b inputs engine data and trains the first AI model by using a fourth acceleration waveform acquired from an acceleration sensor mounted on the vehicle when the engine is driven by the engine data as correct data. In the prediction phase, the first AI model is input with engine data at a constant engine speed, and a first trend waveform is acquired from the first AI model. This makes it possible to easily generate a first trend waveform, which is a trend component of the vehicle's acceleration waveform, using the first AI model.

[0067] While the preferred embodiments of the present invention have been described above with reference to the accompanying drawings, it goes without saying that the present invention is not limited to such embodiments. It is clear that those skilled in the art can conceive of various modifications and alterations within the scope of the claims, and it is understood that such modifications and alterations also fall within the technical scope of the present invention.

[0068] The series of processes performed by the machine learning device 300 according to this embodiment described above may be implemented using software, hardware, or a combination of software and hardware. The programs constituting the software are stored in advance in, for example, a non-transitory storage medium provided inside or outside each device. The programs are then read from, for example, a non-transitory storage medium (for example, a ROM) to a transitory storage medium (for example, a RAM) and executed by a processor such as a CPU.

[0069] A program for realizing each function of each of the above devices can be created and installed on the computer of each of the above devices. The processor executes the program stored in memory, thereby performing the processing of each of the above functions. At this time, the program may be shared and executed by multiple processors, or the program may be executed by a single processor. Furthermore, the functions of each of the above devices may be realized by cloud computing using multiple computers interconnected by a communications network. The program may be provided to and installed on the computer of each device by distribution from an external device via a communications network. [Explanation of symbols]

[0070] S Machine Learning System 100 Engine 200 Engine Dynamometer (EDM) 300 Machine Learning Device 300a First correct answer data generation unit 300b First Study Section 300c Trend waveform acquisition section 300d Second correct answer data generation unit 300e Second Learning Section 300f Prediction Department 310 Interface 320 storage device 330 System Bus 340 processor 350 memory

Claims

1. one or more processors; one or more memories coupled to the processor; Equipped with The processor: In the prediction phase, a first trend waveform, which is a trend component of a first acceleration waveform of the vehicle, is input to a first model, and a second acceleration waveform is obtained from the first model by adding an acceleration amplitude component to the first trend waveform; Performing a process including Machine learning device.

2. The processor: In a learning phase, a third acceleration waveform acquired from an acceleration sensor mounted on the vehicle is used as correct data, and a second trend waveform obtained by removing high frequency components from the third acceleration waveform is used as an input to learn the first model. Performing a process including The machine learning device according to claim 1 .

3. The processor: in a learning phase, parameters related to engine operation are input, and when the engine is operated by the parameters, a fourth acceleration waveform acquired from an acceleration sensor mounted on the vehicle is used as correct data, thereby learning a second model; In a prediction phase, the parameters with the engine speed kept constant are input to the second model, and the first trend waveform is obtained from the second model; Performing a process including The machine learning device according to claim 1 .

4. the parameters used in training the second model include varying engine speed; The machine learning device according to claim 3 .

5. the first trend waveform is a waveform that represents a fluctuation trend of the acceleration over time; The machine learning device according to any one of claims 1 to 4.

Images