Estimation device, estimation method and estimation program, and generation device, generation method and generation program of learned machine learning model

Abstract

To appropriately estimate timing when damage occurs at a tire.SOLUTION: An estimation device estimates regarding damage at a tire mounted onto a vehicle according to data on a physical amount obtained during a travel of the vehicle. The estimation device includes a derivation part configured to: enter, into one or a plurality of learned machine learning models, input data at least including data on oscillation generated at the vehicle due to a rotation of the tire, data on a rotational speed or a rotational acceleration of the tire, and data on a strain caused in the tire during travel of the vehicle; and derive output data from the one or plurality of learned machine learning models. The output data corresponds to at least one of: timing when damage occurs at the tire while a vehicle is traveling; a presence or absence of damage at the tire at a point where the data of the physical amount is acquired; and a portion where the damage occurs at the tire.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an estimation device, an estimation method, and an estimation program for estimating the timing of occurrence of damage to a tire, as well as a generation device, a generation method, and a generation program for a trained machine learning model. [Background technology]

[0002] Patent Document 1 discloses a technique for detecting the occurrence of a failure in a tire component before the tire bursts. The technique disclosed in Patent Document 1 includes a measurement step in which the noise generated when the tire rolls on the road surface is measured by changing the tire's rolling speed stepwise; a calculation step in which a change amount or rate of change in the measurement result is calculated based on the measurement result obtained in the previous measurement step; and a determination step in which the change amount or rate of change is compared with a predetermined threshold and, if this change amount or rate of change exceeds the threshold, a determination is made that a failure has occurred in a tire component. Patent Document 1 also discloses that, instead of noise, the axial force of the tire's mounting hub can be measured and tire failure can be detected based on the measured axial force. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 4259931 Summary of the Invention [Problem to be solved by the invention]

[0004] As disclosed in Patent Document 1, before a tire bursts, abnormalities may be detected in the sound or changes in the axial force. However, there is more than one type of phenomenon that can be a precursor to a burst, and it has been empirically found that a burst can occur even when no noticeable abnormalities in the noise or axial force are detected. This is thought to be because there are various patterns of partial tire damage that can be the starting point for a burst. For this reason, it remains difficult to properly estimate the timing of a burst and the partial tire damage that can cause a burst, and improvements have been desired.

[0005] The present invention aims to provide a technology for appropriately estimating, based on data of physical quantities acquired while a vehicle is traveling, at least one of the timing at which tire damage will occur, whether tire damage exists at the time the data of the physical quantities is acquired, and the location at which damage will occur in the tire. [Means for solving the problem]

[0006] An estimation device according to a first aspect of the present invention is an estimation device that estimates damage to a tire mounted on a vehicle based on physical quantity data acquired while the vehicle is traveling, and includes a derivation unit that inputs input data, including at least data on vibrations occurring in the vehicle due to the rotation of the tire, data on the rotation speed or rotation acceleration of the tire, and data on strain occurring in the tire while the vehicle is traveling, into one or more trained machine learning models, and derives output data from the one or more trained machine learning models. The output data corresponds to at least one of the timing at which damage occurs to the tire while the vehicle is traveling, the presence or absence of damage to the tire at the time the physical quantity data is acquired, and a location at which damage occurs in the tire.

[0007] An estimation device according to a second aspect of the present invention is the estimation device according to the first aspect, wherein the input data further includes at least one of data on sounds generated due to rotation of the tires, data on external forces applied to the tires, data on the temperature of the tires, and data on the air pressure of the tires.

[0008] An estimation device according to a third aspect of the present invention is the estimation device according to the first or second aspect, wherein the derivation unit inputs the input data to a plurality of trained machine learning models, each of which is associated with a respective one of a plurality of parts of the tire, and the output data derived from each of the plurality of trained machine learning models corresponds to the presence or absence of damage to each of the plurality of parts at the time of acquiring data on the physical quantity.

[0009] An estimation device according to a fourth aspect of the present invention is the estimation device according to any one of the first aspect to the third aspect, further comprising a notification generation unit that generates a notification to notify at least one of the timing at which damage will occur to the tire while the vehicle is traveling, the presence or absence of damage to the tire at the time the data of the physical quantity is acquired, and a location at which damage will occur in the tire.

[0010] An estimation method according to a fifth aspect of the present invention is an estimation method executed by one or more computers for making an estimation regarding damage to a tire mounted on a vehicle based on data of physical quantities acquired while the vehicle is traveling, the estimation method including: inputting input data including at least data on vibrations occurring on the vehicle due to rotation of the tire, data on the rotational speed or rotational acceleration of the tire, and data on strain occurring in the tire while the vehicle is traveling into one or more trained machine learning models; and deriving output data from the one or more trained machine learning models. The output data corresponds to at least one of timing when damage occurs to the tire while the vehicle is traveling, whether or not the tire is damaged at the time the data of the physical quantities is acquired, and a location where the damage occurs in the tire.

[0011] An estimation program according to a sixth aspect of the present invention is an estimation program for making an estimation regarding damage to a tire mounted on a vehicle based on data of physical quantities acquired while the vehicle is traveling, the estimation program causing one or more computers to input input data including at least data on vibrations occurring on the vehicle due to the rotation of the tire, data on the rotation speed or rotation acceleration of the tire, and data on strain occurring in the tire while the vehicle is traveling to one or more trained machine learning models, and deriving output data from the one or more trained machine learning models. The output data corresponds to at least one of the timing at which damage occurs to the tire while the vehicle is traveling, the presence or absence of damage to the tire at the time the data of the physical quantities is acquired, and a location at which the damage occurs in the tire.

[0012] A seventh aspect of the present invention provides an apparatus for generating a trained machine learning model, the apparatus comprising: an acquisition unit that acquires a training dataset including a plurality of first data, the first data being data on physical quantities acquired while a vehicle is traveling, the first data including at least data on vibrations occurring in the vehicle due to the rotation of tires mounted on the vehicle, data on the rotation speed or rotation acceleration of the tires, and data on strain occurring in the tires while the vehicle is traveling; and second data that has been combined with the first data in advance; and a learning unit that inputs the first data to a machine learning model, derives output data from the machine learning model, and adjusts parameters that define the machine learning model so as to reduce an error between the output data and the second data. The second data includes at least one of the timing at which damage occurred to the tire while the vehicle was traveling, whether or not the tire was damaged at the time the data on the physical quantities was acquired, and a location at which the damage occurred in the tire.

[0013] A method for generating a trained machine learning model according to an eighth aspect of the present invention is a method for generating a trained machine learning model executed by one or more computers, and includes: acquiring a training dataset including a plurality of first data, the first data being data on physical quantities acquired while a vehicle is traveling, the first data including at least data on vibrations occurring in the vehicle due to the rotation of tires mounted on the vehicle, data on the rotational speed or rotational acceleration of the tires, and data on strain occurring in the tires while the vehicle is traveling; and second data pre-combined with the first data; inputting the first data into a machine learning model to derive output data from the machine learning model; and adjusting parameters defining the machine learning model so as to reduce an error between the output data and the second data. The second data includes at least one of the timing at which damage occurred to the tire while the vehicle was traveling, whether or not the tire was damaged at the time the data on the physical quantities was acquired, and a location at which the damage occurred in the tire.

[0014] A ninth aspect of the present invention provides a program for generating a trained machine learning model that causes one or more computers to execute the following steps: acquire training datasets including first data, which are data on physical quantities acquired while a vehicle is traveling, the first data including at least data on vibrations occurring in the vehicle due to the rotation of tires mounted on the vehicle, data on the rotational speed or rotational acceleration of the tires, and data on strain occurring in the tires while the vehicle is traveling; and second data, which are pre-combined with the first data; input the first data into a machine learning model to derive output data from the machine learning model; and adjust parameters that define the machine learning model so as to reduce an error between the output data and the second data. The second data includes at least one of the timing at which damage occurred to the tire while the vehicle was traveling, whether or not the tire was damaged at the time the data on the physical quantities was acquired, and a location at which the damage occurred in the tire. [Effects of the Invention]

[0015] According to the present invention, it is possible to use a trained machine learning model to appropriately estimate at least one of the timing at which tire damage will occur, whether tire damage exists at the time the physical quantity data is acquired, and the location of the tire where the damage will occur, based on data on physical quantities acquired while the vehicle is traveling. This allows for more detailed analysis of tire damage and provides more insights. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a diagram showing the overall configuration of a damage notification system according to an embodiment; [Figure 2] FIG. 2 is a block diagram showing the electrical configuration of the estimation device according to an embodiment. [Figure 3] FIG. 1 is a diagram illustrating an example of the configuration of a machine learning model. [Figure 4] 1 is a flowchart showing the flow of an estimation method according to an embodiment. [Figure 5] FIG. 2 is a block diagram showing the electrical configuration of a generating device according to an embodiment. [Figure 6] 1 is a flowchart showing the flow of a generation method according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0017] Hereinafter, an estimation device, an estimation method, and an estimation program, as well as a generation device, a generation method, and a generation program for a trained machine learning model according to embodiments of the present invention will be described.

[0018] <1. Overview of the Damage Notification System> FIG. 1 is a diagram showing the overall configuration of a tire T damage notification system 1 (hereinafter also simply referred to as "system 1") according to one embodiment of the present invention. The system 1 is mounted on a vehicle 3 and is configured to estimate the timing at which damage will occur in a tire T that is a target for damage detection, as well as the location of the damage in the tire T, and to notify a person of the damage before it occurs or in the early stages of the damage. The system 1 includes an estimation device 2 connected to a communication network 4 and capable of mutual data communication, a computer 30 (30a, 30b), and sensors 5 to 10 mounted on the vehicle 3 or the tire T. In this embodiment, the vehicle 3 is a vehicle used for durability evaluation of the tire T, but the system 1 is not limited thereto and can be applied to general vehicles. In the following description, a tire T mounted on the left front wheel hub of the vehicle 3 will be used as an example, but the tires of the vehicle 3 monitored by the system 1 may be mounted anywhere on the vehicle 3, and there may be multiple tires.

[0019] A typical damage to the tire T is a burst. Other than a burst, there is partial damage occurring in each portion, such as the tread portion, the side portion (for example, the shoulder portion and the sidewall portion), and the bead portion. Although such partial damage is often not visible from the outside of the tire T, it can be the cause of a burst, and if the tire T continues to rotate, there is a high possibility that it will eventually burst. The estimation device 2, which will be described later, is configured to estimate the timing at which a burst will occur in the tire T and the presence or absence of damage in each portion at a certain point in time, based on time-series data of physical quantities acquired by the vehicle 3.

[0020] <2. Vehicles> The vehicle 3 includes a power source (not shown), an on-board device 30a, and an alarm 31. The on-board device 30a is an on-board ECU (Electronic Control Unit) built into the vehicle 3, and its hardware is composed of one or more microcomputers. The on-board device 30a controls the operation of the power source of the vehicle 3 and the operation of various systems. The on-board device 30a is also connected to sensors 5 to 10 (described later) so as to be able to communicate data, and receives time-series data output from the sensors 5 to 10 while the vehicle 3 is traveling. The on-board device 30a creates a data set by synchronizing the time-series data, and transmits the time-series data set for a predetermined time range to the estimation device 2. The on-board device 30a also receives estimation results from the estimation device 2 in a timely manner. The on-board device 30a may also be configured to timely acquire road surface information indicating the condition of the road surface on which the vehicle 3 is traveling, weather information for the location where the vehicle 3 is currently located, and the like from various servers via the communication network 4.

[0021] The alarm 31 is connected to the in-vehicle device 30a and operates in response to commands from the in-vehicle device 30a. More specifically, the alarm 31 outputs a message notifying at least one of the following based on the estimation results from the estimation device 2: that the tire T may burst while traveling, the estimated timing at which the tire T is likely to burst, the estimated timing at which partial damage is likely to occur in the tire T, and the estimated location at which damage is likely to occur. The message may be output in any manner, such as a warning sound such as a predetermined buzzer, a voice, or a display of text or graphics. Therefore, the alarm 31 may be configured with a speaker, a liquid crystal element, a liquid crystal display, a plasma display, an organic EL display, or the like. The location at which the alarm 31 is installed is not particularly limited, but it is preferably installed on or near the instrument panel so that the driver of the vehicle 3 can easily recognize it.

[0022] Note that a portable computer 30b such as a smartphone, laptop computer, or tablet carried by the driver of the vehicle 3 or the like may output the message instead of or in addition to the alarm 31. In this case, the computer 30b may store a program that receives the estimation result from the estimation device 2 via the communication network 4 and outputs the message based on the estimation result.

[0023] 3. Sensors The sensors 5 to 10 each detect a predetermined physical quantity while the vehicle 3 is traveling, and output time-series data of the detected physical quantity. These time-series data are synchronized by the on-vehicle device 30a and transmitted to the estimation device 2 as time-series data separated by a predetermined time range. This time range is, for example, 1 to 10 minutes, 1 to 60 seconds, or 0.1 to 1 second, and can be appropriately selected as long as it is a time range that allows estimation by the estimation device 2 and allows advance notification of damage to the tire T (particularly, a burst).

[0024] [Rotational speed sensor] The sensor 5 detects the rotational speed of the tire T (more precisely, the rotational speed of the wheel hub on which the tire T is mounted). The sensor 5 is not particularly limited as long as it can detect the rotational speed of a rotating body. For example, a sensor of the type that measures the rotational speed from the output signal of an electromagnetic pickup can be used, or a sensor of the type that generates electricity using rotation, like a dynamo, and measures the rotational speed from the voltage generated. The mounting position of the sensor 5 is also not particularly limited.

[0025] The on-vehicle device 30a differentiates the acquired time series data of the rotational speed of the tire T to acquire time series data of the rotational acceleration of the tire T. In order to improve the estimation accuracy by the estimation device 2, it is preferable that the on-vehicle device 30a perform a filtering process on the time series data of the rotational speed of the tire T to reduce noise from the road surface. Various techniques for filtering are already known, so a description thereof will be omitted here.

[0026] [Strain sensor] The sensors 6 to 8 each detect strain occurring in a respective portion of the tire T. The sensors 6 to 8 are not particularly limited as long as they can convert strain at the portion where they are attached into an electrical signal, and any type can be used. In this embodiment, sensor 6 is attached to the inside of the tread portion of the tire T, sensor 7 is attached to the inside of the side portion, and sensor 8 is attached to the inside of the bead portion. Accumulation of mechanical strain contributes to damage to the tire T. For this reason, it is considered that the time series data of strain of the tire T is correlated with the timing of occurrence of damage to the tire T. In this embodiment, by acquiring the time series data of strain of the tire T for each portion, the estimation device 2 can more accurately estimate the portion of the tire T where damage will occur. Note that, in order to improve the estimation accuracy by the estimation device 2, it is preferable that the on-vehicle device 30a perform a filtering process to reduce noise on the time series data of strain at each portion of the tire T.

[0027] [Sound sensor] The sensor 9 detects sound generated due to the rotation of the tire T. The sensor 9 is not particularly limited as long as it can convert sound pressure into an electrical signal, and a known microphone or the like can be used. The sound generated due to the rotation of the tire T may be air noise inside the tire T, the rotation sound of a wheel hub to which the tire T is attached, or the sound of the tire T colliding with the road surface. In this embodiment, the sensor 9 is attached to the wheel hub to which the tire T is attached or in its vicinity. However, the attachment location of the sensor 9 is not limited thereto and may be, for example, inside the tire T or a portion of the body of the vehicle 3 close to the road surface, and can be selected as appropriate. When damage occurs to any part of the tire T, unusual characteristics may appear in the time-series data of the sound generated due to the rotation of the tire T. Furthermore, abnormal noise may occur even before the tire T bursts. For this reason, it is considered that the time-series data of the sound is correlated with the timing of occurrence of damage to the tire T.

[0028] [External force sensor] The sensor 10 detects an external force acting on the tire T. The sensor 10 is not particularly limited as long as it can convert the external force acting on the tire T into an electrical signal; for example, a known six-component wheel force meter or the like can be used. The time-series data output by the sensor 10 may include only a normal force in the vertical direction (z-axis direction), or may include at least one of a load in a direction perpendicular to the rotation axis of the tire T (x-axis direction), a load in a direction parallel to the rotation axis (y-axis direction), and torque around the x, y, and z axes. The normal force acting on the tire T from the road surface is proportional to the driving force of the vehicle 3. However, if damage occurs to the tire T, the normal force may deviate from a normal range due to the energy loss or wobble of the rotation axis of the tire T. For this reason, it is believed that the time-series data of external forces, including the normal force, is correlated with the timing of occurrence of damage to the tire T.

[0029] [Vibration sensor] The sensor 11 detects vibrations occurring in the vehicle 3 due to the rotation of the tire T. The sensor 11 is not particularly limited as long as it can convert the frequency of the vibration and at least one of the displacement, velocity, and acceleration into an electrical signal. For example, a piezoelectric sensor, an electrodynamic sensor, a servo sensor, an eddy current sensor, a capacitance sensor, an optical sensor, or the like can be used. In this embodiment, the sensor 11 is attached to an arm 32 of a suspension system that supports a wheel hub on which the tire T is mounted. When wobbling is transmitted to the arm 32 due to damage to the tire T or before the tire T bursts, vibrations with unusual characteristics may appear in at least a part of the suspension system. This vibration may be caused by a standing wave, which is a precursor to a burst. Therefore, the time-series data of the vibration is considered to be correlated with the timing of occurrence of damage to the tire T. In order to improve the estimation accuracy of the estimation device 2, it is preferable that the on-vehicle device 30a perform a filtering process on the time-series data of the vibration to reduce noise caused by vibrations from the road surface.

[0030] In addition to the sensors 5 to 11, the system 1 may include at least one of an air pressure sensor that detects the air pressure of the tire T and a temperature sensor that detects the temperature of the tire T. The air pressure sensor is not particularly limited, and for example, an air pressure sensor that constitutes a known tire pressure monitoring system (TPMS) can be used. If the air pressure of the tire T is insufficient, standing waves are likely to occur, and if the air pressure is excessive, it is likely to burst or cause partial damage due to impact from the road surface. Therefore, it can be said that the time series data of the air pressure of the tire T indicates whether or not the tire T is in a condition in which it is likely to burst or cause partial damage.

[0031] The temperature sensor is not particularly limited as long as it can convert the temperature of at least one of the rubber and non-rubber components that make up the tire T into an electric signal, and any type can be used. The location where the temperature sensor is attached is also not particularly limited. Damage to the tire T occurs due to the accumulation of heat generation and mechanical strain. For this reason, it is believed that time-series data on the temperature of a specific part of the tire T is correlated with the timing of damage occurrence in that part. It is also known that the temperature of the tire T rises sharply when a standing wave occurs in the tire T. Note that the temperature sensor may be installed outside the tire T instead of inside the tire T, and configured to detect the temperature of the tire T in a non-contact manner.

[0032] 4. Configuration of the estimation device The estimating device 2 is a general-purpose computer in terms of hardware, and is realized as an information processing terminal such as a desktop personal computer, a laptop personal computer, a tablet, or a smartphone. The estimating device 2 is manufactured by installing a program 232 into the general-purpose computer from a computer-readable storage medium 233 such as a CD-ROM or a USB memory, or via a network. The program 232 causes the estimating device 2 to perform the operations described below.

[0033] The estimation device 2 includes a control unit 20, a display unit 21, an input unit 22, a storage unit 23, and a communication unit 24. These units 20 to 24 are connected to one another via a bus line and are capable of communicating with one another. The display unit 21 can be configured with a liquid crystal display, an organic EL display, a plasma display, a touch panel display, or the like, and displays output data derived from a machine learning model (described later). This display can be confirmed by, for example, a user (such as a developer of tire T) who performs durability evaluation of tire T. The input unit 22 can be configured with a mouse, keyboard, touch panel, or the like, and accepts operations on the estimation device 2. The display unit 21 and the input unit 22 may both be configured with the same touch panel display.

[0034] The storage unit 23 can be configured with a non-volatile memory such as a hard disk or a flash memory. In addition to storing a program 232, the storage unit 23 also stores parameters that define trained machine learning models 231A to 231D (hereinafter simply referred to as "trained models 231A to 231D") constructed by machine learning. The trained models 231A to 231D will be described later.

[0035] The control unit 20 may be configured with a CPU, a GPU (Graphics Processing Unit), a ROM, a RAM, etc. The control unit 20 reads and executes a program 232 stored in the storage unit 23, thereby virtually operating as an acquisition unit 20A, a derivation unit 20B, and a notification generation unit 20C. The acquisition unit 20A acquires time-series data of the sensors 5 to 11 transmitted from the in-vehicle device 30a via the communication unit 24, etc. The derivation unit 20B inputs input data (described later) to trained models 231A to 231D and derives output data from the trained models 231A to 231D. The notification generation unit 20C generates a notification notifying of (estimated) damage to the tire T based on the output data derived by the derivation unit 20B, and transmits the notification to the in-vehicle device 30a. The communication unit 24 functions as a communication interface for data communication via the communication network 4 or other networks, and data communication with external devices.

[0036] <5. Configuring the trained model> The configurations of the trained models 231A to 231D are described below. In this embodiment, the trained models 231A to 231D share a common layer structure, but are defined by different parameters and have different input and output data. More specifically, the input data to the trained model 231A is a dataset of time-series data for a predetermined time range created by the on-board device 30a based on the time-series data output from the sensors 5 to 11 while the vehicle 3 is traveling. The input data to the trained model 231B is a dataset of time-series data for a predetermined time range created by the on-board device 30a based on the time-series data output from the sensors 5, 6, 9 to 11 while the vehicle 3 is traveling. The input data to the trained model 231C is a dataset of time-series data for a predetermined time range created by the on-board device 30a based on the time-series data output from the sensors 5, 7, 9 to 11 while the vehicle 3 is traveling. The input data to the trained model 231D is a data set of time-series data for a predetermined time range created by the on-vehicle device 30a based on the time-series data output from the sensors 5, 8 to 11 while the vehicle 3 is traveling.

[0037] The output data output from the trained model 231A corresponds to the timing at which the tire T begins to burst. This output data may be a value corresponding to the number of seconds from when the vehicle 3 starts traveling until when the tire T begins to burst, for example.

[0038] The output data output from trained models 231B, 231C, and 231D correspond to the presence or absence of damage to the tread, side, and bead portions of tire T at the time when the time series data is output. That is, trained models 231B, 231C, and 231D are each constructed to estimate the presence or absence of damage to a specific portion of a tire at a certain point in time. These output data can be values corresponding to, for example, the probability that a specific portion has damage (or the probability that a specific portion does not have damage).

[0039] 3 is a diagram illustrating the configuration of trained models 231A to 231D according to this embodiment. As shown in FIG. 3, trained models 231A to 231D are each a model based on a neural network (NN). In this embodiment, parameters defining trained models 231A to 231D (for example, weight w and bias b, which will be described later) are different from each other.

[0040] Each of the machine learning models 231A to 231D includes an input layer 2310 having a predetermined number of nodes, a plurality of intermediate layers 2311, and an output layer 2312. The input layer 2310 is a layer for reading input data, and has the same number of nodes as the number of data included in the input data. Input values are output as is from each node of the input layer 2310, and are input to each node of the first intermediate layer 2311, which is connected immediately after the input layer 2310.

[0041] Each node in the first hidden layer 2311 calculates a value by multiplying the input by a weight w and adding a bias b. The calculated value is then converted using an activation function and output to the node in the subsequent hidden layer 2311. The input here is a plurality of values read by the input layer 2310, and it is possible to appropriately set which node's value in the input layer 2310 is input. The weight w has been adjusted for each value passed from each node in the input layer 2310 to each node in the first hidden layer 2311 through a learning process described below. In other words, for each node in the first hidden layer 2311, weights w corresponding to the number of values input thereto have been optimized. In addition, biases b corresponding to the number of nodes have also been optimized through a learning process described below.

[0042] The activation function is a function for realizing nonlinear transformation. There are no particular limitations on the activation function, and any known activation function such as a tanh function, a sigmoid function, a ReLU function, a step function, an ELU function, or a Softmax function can be used.

[0043] The values output from each node of the first hidden layer 2311 are input to each node of the subsequent hidden layer 2311. The subsequent hidden layer 2311 also has a predetermined number of nodes, and for each node, the weight w and bias b are optimized by the learning process described below. Like the first hidden layer 2311, each node outputs a value calculated for the input value using the corresponding weight w and bias b and an activation function. That is, in principle, each hidden layer 2311 repeats the same processing as the first hidden layer 2311. However, in the last hidden layer 2311, a linear transformation is performed on the input using the optimized weight w and bias b, but a nonlinear transformation using an activation function is not performed. That is, the linearly transformed values output from each node of the last and subsequent hidden layer 2311 are input to each node of the output layer 2312. The number of hidden layers 2311 is preferably two or more, but may be one. It is not particularly limited and can be set appropriately.

[0044] The output layer 2312 has nodes whose number corresponds to the output data, and each node typically outputs the input value as is. In this embodiment, the output layer 2312 has one node. The data output from the output layer 2312 becomes the output data of each of the trained models 231A to 231D. The output from the output layer 2312 of the trained model 231A is a value corresponding to the number of seconds from when the tire starts rotating to when the tire T starts to burst. The output from the output layer 2312 of the trained model 231B is a value corresponding to the probability that there is damage to the tread portion of the tire T at the time of acquisition of the time series data on which the input data is based. The output from the output layer 2312 of the trained model 231C is a value corresponding to the probability that there is damage to the side portion of the tire T at the time of acquisition of the time series data on which the input data is based. The output from the output layer 2312 of the trained model 231D is a value corresponding to the probability that there is damage to the bead portion of the tire T at the time of acquisition of the time series data on which the input data is based.

[0045] <6. Estimation method> 4 is a flowchart showing the flow of the estimation method executed by the estimation device 2 according to this embodiment. The following estimation method starts, for example, when the vehicle 3 starts traveling, is repeatedly performed almost in real time until the vehicle 3 stops traveling, and ends when the vehicle 3 stops traveling.

[0046] First, the acquisition unit 20A sequentially acquires data sets of time-series data transmitted from the in-vehicle device 30a (step S1). This data set is created by the in-vehicle device 30a filtering, as necessary, and synchronizing the time-series data of the rotational speed of the tire T, the distortion of each part of the tire T, the sound generated by the rotation of the tire T, the external force applied to the tire T, and the vibration generated during the rotation of the tire T within a predetermined time range. The acquisition unit 20A temporarily stores the acquired data sets in RAM or stores them in the storage unit 23.

[0047] Next, the acquisition unit 20A creates input data for the trained models 231A to 231D based on the dataset saved in step S1 (step S2). The input data for the trained model 231A is the dataset acquired in step S1. The input data for the trained model 231B is the dataset acquired in step S1, excluding time series data on strain in the side and bead portions. The input data for the trained model 231C is the dataset acquired in step S1, excluding time series data on strain in the tread and bead portions. The input data for the trained model 231D is the dataset acquired in step S1, excluding time series data on strain in the tread and side portions.

[0048] Next, the derivation unit 20B inputs each piece of input data created in step S2 into the trained models 231A to 231D, respectively, and derives each piece of output data for each piece of input data from the trained models 231A to 231D (step S3).

[0049] Next, the derivation unit 20B estimates whether the tire T will burst while the vehicle 3 is traveling (currently) and whether any part of the tire T has been damaged (step S4) based on the output data derived in step S3. More specifically, the derivation unit 20B estimates whether the tire T will burst while the vehicle 3 is traveling by comparing the output data from the trained model 231A with a predetermined time threshold. The derivation unit 20B also estimates whether damage has been caused to at least one of the tread portion, the side portion, and the bead portion by comparing the output data from the trained models 231B to 231D with a predetermined probability threshold.

[0050] If it is estimated in step S4 that the tire T will burst while the vehicle 3 is traveling, or if it is estimated that damage has occurred in any part of the tire T in addition to or instead of this (YES), the notification generation unit 20C generates a notification to notify that effect (step S5). The generated notification is immediately transmitted to the in-vehicle device 30a via the communication network 4. The notification may include information on the timing at which the tire T is estimated to burst and information identifying the part where the damage has occurred. After step S5, step S6 is executed.

[0051] When the in-vehicle device 30a receives the notification generated in step S5, it outputs the notification to the alarm 31. This allows the driver of the vehicle 3 to take action such as stopping the vehicle 3 and inspecting the tire T.

[0052] If it is estimated in step S4 that the tire T will not burst while the vehicle 3 is traveling and that no damage has occurred to any part of the tire T (NO), step S5 is not executed and step S6 is executed.

[0053] In step S6, the acquisition unit 20A determines whether the vehicle 3 is currently stopped. This determination can be made based on a signal from the in-vehicle device 30a, for example. When the acquisition unit 20A determines that the vehicle 3 is not stopped (NO), it executes the processes from step S1 onwards again. That is, based on the latest data set transmitted from the in-vehicle device 30a, it is estimated whether the tire T is in a state where it may burst and whether any part of the tire T has been damaged. In this way, if it is estimated that the vehicle 3 will continue to run, that the tire T will not burst, and that no part of the tire T has been damaged, steps S1 to S4 are repeated.

[0054] On the other hand, when the acquisition unit 20A determines that the vehicle 3 has stopped (YES), the estimation process by the estimation device 2 ends.

[0055] According to the above steps S1 to S6, although there is a time lag strictly speaking with respect to the time when the time series data of the sensors 5 to 11 is output, it is possible to notify the driver of the vehicle 3 before a burst occurs in the tire T or before or after partial damage occurs in the tire T.

[0056] Note that, in the case where a large time lag occurs between the acquisition of time-series data by sensors 5-11 and the derivation of an estimation result for input data based on this time-series data due to the calculation speed of estimation device 2, the following can also be done. For example, based on data from previous road travel by vehicle 3 equipped with tires of similar configuration, vehicle 3 is driven under conditions that are unlikely to cause tire T to burst, and a dataset of time-series data from sensors 5-11 is acquired. After vehicle 3 stops traveling, input data created based on the above dataset is input to trained models 231A-231D to estimate the timing of tire T bursting and whether or not damage has occurred in the tread, side, or bead portions.

[0057] <7.Generation device> FIG. 5 is a block diagram showing the electrical configuration of a generating device 2X configured to execute a method for generating trained models 231A to 231D. The generating device 2X is a general-purpose computer in terms of hardware, and is realized as an information processing terminal such as a desktop personal computer, a laptop personal computer, a tablet, or a smartphone. The generating device 2X is manufactured by installing a program 232X into a general-purpose computer from a computer-readable storage medium 233X such as a CD-ROM or a USB memory, or via a network. The program 232X causes the generating device 2X to perform the operations described below.

[0058] The generation device 2X includes a control unit 20X, a display unit 21X, an input unit 22X, a storage unit 23X, and a communication unit 24X. These units 20X to 24X are connected to one another via a bus line and can communicate with each other. The same explanations as for the display unit 21, input unit 22, and communication unit 24 of the estimation device 2 apply to the display unit 21X, input unit 22, and communication unit 24, and therefore further description will be omitted.

[0059] The control unit 20X can be configured with a CPU, a GPU, a ROM, a RAM, etc. The control unit 20X reads and executes a program 232X in the storage unit 23X, thereby virtually operating as an acquisition unit 20D and a learning unit 20E. The acquisition unit 20D acquires training data sets 234A to 234D from outside the generation device 2X via a storage medium 233X, a communication unit 24X, etc. The training data sets 234A to 234D are data sets for adjusting parameters of the trained models 231A to 231D, respectively.

[0060] The storage unit 23X can be configured with a nonvolatile memory such as a hard disk and a flash memory, etc. In the storage unit 23X, a program 232X is stored, and also learning data sets 234A to 234D are stored.

[0061] The training datasets 234A to 234D are datasets created from data on durability evaluations of multiple tires using the vehicle 3. The multiple tires preferably have a common size and the same or similar tread pattern. Furthermore, the durability evaluation conditions (road surface conditions, vehicle speed of the vehicle 3, and duration of a specific vehicle speed) are assumed to be common to the multiple tires. The training datasets 234A to 234D each include multiple first data and multiple second data to be combined with the first data. The first data of the training datasets 234A to 234D corresponds to the input data of the trained models 231A to 231D, respectively. Specifically, the first data of the training dataset 234A is created from time-series data output from the sensors 5 to 11 during a predetermined time range from 0 to t seconds during the durability evaluation of one tire. 0 second is the start time of the vehicle 3's travel, and t seconds represents the specified duration of the durability evaluation (the continuous travel time of the vehicle 3). However, for tires that burst during the durability test, the data from the time t0 seconds after the bursting began is assumed to be compensated for with a predetermined value.

[0062] The time series data output from the sensors 5 to 11 are synchronized with each other within a time range that can be considered to be the same time, and then grouped in chronological order within a predetermined time range (sufficiently shorter than t seconds). The grouped time series data is subjected to filtering similar to that performed by the in-vehicle device 30a, as necessary. The number of groups is the same for multiple tires. The multiple first data included in the training dataset 234A is a collection of the above groups for multiple tires.

[0063] The plurality of first data included in training data set 234B is similarly created based on the time series data output from sensors 5, 6, and 9 to 11. The plurality of first data included in training data set 234C is similarly created based on the time series data output from sensors 5, 7, and 9 to 11. The plurality of first data included in training data set 234D is similarly created based on the time series data output from sensors 5, 8 to 11.

[0064] The second data is data to be combined with each group of the first data, and has the same number of values per tire as the number of groups of the first data. The second data in the training dataset 234A is the timing (number of seconds t0) at which the tire began to burst after the vehicle 3 started traveling. If the tire did not burst, the second data is number of seconds t.

[0065] The second data in the training dataset 234B is the probability that damage has occurred in the tread portion of the tire. The second data is, for example, a value ranging from 0 (no damage to the tread portion) to 1 (damage to the tread portion). In other words, all of the second data for tires for which no damage to the tread portion was confirmed after the durability evaluation is 0. On the other hand, at least some of the second data for tires for which damage to the tread portion was confirmed after the durability evaluation is 1. The second data for tires for which damage to the tread portion was confirmed can be created, for example, as follows: First, all of the second data to be combined with a group of first data within a first time range in which no damage to the tread portion is estimated to have occurred is set to 0. Next, all of the second data to be combined with a group within a second time range in which damage to the tread portion is estimated to have definitely occurred is set to 1. If the first time range and the second time range are not consecutive, all of the second data to be combined with a group of first data within the time range between them is set to 0.5.

[0066] The second data of training data set 234C is the probability that damage has occurred in the side portion of the tire of the corresponding first data. The second data of training data set 234D is the probability that damage has occurred in the bead portion of the tire of the corresponding first data. The second data of training data sets 234C and 234D can each be created in the same manner as the second data of training data set 234B.

[0067] <8. Generation method> 6 is a flowchart showing the flow of a method for generating the trained models 231A to 231D, which is executed by the generation device 2X according to this embodiment. The method for generating the trained models 231A to 231D will be described below.

[0068] First, the acquiring unit 20D acquires the training data sets 234A to 234D from outside the generating device 2X via the storage medium 233X, the communication unit 24X, etc. (Step S11). The acquiring unit 20D stores the acquired training data sets 234A to 234D in the storage unit 23X. At this time, the acquiring unit 20D divides the training data sets 234A to 234D into a training data set and a test data set according to a predetermined ratio and stores them.

[0069] Next, the learning unit 20E divides each of the training data sets 234A-234D into data sets for a predetermined number of tires, and creates multiple subsets each including data sets for the same number of tires (step S12). The predetermined number is the number of data to be continuously input to each machine learning model in the next step S13, and can be set as appropriate. Each machine learning model becomes a trained model 231A-231D by repeating steps S12-S16. In other words, each machine learning model has the layer structure shown in FIG. 3 and is a model in a state where parameters have not been adjusted (i.e., before learning).

[0070] Next, the learning unit 20E selects one subset from the training data sets 234A to 234D, sequentially inputs the first data included in the selected subset to each machine learning model, and derives output data from each machine learning model (step S13). The output data from each machine learning model corresponds to the second data combined with the first data input as input data.

[0071] Next, the learning unit 20E adjusts and updates parameters so that the value of the error function between the output data derived in step S13 and the second data combined with the first data input in step S13 becomes minimum (step S14). More specifically, the learning unit 20E adjusts and updates weight coefficients, biases, etc. in the intermediate layer of each machine learning model by gradient descent or the like.

[0072] Next, the learning unit 20E determines whether one epoch of learning has been completed (step S15). In this embodiment, it is determined that one epoch of learning has been completed when steps S13 and S14 have been performed once for each subset created in step S12. If it is determined that one epoch of learning has not been completed (NO), the learning unit 20E repeats steps S13 to S14 using a subset that has not yet been used. On the other hand, if it is determined that one epoch of learning has been completed (YES), step S16 is executed. Note that in this process, a dropout method in which nodes are randomly eliminated may be applied to prevent overlearning.

[0073] In the following step S16, the learning unit 20E determines whether learning of all epochs has been completed. The total number of epochs is not particularly limited and can be set as appropriate. If it is determined that learning of all epochs has not been completed (NO), the learning unit 20E executes step S12 again to generate a subset of new combinations for each of the training data sets of the learning data sets 234A to 234D. Thereafter, steps S13 to S16 are executed. On the other hand, if it is determined that learning of all epochs has been completed (YES), the learning unit 20E stores the latest parameters in the storage unit 23X and sets these as parameters that define the trained models 231A to 231D. In other words, the trained models 231A to 231D are generated by the above procedure.

[0074] The learning unit 20E may input the first data of the test data in the learning data sets 234A-234D to each machine learning model at the appropriate time, calculate the error between each output data and the second data of each test data, and display the calculation result on the display unit 11X. In this way, if it is considered that the output error of the machine learning model has converged within a predetermined range before the completion of learning for all epochs, the learning may be terminated at that point.

[0075] <9. Features> According to the system 1 of the above embodiment, the timing at which the tire T will burst and whether or not damage has occurred in at least one of the tread portion, side portion, and bead portion are estimated based on data on physical quantities acquired while the vehicle 3 is traveling. If it is estimated that the tire T may burst while the vehicle 3 is traveling, or that damage has occurred (or will occur in the near future) in at least one of the tread portion, side portion, and bead portion, a notification of this fact is sent to the driver of the vehicle 3, etc. This allows the driver to take measures, such as stopping the vehicle 3 before the tire T bursts, such as adjusting the camber or checking the condition of the tire T. Furthermore, developers of the tire T can analyze the mechanisms of damage and burst of the tire T. Note that, while individual phenomena such as standing waves are known as signs of a tire burst, there is not just one type of phenomenon observed before a burst; for example, a burst may occur without the occurrence of standing waves. Furthermore, if the tire pressure is not properly adjusted, it is believed that a tire is more likely to burst, but it is difficult to estimate the occurrence of a burst or partial tire damage from air pressure data alone. In this regard, the estimation device 2 according to the above embodiment estimates a burst and partial tire damage based on a combination of multiple time-series data and taking into account multiple factors, which is expected to improve estimation accuracy.

[0076] <10. Variations> Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment, and various modifications are possible without departing from the spirit of the present invention. The gist of the following modifications can be combined as appropriate.

[0077] (1) In the above embodiment, NN-based models were used as the trained models 231A to 231D. However, the machine learning models are not limited to this. Other machine learning models, such as models using support vector machines (SVMs), convolutional neural network (CNN) models, K-NN models, clustering, k-means, decision trees, and logistic regression models, and models combining these may also be used. Furthermore, even in the case of NN-based models, the layer structure and the like may be changed as appropriate. The method of generating the trained models 231A to 231D is not particularly limited to the above embodiment, and known methods may be applied.

[0078] (2) In the above embodiment, the input data of the trained models 231A to 231D is a data set of continuous time-series data over a predetermined time range, which is created from a plurality of synchronized time-series data. However, the input data of the trained models 231A to 231D may be a data set synchronized at a certain time.

[0079] (3) In the above embodiment, the output data of the trained model 231A corresponds to the time until the tire T begins to burst. However, the output data of the trained model 231A may correspond to the burst occurrence probability at the time when the time series data based on the input data is acquired, similar to the output data of the trained models 231B to 231D in the above embodiment. In this case, the second data of the training dataset 234A can be created in the same manner as the second data of the training dataset 234B in the above embodiment. Furthermore, the time until the tire T begins to burst does not have to start from the time when the vehicle 3 starts traveling. For example, the start point may be the time when the vehicle 3 starts traveling at a predetermined speed or higher.

[0080] (4) Any of the trained models 231A to 231D of the estimation device 2 may be omitted. That is, the estimation device 2 may be configured to estimate any of the following: the timing at which a tire begins to burst; the presence or absence of damage to the tread portion at the time of acquiring the time-series data; the presence or absence of damage to the side portion at the time of acquiring the time-series data; or the presence or absence of damage to the bead portion at the time of acquiring the time-series data. Furthermore, the trained models 231B to 231D may be configured as a single integrated trained model. In this case, for example, three nodes corresponding to the tread portion, the side portion, and the bead portion may be created in the output layer of one trained model, and each node may output a value corresponding to the probability of damage occurring at the time of acquiring the time-series data. That is, the output data in this case corresponds to both the presence or absence of damage to the tire T at the time of acquiring the time-series data and the location of the damage occurring on the tire T.

[0081] (5) The input data to the trained models 231A to 231D may include at least data on vibrations occurring in the vehicle 3, data on the rotational speed or rotational acceleration of the tire T, and data on strain occurring in the tire T. In other words, at least one of the sensors 6 to 10 and 12 may be omitted. The input data to the trained models 231A to 231D may also include at least one of time-series data on the air pressure of the tire T, time-series data on the temperature of the tire T, road surface information representing the condition of the road surface on which the vehicle 3 is traveling (whether the road surface is dry or wet, whether the road surface is uneven or not, etc.), and meteorological information (weather, temperature, etc.) for the location where the vehicle 3 is currently located. The same applies to the training datasets 234A to 234D.

[0082] (6) The input data to the trained models 231A to 231D may include time-series data on the driving force of the vehicle 3 in addition to or instead of the time-series data on the physical quantities detected by the sensors 5 to 11. The time-series data on the driving force is, for example, time-series data on the engine torque or time-series data on the power consumption of the drive motor. The same applies to the training datasets 234A to 234D.

[0083] (7) In the above embodiment, the estimation device 2 and the generation device 2X are each configured as a single device. However, they may be configured as a single integrated device. Alternatively, the functions of the units 20A to 20C, 20D, and 20E, the memory unit 23, and the memory unit 23X may be distributed across multiple devices. For example, learning of a machine learning model may be performed using a service provided over a network, and the constructed trained models 231A to 231D may be stored in the estimation device 2. Furthermore, the trained models 231A to 231D may be constructed in a distributed manner across multiple devices. Furthermore, for example, at least a portion of the processing performed by the on-vehicle device 30a in the above embodiment may be performed by the computer 30b, the estimation device 2, or another computer.

[0084] (8) The control units 20 and 20X may be configured to include a CPU, a GPU, a vector processor, an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), other chips dedicated to artificial intelligence, etc. Furthermore, the operations of the control units 20 and 20X may be executed by one or more processors.

[0085] 1 System 2 Estimation device 2X generator 3 vehicles 5~11 sensors 231A~231D Trained Models

Claims

1. An estimation device that estimates damage to tires mounted on a vehicle based on data of physical quantities acquired while the vehicle is traveling, the device comprising: a derivation unit that inputs input data including at least data on vibrations generated in the vehicle due to rotation of the tire, data on the rotation speed or rotation acceleration of the tire, and data on strain generated in the tire while the vehicle is traveling into one or more trained machine learning models, and derives output data from the one or more trained machine learning models. Equipped with the output data corresponds to at least one of a timing at which damage occurs to the tire while the vehicle is traveling, whether or not the tire is damaged at the time when the data of the physical quantity is acquired, and a location at which damage occurs in the tire. Estimation device.

2. the input data further includes at least one of data on sounds generated due to rotation of the tire, data on external forces applied to the tire, data on the temperature of the tire, and data on the air pressure of the tire. The estimation device according to claim 1 .

3. The derivation unit inputs the input data into a plurality of trained machine learning models; each of the plurality of trained machine learning models is associated with each of a plurality of portions of the tire; the output data derived from each of the plurality of trained machine learning models corresponds to the presence or absence of damage to each of the plurality of parts at the time of acquiring the data of the physical quantities. The estimation device according to claim 1 or 2.

4. a notification generating unit that generates a notification to notify at least one of the timing at which damage will occur to the tire while the vehicle is traveling, the presence or absence of damage to the tire at the time of acquiring the data of the physical quantity, and a portion at which damage will occur in the tire; Further provided with The estimation device according to claim 1 or 2.

5. 1. An estimation method for estimating damage to tires mounted on a vehicle based on data of physical quantities acquired while the vehicle is traveling, the method being executed by one or more computers, the method comprising: inputting input data including at least data on vibrations occurring in the vehicle due to the rotation of the tire, data on the rotation speed or rotation acceleration of the tire, and data on strain occurring in the tire while the vehicle is traveling into one or more trained machine learning models, and deriving output data from the one or more trained machine learning models; Including, the output data corresponds to at least one of a timing at which damage occurs to the tire while the vehicle is traveling, whether or not the tire is damaged at the time when the data of the physical quantity is acquired, and a location at which damage occurs in the tire. Estimation method.

6. An estimation program for estimating damage to tires mounted on a vehicle based on data of physical quantities acquired while the vehicle is traveling, the program comprising: inputting input data including at least data on vibrations occurring in the vehicle due to the rotation of the tire, data on the rotation speed or rotation acceleration of the tire, and data on strain occurring in the tire while the vehicle is traveling into one or more trained machine learning models, and deriving output data from the one or more trained machine learning models; Execute the output data corresponds to at least one of a timing at which damage occurs to the tire while the vehicle is traveling, whether or not the tire is damaged at the time when the data of the physical quantity is acquired, and a location at which damage occurs in the tire. Estimation program.

7. an acquisition unit that acquires a learning dataset including a plurality of first data, the first data being data of physical quantities acquired while the vehicle is traveling, the first data including at least data of vibrations occurring in the vehicle due to the rotation of tires mounted on the vehicle, data of the rotation speed or rotation acceleration of the tires, and data of strain occurring in the tires while the vehicle is traveling, and second data that is combined with the first data in advance; a learning unit that inputs the first data into a machine learning model, derives output data from the machine learning model, and adjusts parameters that define the machine learning model so that an error between the output data and the second data is reduced; Equipped with the second data includes at least one of a timing when damage occurred to the tire while the vehicle was traveling, whether or not the tire was damaged at the time when the data of the physical quantity was acquired, and a location where the damage occurred in the tire. A device for generating trained machine learning models.

8. 1. A method for generating a trained machine learning model, executed by one or more computers, comprising: acquiring a learning dataset including a plurality of first data, the first data being data of physical quantities acquired while the vehicle is traveling, the first data including at least data of vibrations occurring in the vehicle due to rotation of tires mounted on the vehicle, data of rotational speed or rotational acceleration of the tires, and data of strain occurring in the tires while the vehicle is traveling, and second data combined with the first data in advance; inputting the first data into a machine learning model and deriving output data from the machine learning model; adjusting parameters that define the machine learning model so that an error between the output data and the second data is reduced; Including, the second data includes at least one of a timing when damage occurred to the tire while the vehicle was traveling, whether or not the tire was damaged at the time when the data of the physical quantity was acquired, and a location where the damage occurred in the tire. How to generate a trained machine learning model.

9. A program for generating a trained machine learning model, which is installed on one or more computers: acquiring a learning dataset including a plurality of first data, the first data being data of physical quantities acquired while the vehicle is traveling, the first data including at least data of vibrations occurring in the vehicle due to rotation of tires mounted on the vehicle, data of rotational speed or rotational acceleration of the tires, and data of strain occurring in the tires while the vehicle is traveling, and second data combined with the first data in advance; inputting the first data into a machine learning model and deriving output data from the machine learning model; adjusting parameters that define the machine learning model so that an error between the output data and the second data is reduced; Execute The second data includes at least one of the timing when damage occurred to the tire while the vehicle was traveling, whether or not the tire was damaged at the time the physical quantity data was acquired, and the location where the damage occurred in the tire.

Images