Training data creation device, training data creation method, and program

Abstract

The present application relates to a training data creation device, a training data creation method, and a program. The training data creation device includes an acquisition unit that acquires a plurality of sensor values from a plurality of sensors; a calculation unit that calculates a calculation value corresponding to a sensor value other than the used sensor value using a sensor value included in the plurality of sensor values; a determination unit that determines whether each of the plurality of sensors is normal; and a training data creation unit that creates training data from an input / output data set including at least one calculation value corresponding to a sensor value of a sensor determined to be abnormal in a case where at least one sensor is determined to be abnormal by the determination unit.

Description

TECHNICAL FIELD

[0001] The present application relates to a training data creation device, a training data creation method, and a program. BACKGROUND

[0002] In Japanese Patent Application Publication No. 2018-229, a training data creation device is disclosed. In the training data creation device, a user's posture and motion state are recognized based on acceleration information acquired from two or more sensors attached to portions that generate qualitatively or quantitatively different accelerations depending on the user's posture and motion state, and training data is created by associating the recognized posture and motion state with acceleration information acquired from at least one of the two or more sensors.

[0003] In the disclosed training data creation device, in a case where there is a sensor that has failed or whose sensing accuracy has decreased abnormally, it is feared that training data cannot be created. SUMMARY

[0004] The present application was achieved in view of the above-described circumstances, and aims to provide a training data creation device, a training data creation method, and a program that can create training data even in a case where there is an abnormal sensor.

[0005] One aspect of the present application is a training data creation device including: an acquisition unit that acquires a plurality of sensor values from a plurality of sensors; a calculation unit that calculates a calculation value corresponding to a sensor value other than the used sensor value using a sensor value included in the plurality of sensor values; a determination unit that determines whether each of the plurality of sensors is normal; and a training data creation unit that creates training data from an input / output data set including the calculation value corresponding to the sensor value of the sensor determined to be abnormal, in a case where at least one sensor is determined to be abnormal by the determination unit.

[0006] Thus, training data is created from an input / output data set including a calculation value corresponding to a sensor value of a sensor determined to be abnormal, so training data can be created even in a case where there is an abnormal sensor.

[0007] It can also be that the determination unit determines the sensor to be abnormal in a case where the sensor has failed or its sensing accuracy has decreased by a predetermined degree or more.

[0008] Thus, both a case where a sensor has failed and a case where the accuracy of a sensor has decreased can be dealt with as sensor abnormalities.

[0009] It can also be that the determination unit compares the sensor value and the calculation value to determine the sensing accuracy of the sensor.

[0010] Thus, the sensing accuracy of the sensor can be determined by comparing the sensor value with the operation value, and thus the accuracy reduction of the sensor can be more appropriately determined.

[0011] The training data production section can produce the training data from an input / output data set including the operation value corresponding to the sensor value of the sensor determined to be normal by the determination section.

[0012] Thus, the training data of high quality can be produced using the operation value considered to have a small deviation from the sensor value obtained when the sensor was determined to be normal in the past.

[0013] The training data production section can include the sensor value of the sensor determined to be normal by the determination section in the input / output data set.

[0014] Thus, the training data of high quality can be produced using the sensor value, which is valuable raw data, effectively.

[0015] One aspect of the present application is a training data production method including: acquiring a plurality of sensor values from a plurality of sensors, operating an operation value corresponding to a sensor value other than the used sensor value using a sensor value included in the plurality of sensor values, storing the operation value in a storage section, determining whether each of the plurality of sensors is normal, and producing training data from an input / output data set including the operation value corresponding to the sensor value of the sensor determined to be abnormal read from the storage section in a case where at least one sensor is determined to be abnormal.

[0016] Thus, the training data is produced from the input / output data set including the operation value corresponding to the sensor value of the sensor determined to be abnormal, and thus the training data can be produced even in a case where there is an abnormal sensor.

[0017] One aspect of the present application is a program causing a processor having hardware to execute: acquiring a plurality of sensor values from a plurality of sensors, operating an operation value corresponding to a sensor value other than the used sensor value using a sensor value included in the plurality of sensor values, storing the operation value in a storage section, determining whether each of the plurality of sensors is normal, and producing training data from an input / output data set including the operation value corresponding to the sensor value of the sensor determined to be abnormal read from the storage section in a case where at least one sensor is determined to be abnormal.

[0018] Thus, the training data is created from the input / output data set including the operation value corresponding to the sensor value of the sensor determined to be abnormal, and therefore the training data can be created even when there is an abnormal sensor.

[0019] According to the present application, the training data can be created even when there is an abnormal sensor. BRIEF DESCRIPTION OF DRAWINGS

[0020] Hereinafter, features, advantages, and technical and industrial significance of exemplary embodiments of the present application will be described with reference to the accompanying drawings, in which like numerals refer to like elements, and wherein:

[0021] Figure 1 is a schematic view showing the configuration of an apparatus that is a training data creation device according to the embodiment.

[0022] Figure 2 is a view explaining an outline of creation of training data.

[0023] Figure 3 is a flowchart showing a process of determining whether a sensor is normal.

[0024] Figure 4 is a flowchart showing a process of determining the accuracy of a sensor.

[0025] Figure 5 is a flowchart showing a process of creating training data.

[0026] Figure 6 is a flowchart showing a process of executing machine learning.

[0027] Figure 7 is a block diagram explaining control of cruise control in a vehicle.

[0028] Figure 8 is a block diagram of an F / F control section.

[0029] Figure 9 is a view explaining a sensor value and an operation value. DETAILED DESCRIPTION

[0030] Hereinafter, the embodiment of the present application will be specifically described with reference to the accompanying drawings. Note that, in the drawings, the same reference numerals are appropriately assigned to the same or corresponding constituent elements, and repetitive description will be omitted.

[0031] <SYSTEM CONFIGURATION>

[0032] Figure 1is a schematic diagram showing the configuration of an apparatus that is a training data creation device according to an embodiment. The apparatus 100 includes a sensor group 10, a training data creation device 20, and a machine learning device 30. The sensor group 10, the training data creation device 20, and the machine learning device 30 are connected to be able to communicate with each other through a communication network such as a CAN (Controller Area Network). The apparatus 100 is not particularly limited and is a robot apparatus, a conveyance apparatus, a wearable apparatus disclosed in Patent Literature 1, or the like.

[0033] <SENSOR GROUP>

[0034] The sensor group 10 is configured by a plurality of sensors that detect the state and characteristics of the apparatus 100. Each sensor of the sensor group 10 transmits a sensor value as a detection result to the training data creation device 20 through a network such as a CAN.

[0035] <TRAINING DATA CREATION DEVICE>

[0036] The training data creation device 20 includes a control section 21, a storage section 22, and a communication section 23.

[0037] The control section 21 includes, for example, a processor such as a CPU (Central Processing Unit), a DSP (Digital Signal Processor), an FPGA (Field-Programmable Gate Array), and a main storage section such as a RAM (Random Access Memory) and a ROM (Read Only Memory). The control section 21 reads out a program stored in the storage section 22 to a work area of the main storage section and executes it, controls each configuration section and the like by execution of the program realized by the processor, and thereby realizes a function module that meets a prescribed purpose in cooperation with hardware and software.

[0038] The control section 21 has an acquisition section 21a, a calculation section 21b, a determination section 21c, and a training data creation section 21d as functional modules. The acquisition section 21a acquires a plurality of sensor values from the sensor group 10. The calculation section 21b calculates a calculation value corresponding to a sensor value other than the used sensor value using the sensor value included in the plurality of sensor values, and stores the calculation value in the storage section 22. The determination section 21c determines whether each of the sensors of the sensor group 10 is normal. Further, the determination section 21c determines the sensing accuracy of the sensor by comparing the sensor value with the calculation value. The training data creation section 21d creates training data from an input / output data set including the sensor value or the calculation value. The control section 21 stores the created training data in the storage section 22. Further, the control section 21 reads out the training data stored in the storage section 22, and transmits the training data to the communication section 23.

[0039] The storage section 22 is constituted by a RAM, a hard disk drive (HDD), a removable medium, or the like, and is also called an auxiliary storage section. Note that the removable medium is, for example, a USB (Universal Serial Bus) memory or a disk recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), and a BD (Blue-ray (registered trademark) Disc). Further, the storage section 22 can be constituted using a computer-readable recording medium such as a storage card that can be attached from the outside. An operating system (OS), various programs, various tables, various databases, and the like for realizing the functions of the training data creation device 20 are stored in advance in the storage section 22, or are stored in the storage section 22 by being downloaded via a communication network.

[0040] The communication section 23 is configured to include, for example, a communication interface such as a CAN interface board, and communicates with the sensor group 10 and the machine learning device 30. The communication section 23 receives data of the sensor values from the sensor group 10, and transmits the training data to the machine learning device 30.

[0041] <Machine Learning Device>

[0042] The machine learning device 30 has a control section 31, a storage section 32, and a communication section 33.

[0043] The control section 31 has a processor and a main storage section similarly to the control section 21 of the training data creation device 20. The control section 31 reads out a program stored in the storage section 32 to a work area of the main storage section and executes, controls each of the constituent sections and the like by the execution of the program, and thereby realizes functional modules in accordance with the prescribed purpose.

[0044] The control section 31 has an acquisition section 31a and a learning section 31b as functional modules. The acquisition section 31a reads and acquires the training data transmitted from the training data production device 20 from the storage section 32. The learning section 31b performs machine learning using the training data, and generates a learned model.

[0045] The storage section 32 is configured by the same storage medium as the storage section 22 of the training data production device 20. In the storage section 32, an OS, various programs, various tables, various databases, and the like for realizing the functions of the machine learning device 30 can be stored. Further, the storage section 32 stores the training data transmitted from the training data production device 20, and the learned model generated by the learning section 31b.

[0046] The learned model is, for example, a model produced by machine learning using deep learning of a forward propagation type neural network. In this case, the storage section 32 storing the learned model means storing information of network parameters, algorithms of operations, and the like in the learned model. Further, the following transmission, reception, or readout, and the like of the learned model also mean transmission, reception, or readout, and the like of the information of network parameters, algorithms of operations, and the like.

[0047] The communication section 33 is configured to include a communication interface like the communication section 23 of the training data production device 20, and communicates with the training data production device 20. For example, the communication section 33 receives the training data transmitted from the training data production device 20.

[0048] <Function of Training Data Production Device>

[0049] Next, the functions of the training data production device 20 are described more specifically. In the control section 21 of the training data production device 20, the operation section 21b operates an operation value equivalent to a sensor value other than the input sensor value using the sensor value included in the plurality of sensor values as the input sensor value. The operation section 21b, for example, reads a program for executing an operation model stored in the storage section 22 to perform an operation of taking at least one input sensor value as an input value and at least one operation value as an output value. The operation model is optimized so that the operation value is within a prescribed error range with respect to the corresponding sensor value. Such an operation model can be a learned model obtained by machine learning or the like, or a physical model constructed using an operation formula (equation of motion or the like) representing a physical law.

[0050] The determination unit 21c determines whether each sensor in the sensor group 10 is functioning normally. For example, the determination unit 21c reads data stored in the storage unit 22 indicating the acceptable range of sensor values ​​for each sensor. If the acquired sensor value falls outside this range, the determination unit 21c determines that the sensor has malfunctioned. Furthermore, the determination unit 21c compares the sensor value of a sensor determined to be functioning normally with a calculated value corresponding to the sensor value to determine the sensing accuracy of the sensor. For example, if the difference between the sensor value and the calculated value is greater than a specified value, the determination unit 21c determines that the sensing accuracy of the sensor has decreased by more than a specified degree. The specified value is an indicator used to determine the sensing accuracy of the sensor and is stored, for example, in the storage unit 22. Ultimately, the determination unit 21c determines that a sensor that has malfunctioned or whose sensing accuracy has decreased by more than a specified degree is abnormal.

[0051] When the determination unit 21c determines that at least one sensor is abnormal, the training data generating unit 21d generates training data based on an input / output data set including a calculation value corresponding to the sensor value of the sensor determined to be abnormal.

[0052] Figure 2 The diagram is used to explain the outline of the training data generation unit 21d. First, when all sensors are judged to be normal, the training data generation unit 21d generates an input / output data set (X1, ..., X2) based on the sensor values. n ∶Y1,……,Y m ). n, m are integers greater than 1, and the number of data is (n+m). Figure 2 In the example shown in FIG1 when the sensor is normal, the data included in the input and output data sets are all sensor values ​​as raw data. n is the input data set. Y1, ..., Y m The training data generating unit 21d appropriately performs known pre-processing such as deletion or completion of missing data, normalization or standardization of data on the input / output data set to generate training data.

[0053] Next, when the training data generating unit 21d determines that at least one sensor is abnormal, it generates an input / output data set (X1, ..., X2) based on the sensor value and the calculation value. n ∶y1,……,y m ).exist Figure 2 In the example shown, y1, ..., y m is the calculated value. m are the output data sets, which are the sensor values ​​Y1, ..., Y m Equivalent operation value. Figure 2In the example shown as a case of sensor abnormality, when at least one sensor detecting sensor values ​​that should constitute the output dataset is determined to be abnormal, all sensor values ​​that should constitute the output dataset are replaced with calculated values. The training data creation unit 21d appropriately performs known preprocessing on the input and output datasets, such as deletion or completion of missing data and data normalization or standardization, to create training data.

[0054] In the training data creation device 20 configured as described above, the acquisition unit 21a acquires multiple sensor values ​​from multiple sensors in the sensor group 10. The calculation unit 21b then uses sensor values ​​included in the multiple sensor values ​​to calculate a calculated value corresponding to sensor values ​​other than the used sensor values ​​and stores the calculated value in the storage unit 22. The determination unit 21c determines whether each of the multiple sensors is normal. If the determination unit 21c determines that at least one sensor is abnormal, the training data creation unit 21d reads a calculated value corresponding to the sensor value of the sensor determined to be abnormal from the storage unit 22 and creates training data based on the input / output data set containing the read calculated value.

[0055] Thus, even when there is an abnormal sensor, the training data creation device 20 can create training data based on the input-output data set including the calculation value corresponding to the sensor value.

[0056] Furthermore, the determination unit 21c determines that a sensor is abnormal when a sensor fails or the sensing accuracy has decreased by a predetermined level or more. This can therefore cope with both sensor anomalies, namely, sensor failure and sensor accuracy decrease.

[0057] Furthermore, since the determination unit 21 c determines the sensing accuracy of the sensor by comparing the sensor value with the calculated value, the sensing accuracy of the sensor can be determined based on the calculated value.

[0058] It should be noted that in Figure 2 In the example shown as a sensor abnormality, when it is determined that at least one sensor that detects the sensor values ​​that should constitute the output data set is abnormal, all the sensor values ​​that should constitute the output data set are replaced with calculated values, but the replacement of sensor values ​​with calculated values ​​is not limited to this.

[0059] For example, the training data generator 21d may include the sensor values ​​of the sensors determined to be normal by the determination unit 21c in the input / output data set. This allows the sensor values, which are valuable raw data, to be effectively utilized to generate high-quality training data, thereby enabling the machine learning device 30 to quickly improve learning accuracy.

[0060] Further, the storage section 22 can store the determination result obtained by the determination section 21c, and the training data production section 21d can produce the operation value to be included in the input / output data set as the operation value corresponding to the sensor value of the sensor that is normal in the past determination result stored in the storage section 22. Thus, the training data of high quality can be produced using the operation value that is considered to have a small deviation from the sensor value obtained when the sensor that was determined to be normal in the past is normal.

[0061] Further, the conversion of the sensor value into the operation value is not limited to the data that should constitute the output data set, but can be performed on the data that should constitute the input data set. Thus, the training data can be produced even when the abnormal sensor acquires the sensor value of either the input data or the output data.

[0062] <PROCESS FLOW>

[0063] Various processes performed in the training data production device 20 or the machine learning device 30 will be described. In the flowchart below, the sensors of the sensor group 10 and the corresponding sensor values, operation values, and flags are added with a suffix i so as to correspond to the count i.

[0064] Figure 3 is a flowchart showing a process of determining whether the sensor is normal.

[0065] First, in step S101, the determination section 21c initializes the count i to i = 0.

[0066] Next, in step S102, the determination section 21c determines whether the sensor (i) is normal. In the case where it is determined to be normal (YES in step S102), the determination section 21c sets the flag (i) to 1 in step S103, and in the case where it is determined to be abnormal, that is, a failure (NO in step S102), the determination section 21c sets the flag (i) to 0 in step S104.

[0067] Next, in step S105, the determination section 21c determines whether the count i is (the number of data - 1). In the case where the count i is (the number of data - 1) (YES in step S105), the determination section 21c ends the process. In the case where the count i is not (the number of data - 1) (NO in step S105), the determination section 21c increments the count i in step S106 and returns to step S102. Thus, the determination of whether the sensor is normal or a failure is performed for all the sensors.

[0068] Figure 4 is a flowchart showing a process of determining the accuracy of the sensor.

[0069] First, in step S201, the determination section 21c initializes the count i to i = 0.

[0070] Next, in step S202, the determination section 21c confirms whether the flag (i) is 1. In the case where the flag (i) is 1 (YES in step S202), the processing proceeds to step S203. In the case where the flag (i) is 0 (NO in step S202), the processing proceeds to step S206.

[0071] In step S203, the determination section 21c determines whether the difference between the sensor value (i) and the operation value (i) is less than a prescribed value. The prescribed value is stored in the storage section 22, for example, in advance, and is set in accordance with the precision allowed for the sensor. In the case where the difference is less than the prescribed value (YES in step S203), the determination section 21c determines that the precision of the sensor (i) is the prescribed degree or more (S204). In the case where the difference is the prescribed value or more (NO in step S203), the determination section 21c changes the flag (i) to 0 in step S205. Thereafter, the processing proceeds to step S206.

[0072] Next, in step S206, the determination section 21c determines whether the count i is (the number of data - 1). In the case where the count i is (the number of data - 1) (YES in step S206), the determination section 21c ends the processing. In the case where the count i is not (the number of data - 1) (NO in step S206), the determination section 21c increments the count i in step S207 and returns to step S202. Thus, the determination of whether the sensor is normal or faulty and whether the sensing precision has decreased by the prescribed degree or more is performed for all the sensors.

[0073] Figure 5 is a flowchart showing the processing of making training data.

[0074] First, in step S301, the training data making section 21d initializes the count i to i = 0.

[0075] Next, in step S302, the training data making section 21d confirms whether the flag (i) is 1. In the case where the flag (i) is 1 (YES in step S302), the training data making section 21d determines that the sensor value (i) is used in step S303. In the case where the flag (i) is 0 (NO in step S302), it is determined that the operation value (i) is used in step S304.

[0076] Next, in step S305, the training data creating section 21d determines whether the count i is (the number of data - 1). In the case where the count i is (the number of data - 1) (YES in step S305), the process proceeds to step S307. In the case where the count i is not (the number of data - 1) (NO in step S305), the training data creating section 21d increments the count i in step S306 and returns to step S302.

[0077] In step S307, the training data creating section 21d creates training data using the sensor value (i) or the operation value (i) determined to be used, stores the training data in the storage section 22, and ends the process.

[0078] After that, the control section 21 reads out the training data stored in the storage section 22 and transmits the training data to the communication section 23. This transmission process can be performed sequentially after the training data is stored or can be performed collectively after a certain number of training data is stored in the storage section 22.

[0079] Figure 6 is a flowchart showing a process of executing machine learning.

[0080] First, the learning section 31b determines whether the storage section 32 has accumulated training data of a prescribed amount or more. The prescribed amount is a value that is an index of setting the quality of machine learning to a permissible level and is stored in the storage section 32 in advance. In the case where the training data of the prescribed amount or more is accumulated (YES in step S401), the learning section 31b performs machine learning in step S402 and ends the process. Further, in the case where the training data of the prescribed amount or more is not accumulated (NO in step S401), the learning section 31b ends the process.

[0081] <Example of Sensor, Sensor Value, Operation Value>

[0082] Next, one example of a method of operating an operation value equivalent to a sensor value other than the used sensor value using the sensor value will be described. Hereinafter, a case where a device is a vehicle that performs cruise control using an engine as a power source and a sensor detects a state or characteristics of the device will be described.

[0083] Figure 7 is a block diagram showing control of cruise control in a vehicle. The vehicle 200 is provided with a controller 40, a sensor group 50, and an actuator group 60.

[0084] The controller 40 is, for example, an ECU (Electronic Control Unit) that has a microcomputer configured mainly of a processor such as a CPU and a storage section such as a ROM, a RAM, and the like. The controller 40 has, as functional modules, a subtracter 41, an F / F control section 42, an F / B control section 43, and an adder 44, which are realized by the processor reading out a program stored in the storage section to a work area of the storage section and executing it. As illustrated in a block diagram in Figure 8 The F / F control section 42 has an operation amount inverse operation model 42a and an F / F gain imparting section 42b. The controller 40 performs control of the cruise control, the details of which will be described later.

[0085] The sensor group 50 has sensor groups 51, 52, 53. The sensor group 51 includes sensors that detect vehicle state / environmental information of the vehicle 200. The vehicle state / environmental information refers to information such as vehicle speed, outside air temperature, outside air pressure, and the like, and can also include engine speed, for example. The sensor group 51 outputs a signal containing the vehicle state / environmental information to the F / F control section 42. The sensor group 52 includes sensors that detect vehicle speed and acceleration. The sensor group 52 outputs a signal containing information of the vehicle speed and acceleration to the subtracter 41. The sensor group 53 includes sensors that detect and output operation amounts of an accelerator and a brake.

[0086] The actuator group 60 includes actuators that act in accordance with an instruction signal of the operation amount from the controller 40, and specifically includes a throttle actuator and a brake actuator. The throttle actuator controls the amount of supply of air to the engine (throttle opening degree) in accordance with the instruction signal of the operation amount from the controller 40, thereby controlling the driving force of the vehicle 200. The brake actuator controls the brake system in accordance with the instruction signal of the operation amount from the controller 40, thereby controlling the brake force imparted to the wheels of the vehicle 200.

[0087] Next, the control of the cruise control will be described. A signal containing information of a target vehicle speed and a target acceleration for the cruise control is input to the controller 40. The target vehicle speed and the target acceleration are input to the subtracter 41, and the sensed vehicle speed and acceleration from the sensor group 52 are input. The subtracter 41 performs subtraction processing, and outputs a value of the difference between the target vehicle speed and the sensed vehicle speed and a value of the difference between the target acceleration and the sensed acceleration to the F / B control section 43.

[0088] The F / B control section 43 performs feedback control based on the input difference values, and outputs, as an F / B control operation amount, an operation amount of the accelerator and the brake that is calculated in such a way that the difference becomes small to the adder 44.

[0089] The F / F control section 42 is input with a signal containing information of a target acceleration and a signal containing vehicle state / environment information, performs feedforward control based on these information and, for example, a stored map, and outputs an operation amount of an accelerator, a brake as an F / F control operation amount to the adder 44.

[0090] The adder 44 performs addition processing of an F / B control operation amount and an F / F control operation amount, and outputs an operation amount of an accelerator, a brake to the actuator group 60. Thereby, control of cruise control is realized.

[0091] Here, in the F / F control section 42, when a signal containing information of a target acceleration and a signal containing vehicle state / environment information are input, an operation amount inverse operation model 42a outputs an operation amount that is a basis of an F / F control operation amount to an F / F gain imparting section 42b. The F / F gain imparting section 42b imparts a prescribed gain to the input operation amount, and outputs as an F / F control operation amount.

[0092] The operation amount inverse operation model 42a can use a learned model realized by machine learning, a physical model. The physical model is, for example, a model that performs the following operation. That is, the operation amount inverse operation model 42a using a physical model calculates a required driving force from a target acceleration, a vehicle weight, and a vehicle running resistance that is a function of a vehicle speed. Then, the operation amount inverse operation model 42a calculates a required engine torque using a tire radius, a differential ratio, a T / M gear, and the like from the required driving force. Then, the operation amount inverse operation model 42a determines an operation amount to a throttle actuator using an engine torque characteristic map from the required engine torque and an engine rotation speed, and determines a required accelerator opening degree.

[0093] Figure 9 is a diagram that describes a sensor value and an operation value. The operation amount inverse operation model 42a is input with an acceleration (sensor), an outside air temperature (sensor), an outside air pressure (sensor), a vehicle speed (sensor), and the like as input data from the sensor group 50 as a sensor value, and outputs an accelerator operation amount (operation), a brake operation amount (operation) as an operation value. Further, the sensor group 50 outputs an accelerator operation amount (sensor), a brake operation amount (sensor) as a sensor value.

[0094] The operation amount inverse operation model 42a can be used as the operation unit 21b in the training data production device 20. In a case where the operation amount inverse operation model 42a is a learned model, as training data that can be used in the machine learning device 30, the training data produced by the training data production device 20 becomes, for example, the following data set. For example, as Example 1, in a case where all the sensors are normal, the {input data set : output data set} as the training data is {acceleration (sensor), outside air temperature (sensor), outside air pressure (sensor), vehicle speed (sensor) : accelerator operation amount (sensor), brake operation amount (sensor)}.

[0095] Further, as Example 2, in a case where the sensor that senses the accelerator operation amount is abnormal, the {input data set : output data set} as the training data is, for example, {acceleration (sensor), outside air temperature (sensor), outside air pressure (sensor), vehicle speed (sensor) : accelerator operation amount (operation), brake operation amount (sensor)}.

[0096] Further, as Example 3, in a case where the sensor that senses the brake operation amount is abnormal, the {input data set : output data set} as the training data is, for example, {acceleration (sensor), outside air temperature (sensor), outside air pressure (sensor), vehicle speed (sensor) : accelerator operation amount (sensor), brake operation amount (operation)}.

[0097] Further, even in a case where the operation amount inverse operation model 42a is a physical model, the operation amount inverse operation model 42a can be used as the operation unit 21b in the training data production device 20. The operation amount inverse operation model 42a operates the operation values of the accelerator operation amount (operation), the brake operation amount (operation), and the like. The training data production unit 21d can appropriately produce the training data of Example 2 or Example 3 using these operation values, as training data for the machine learning device 30 to produce a learned model that can be used as the operation amount inverse operation model 42a.

[0098] Note that, in the above-described embodiments, the sensing accuracy of the sensor is determined by comparing the difference between the sensor value and the operation value with the prescribed value, but the sensing accuracy of the sensor can be determined by comparing the difference between the cumulative value of the sensor value and the cumulative value of the operation value, the ratio of inclination, the difference of inclination, or the like, with the prescribed value.

[0099] Furthermore, in the above-described embodiment, the training data creation device 20 and the machine learning device 30 are mounted on the device 100, but the machine learning device 30 can be installed on, for example, a server device, and the training data creation device 20 and the machine learning device 30 can be installed on, for example, a server device. In this case, data can be transmitted and received between the device 100 and the server device through wireless communication via a communication network. Furthermore, the training data creation device 20 can have a learning unit to function as the machine learning device. In this case, the machine learning device 30 can not be present.

[0100] Furthermore, the present application is not limited to the above-described embodiments. The present application also includes schemes in which the above-described respective components are appropriately combined. Furthermore, further effects, modifications can be easily derived by those skilled in the art. Thus, the present application is not limited to the above-described embodiments, and various changes can be made.

Claims

1. A training data creation device comprising: an acquisition unit that acquires a plurality of sensor values from a plurality of sensors of a vehicle; a calculation unit that calculates a calculation value corresponding to a sensor value other than the used sensor value using a sensor value included in the plurality of sensor values; a determination unit that determines whether each of the plurality of sensors is normal; and a training data creation unit that creates training data from an input / output data set including the calculation value corresponding to the sensor value of the sensor determined to be abnormal in a cruise control of the vehicle when at least one sensor is determined to be abnormal by the determination unit, wherein the training data creation unit includes the sensor value of the sensor determined to be normal by the determination unit in the input / output data set, the training data creation device further comprises a storage unit that stores a determination result obtained by the determination unit, and the training data creation unit uses a calculation value corresponding to the sensor value of the sensor of which the past determination result stored in the storage unit is normal as the calculation value to be included in the input / output data set.

2. The training data creation device according to claim 1, wherein the determination unit determines the sensor to be abnormal when the sensor has a failure or a sensing accuracy is reduced by a predetermined degree or more.

3. The training data creation device according to claim 1 or 2, wherein the determination unit determines a sensing accuracy of the sensor by comparing the sensor value with the calculation value.

4. A training data creation method comprising: acquiring a plurality of sensor values from a plurality of sensors of a vehicle; calculating a calculation value corresponding to a sensor value other than the used sensor value using a sensor value included in the plurality of sensor values; storing the calculation value in a storage unit; determining whether each of the plurality of sensors is normal in a cruise control of the vehicle and storing a determination result in the storage unit; creating training data from an input / output data set including the calculation value corresponding to the sensor value of the sensor determined to be abnormal read out from the storage unit when at least one sensor is determined to be abnormal, using a calculation value corresponding to the sensor value of the sensor of which the past determination result stored in the storage unit is normal as the calculation value to be included in the input / output data set; and including the sensor value of the sensor determined to be normal in the input / output data set.

5. A computer program product comprising a program that causes a processor having hardware to perform: acquiring a plurality of sensor values from a plurality of sensors of a vehicle; calculating a calculation value corresponding to a sensor value other than the used sensor value using a sensor value included in the plurality of sensor values; storing the calculation value in a storage unit; determining whether each of the plurality of sensors is normal in a cruise control of the vehicle and storing a determination result in the storage unit; creating training data from an input / output data set including the calculation value corresponding to the sensor value of the sensor determined to be abnormal read out from the storage unit when at least one sensor is determined to be abnormal, using a calculation value corresponding to the sensor value of the sensor of which the past determination result stored in the storage unit is normal as the calculation value to be included in the input / output data set; and including the sensor value of the sensor determined to be normal in the input / output data set. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ In a case where it is determined that at least one sensor is abnormal, training data is created from an input / output data set including the operation value corresponding to the sensor value of the sensor determined to be abnormal, which is read from the storage section, and the operation value corresponding to the sensor value of the sensor determined to be normal in the past determination result stored in the storage section is included as the operation value to be included in the input / output data set. The sensor value of the sensor determined to be normal is included in the input / output data set.

Images