Tire physical information estimation system, computational model generation system, and tire physical information estimation method

Abstract

To provide a tire physical information estimation system, a calculation model generation system and a tire physical information estimation method which can construct a simplified calculation model to estimate physical information concerning a tire.SOLUTION: A tire physical information estimation system 100 comprises a physical information estimation part 32 and a data acquisition part 31. The physical information estimation part 32 has a learning-type calculation model 32a extending from an input layer to an output layer, which estimates tire physical information that is generated by motion of a tire 10. The data acquisition part 31 acquires data input to the input layer. The calculation model 32a has a feature extraction part that executes convolution calculation in which a filter is set in a first combination of the tire 10, a vehicle and a road surface condition, and a fully coupled part learnt in a second combination different from the first combination.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a tire physical information estimation system, a calculation model generation system, and a tire physical information estimation method. [Background technology]

[0002] Recently, research has been conducted into systems that input information measured on tires, vehicles, etc. into a learning-type calculation model to estimate tire physical information such as tire force.

[0003] Patent Document 1 describes a conventional tire physical information estimation system. The tire physical information estimation system includes a physical information estimation unit and a data acquisition unit. The physical information estimation unit has a learning-type computational model extending from an input layer to an output layer to estimate physical information related to a tire that is generated by tire motion. The data acquisition unit acquires input data for the input layer. The computational model has a feature extraction unit that extracts feature quantities by performing convolution computations during intermediate computations from the input layer to the output layer. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-46080 Summary of the Invention [Problem to be solved by the invention]

[0005] The tire physical information estimation system described in Patent Document 1 ensures real-time estimation of tire physical information. The inventors of the present application have considered the repurposing of a computational model for estimating tire physical information and have realized that there is room for improvement in terms of reducing the cost and time required to build the computational model.

[0006] The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a tire physical information estimation system, a computational model generation system, and a tire physical information estimation method that are capable of easily constructing a computational model and estimating physical information related to a tire. [Means for solving the problem]

[0007] One aspect of the present invention is a tire physical information estimation system, which includes a learning-type computational model extending from an input layer to an output layer, a physical information estimation unit that estimates tire physical information generated by tire motion, and a data acquisition unit that acquires input data for the input layer, and the computational model includes a feature extraction unit that performs a convolution operation in which a filter is set for a first combination of tire, vehicle, and road surface conditions, and a full connection unit that has been trained for a second combination different from the first combination.

[0008] Another aspect of the present invention is a computational model generation system, which has a learning-type computational model extending from an input layer to an output layer, and includes: a physical information estimation unit that estimates tire physical information generated by tire motion; a data acquisition unit that acquires input data to the input layer; and a learning processing unit that trains the computational model using the tire physical information measured at the tire as training data, wherein the computational model includes a feature extraction unit that executes a convolution operation in which a filter is set for a first combination of tire, vehicle, and road surface conditions, and a full connection unit that trains the computational model for a second combination different from the first combination.

[0009] Another aspect of the present invention is a tire physical information estimation method, which includes a learning-type computational model extending from an input layer to an output layer, and includes a physical information estimation step of estimating tire physical information generated by tire motion, and a data acquisition step of acquiring input data for the input layer, wherein the computational model includes a feature extraction unit that performs a convolution operation in which a filter is set for a first combination of tire, vehicle, and road surface conditions, and a full connection unit that has been trained for a second combination different from the first combination. [Effects of the Invention]

[0010] According to the present invention, it is possible to easily construct a calculation model and estimate physical information about a tire. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a schematic diagram for explaining an overview of a tire physical information estimation system according to an embodiment; [Figure 2] 1 is a block diagram showing a functional configuration of a tire physical information estimation system according to an embodiment. [Figure 3] FIG. 2 is a schematic diagram showing the configuration of a calculation model. [Figure 4] FIG. 1 is a block diagram showing a functional configuration of a computation model generation system. [Figure 5] 4 is a flowchart showing a procedure of a tire physical information estimation process performed by the tire physical information estimation device. [Figure 6] 10 is a graph showing an example of a verification result of transfer learning when a vehicle is changed. [Figure 7] 10 is a graph showing an example of a verification result of transfer learning with different tire and road surface conditions. [Figure 8] FIG. 10 is a block diagram showing a functional configuration of a tire physical information estimation system according to a modified example. DETAILED DESCRIPTION OF THE INVENTION

[0012] The present invention will be described below based on preferred embodiments with reference to Figures 1 to 8. The same or equivalent components and members shown in each drawing are designated by the same reference numerals, and duplicate descriptions will be omitted where appropriate. The dimensions of the members in each drawing are enlarged or reduced as appropriate to facilitate understanding. Some members that are not important for explaining the embodiments will be omitted from the drawings.

[0013] (Embodiment) 1 is a schematic diagram illustrating an overview of a tire physical information estimation system 100 according to an embodiment. The tire physical information estimation system 100 includes a sensor 20 disposed on a tire 10 and a tire physical information estimation device 30. The tire physical information estimation system 100 may also include a server device 40 that acquires and stores tire physical information, such as tire force F, road friction coefficient, and moments acting on the tire 10 about three axes, estimated by the tire physical information estimation device 30 via a communication network 91 and monitors the tire physical information.

[0014] The sensor 20 measures physical quantities of the tire 10, such as the acceleration and strain of the tire 10, the tire air pressure, and the tire temperature, and outputs the measured data to the tire physical information estimation device 30. The tire physical information estimation device 30 applies a computation model trained on one combination (hereinafter referred to as a first combination) of the tire 10, the vehicle, and the road surface condition to a combination (hereinafter referred to as a second combination) different from the first combination, and estimates the tire physical information based on the data measured by the sensor 20. The tire physical information estimation device 30 uses the data measured by the sensor 20 in the computation to estimate the tire physical information, but may also obtain information from the vehicle side, such as vehicle acceleration, from a vehicle control device 90 or the like and use the information in the computation to estimate the tire physical information.

[0015] The tire physical information estimation device 30 outputs tire physical information such as the estimated tire force F, road friction coefficient, and moments acting on the tire 10 about three axes to, for example, a vehicle control device 90. The vehicle control device 90 uses the tire physical information input from the tire physical information estimation device 30, for example, to estimate braking distance, apply it to vehicle control, and further notify the driver of information related to safe vehicle driving. The vehicle control device 90 can also provide information related to safe vehicle driving in the future using map information, weather information, etc. Furthermore, when the vehicle control device 90 has a function for automatically driving the vehicle, the tire physical information estimation system 100 provides the estimated tire physical information to the vehicle control device 90 as data to be used for vehicle speed control and the like during automatic driving.

[0016] The server device 40 acquires, from the tire physical information estimation device 30, the physical quantities of the tire 10 measured by the sensor 20, and tire physical information such as the tire force F and road friction coefficient estimated for the tire 10. The server device 40 may accumulate the physical quantities measured in the tire 10 and the tire physical information estimated by the tire physical information estimation device 30 from a plurality of vehicles.

[0017] 2 is a block diagram showing the functional configuration of a tire physical information estimation system 100 according to an embodiment. The sensor 20 of the tire physical information estimation system 100 includes an acceleration sensor 21, a strain gauge 22, a pressure gauge 23, a temperature sensor 24, etc., and measures physical quantities of the tire 10. These sensors measure physical quantities related to the deformation and movement of the tire 10 as the physical quantities of the tire 10.

[0018] The acceleration sensor 21 and the strain gauge 22 move mechanically together with the tire 10, and measure the acceleration and strain amount, respectively, occurring in the tire 10. The acceleration sensor 21 is disposed, for example, in the tread, side, bead, wheel, etc. of the tire 10, and measures the acceleration in three axes of the tire 10: the circumferential direction, the axial direction, and the radial direction.

[0019] The strain gauges 22 are disposed in the tread, sides, beads, etc. of the tire 10 and measure strain at the locations where they are disposed. The pressure gauge 23 and temperature sensor 24 are disposed, for example, in the air valve of the tire 10 and measure the tire pressure and tire temperature, respectively. The temperature sensor 24 may be disposed directly on the tire 10 to accurately measure the temperature of the tire 10. The tires 10 may be fitted with, for example, an RFID 11 or the like that has unique identification information attached thereto in order to identify each tire.

[0020] The tire physical information estimation device 30 includes a data acquisition unit 31, a physical information estimation unit 32, and a communication unit 33. The tire physical information estimation device 30 is an information processing device such as a PC (personal computer). Each unit in the tire physical information estimation device 30 can be realized in terms of hardware using electronic elements and mechanical parts such as a computer CPU, and in terms of software using a computer program, but the functional blocks realized by the cooperation of these elements are depicted here. Therefore, it will be understood by those skilled in the art that these functional blocks can be realized in various forms by combining hardware and software.

[0021] The data acquisition unit 31 acquires information on acceleration, strain, air pressure, and temperature measured by the sensor 20 via wireless communication or the like. The communication unit 33 communicates with external devices such as the vehicle control device 90 and the server device 40 via wired or wireless communication or the like. The communication unit 33 transmits the physical quantities of the tire 10 measured by the sensor 20, tire physical information estimated about the tire 10, and the like to the external devices via a communication line, for example, a CAN (Control Area Network), the Internet, or the like.

[0022] The physical information estimation unit 32 has a calculation model 32a, and inputs information from the data acquisition unit 31 to the calculation model 32a to estimate tire physical information such as tire force F, road friction coefficient, and moments about three axes acting on the tire 10. As shown in FIG. 2, the tire force F has three axial components: a longitudinal force Fx in the longitudinal direction of the tire 10, a lateral force Fy in the lateral direction, and a load Fz in the vertical direction. The physical information estimation unit 32 may calculate all of these three axial components, or may calculate at least one of the components or two components in any combination.

[0023] The computational model 32a uses a learning model such as a neural network. Fig. 3 is a schematic diagram showing the configuration of the computational model 32a. The computational model 32a is a convolutional neural network (CNN) type, and is a learning model that includes convolution operations and pooling operations used in its prototype, so-called DenseNet. Fig. 3 shows an example in which acceleration data in three axial directions is used as input data to the computational model 32a, and tire forces Fx, Fy, and Fz in three axial directions are output.

[0024] The computational model 32a includes an input layer 50, a stem block 51, a feature extraction unit 52, a middle layer 53, a full connection unit 54, and an output layer 55. Time-series data of acceleration in three axial directions acquired by the data acquisition unit 31 is input to the input layer 50. The acceleration data is measured in time series by the sensor 20, and data within a certain time interval is extracted using a window function to be used as input data. The input data may be, for example, 128 pieces of acceleration data included in a certain time interval along each axial direction.

[0025] The acceleration measured at the tire 10 is periodic for each rotation of the tire 10. The time interval of the input data extracted by the window function may be, for example, a time period corresponding to the rotation period of the tire 10, so that the input data itself has periodicity. Note that the window function may extract input data in a time interval shorter or longer than one rotation of the tire 10, and as long as the extracted input data contains at least periodic information, learning of the calculation model 32a is possible.

[0026] The Stem block 51 reduces the size of the input data while maintaining the characteristics of the input data. The Stem block 51 performs, for example, a convolution operation or a pooling operation.

[0027] The feature extraction unit 52 repeats the Dense block 52a, the convolution operation 52b, and the pooling operation 52c multiple times (denoted as n times in FIG. 3) to extract features and transmit them to each node in the intermediate layer 53. In the example of the feature extraction unit 52 shown in FIG. 3, the Dense block 52a performs a convolution operation with an appropriate filter length on input data, and then performs the subsequent convolution operation 52b and pooling operation 52c.

[0028] The convolution operation in the dense block 52a uses a filter of length 1 to 5 as appropriate, and multiplies and adds the input data arranged in time series while moving the filter. The convolution operation multiplies consecutive data of the filter length (e.g., A1, A2, A3) in the time series input data by each value in the filter (f1, f2, f3), and then adds the values ​​obtained by multiplication to obtain A1×f1+A2×f2+A3×f3. Note that the convolution operation may be performed by zero-padding, which adds "0" data to the end of the input data. Furthermore, the amount of filter movement in the convolution operation is usually set to one input data, but this can be changed as appropriate to reduce the size of the operation model 32a.

[0029] The convolution operation 52b performs a convolution operation with a filter length of 1. The pooling operation 52c performs an average value pooling operation on the data after the convolution operation 52b. The pooling operation 52c calculates the average value of two values ​​arranged in time series, for example. The feature extraction unit 52 repeats the operations by the Dense block 52a, the convolution operation 52b, and the pooling operation 52c multiple times, and outputs 64 pieces of data as a result to the intermediate layer 53.

[0030] The full connection unit 54 fully connects data from each node in the intermediate layer 53 at multiple hierarchical levels, and outputs tire forces Fx, Fy, and Fz to each node in the output layer 55. The full connection unit 54 performs calculations using a full connection path that executes linear calculations using weighting, etc., but may also execute nonlinear calculations using an activation function, etc., in addition to linear calculations.

[0031] Each node of the output layer 55 may output tire physical information such as the tire forces in the three axial directions, the road friction coefficient, and the moments about three axes acting on the tire 10. The output layer 55 may output one type of tire physical information or multiple types of tire physical information in any combination among the tire forces in the three axial directions, the road friction coefficient, and the moments about three axes acting on the tire 10.

[0032] In addition, when estimating the road surface friction coefficient, the output layer 55 may output an estimated value of the road surface friction coefficient, or may classify the road surface friction coefficient into categories such as dry, wet, snowy, or icy conditions and output which category it falls into.

[0033] 4 is a block diagram showing the functional configuration of the computational model generation system 110. The computational model generation system 110 includes a tire physical information measurement device 60 and a computational model generation device 70 having a learning processing unit 71. The computational model generation device 70 has the learning processing unit 71 in addition to the components of the tire physical information estimation device 30. The components of the computational model generation device 70 that correspond to the components of the tire physical information estimation device 30 have the same functions as those of the tire physical information estimation device 30, but the computational model 32a is either pre-learning or currently being learned.

[0034] The tire physical information measurement device 60 measures tire physical information such as the tire force F in three axial directions, the road friction coefficient, and the moment about three axes acting on the tire 10. The learning processing unit 71 uses the tire physical information measured by the tire physical information measurement device 60 as training data to train the calculation model 32a. In the learning process of the calculation model 32a, the tire physical information is estimated by the calculation model 32a based on input information and compared with the training data.

[0035] The learning processing unit 71 compares the tire physical information estimated by the computation model 32a with the teacher data, newly sets various coefficients in the computation process, such as weighting, to the computation model 32a, and repeatedly updates the model to perform learning. The tire physical information estimation system 100 estimates tire physical information using the computation model 32a that has been trained by the computation model generation system 110.

[0036] The computational model 32a is trained on a first combination of the tire 10, the vehicle, and the road surface condition. Here, the tire 10 has a certain specification, and the vehicle has a certain model name. The road surface condition is, for example, dry, wet, snowy, or icy. The specifications of the tire 10 include information on tire performance, such as tire size, tire width, aspect ratio, tire strength, tire outer diameter, load index, and manufacturing date.

[0037] After completing learning for the first combination, the computation model 32a executes learning for a second combination in which two of the tire 10, vehicle, and road surface conditions are the same as those in the first combination and one is different from those in the first combination. Upon completion of learning for the first combination, the filter length and various parameters in the Stem block 51 and feature extraction unit 52 of the computation model 32a are set.

[0038] The computational model 32a of the second combination uses the Stem block 51 and feature extraction unit 52 in the trained first combination to learn the full connection unit 54. In training the computational model 32a of the second combination, the Stem block 51 and feature extraction unit 52 in the trained first combination are reused to train the full connection unit 54, thereby reducing the cost and time of training and simplifying the construction of the computational model.

[0039] The second combination is a tire with different tire 10 specifications, a vehicle with a different vehicle name, or a road surface condition different from that of the first combination. Similarly, a computation model of a third combination, in which any one of the tire 10, the vehicle, and the road surface condition is different from that of the second combination, may be subjected to transfer learning based on the computation model of the second combination that has already been trained.

[0040] Furthermore, a computation model for a fourth combination, in which any one of the tire 10, the vehicle, and the road surface condition is different from that of the third combination, may be subjected to transfer learning based on the learned computation model for the third combination. By repeating such transfer learning, it is possible to learn a computation model for a combination in which all of the tire 10, the vehicle, and the road surface condition are different from those of the first combination.

[0041] Furthermore, after completing learning for the first combination, the computation model 32a may perform transfer learning based on the first combination for a second combination in which one of the tire 10, vehicle, and road surface conditions is the same as the first combination and two are different from the first combination. In this case, it is considered that the estimation accuracy of the tire physical information tends to be lower than when transfer learning is performed for the second combination in which one of the tire 10, vehicle, and road surface conditions is different from the first combination.

[0042] The learning of the calculation model 32a based on the first combination may be performed, for example, at a test site where a vehicle is driven (see FIG. 3). The learning of the calculation model 32a based on the second combination may be performed at the same test site with different tires 10, vehicles, and road surface conditions, or may be performed on a public road other than a test site.

[0043] With regard to the tire 10, it is not necessary to execute learning of the calculation model 32a for each strict specification, and the calculation model 32a may be trained for each type, such as a passenger car tire, a truck tire, etc. Also, with regard to the vehicle, it is not necessary to execute learning of the calculation model 32a for each strict vehicle name and engine type, and the calculation model 32a may be trained for each type, for example, by setting a type based on commonalities such as a drive system and a vehicle weight.

[0044] Next, a description will be given of the operation of the tire physical information estimation system 100. Fig. 5 is a flowchart showing the procedure of tire physical information estimation processing by the tire physical information estimation device 30. The following describes processing for estimating tire physical information using the calculation model 32a of the second combination, after learning has been completed for the tire 10, vehicle, and road surface conditions using the first and second combinations.

[0045] The tire physical information estimation device 30 acquires physical quantities such as acceleration, strain, tire pressure, and tire temperature of the tire 10 measured by the sensor 20 using the data acquisition unit 31 (S1). The physical information estimation unit 32 extracts input data for a certain time interval from the data acquired by the data acquisition unit 31 (S2).

[0046] In estimating tire physical information, acceleration data for at least one axis (for example, the circumferential direction) is required as input data. In addition, in estimating tire physical information, acceleration data for two axes, for example, the circumferential direction and the axial direction of the tire 10, may be used as input data, or acceleration data for three axes may be used as input data. Furthermore, time-series data for at least one of strain in the tire 10, tire pressure, and tire temperature may be included in the input data.

[0047] The computation model 32a performs computation in the Stem block 51 based on the input data extracted in step S2, and inputs the results to the feature extraction unit 52. The feature extraction unit 52 executes processing to extract features from the input data using the Dense block 52a, convolution computation 52b, and pooling computation 52c (S3). As described above, the Stem block 51 and feature extraction unit 52 of the computation model 32a have been diverted from what they learned for the first combination to the second combination.

[0048] The fully connected unit 54 of the computational model 32a estimates tire physical information by performing a fully connected calculation on the feature amounts extracted by the feature extraction unit 52 and input to each node in the intermediate layer 53 (S4), and then ends the processing. As described above, the fully connected unit 54 of the computational model 32a is trained in the second combination, which is a repurposed version of the Stem block 51 and feature extraction unit 52 of the first combination. Parameters such as weighting used in the fully connected calculation of the fully connected unit 54 are determined in the training of the computational model 32a of the second combination. Tire physical information such as tire force F, road friction coefficient, and moments acting on the tire 10 about three axes is output to each node in the output layer 55 by the fully connected calculation.

[0049] Fig. 6 is a graph showing an example of the verification results of transfer learning when changing vehicles. In Fig. 6, the RMSE (root mean square error) between the estimated values ​​and the measured values ​​of tire forces Fx, Fy, and Fz is calculated for an example of adaptation to another vehicle as Comparative Example 1, an example of learning using another vehicle as Comparative Example 2, and an example of transfer learning in this embodiment, and the ratio (%) of the other examples is shown, with the RMSE for Comparative Example 1 being 100%.

[0050] In Comparative Example 1, a computational model trained on a first combination of summer tires with a certain specification, a vehicle with model name A1, and normal dry road conditions is directly applied to a second combination of the same summer tires, a vehicle with model name A2, and the same dry road conditions. Comparative Example 2 uses a computational model trained on the second combination regardless of the first combination. In the transfer learning of this embodiment, a computational model trained on the second combination is used, based on the computational model trained on the first combination.

[0051] 6, the example of transfer learning in this embodiment has roughly the same RMSE as Comparative Example 2, which was trained with the second combination, and it is clear that transfer learning has improved the accuracy of tire force estimation. On the other hand, in Comparative Example 1, the calculation model trained with the first combination is directly used for the second combination to estimate tire force, and it is clear that the estimation accuracy is degraded compared to the example of transfer learning in this embodiment.

[0052] Fig. 7 is a graph showing an example of the verification results of transfer learning with different tire and road surface conditions. In Fig. 7, the RMSEs of the estimated and measured values ​​of tire forces Fx, Fy, and Fz are calculated for Comparative Example 3 (reference estimation), Comparative Example 4 (repurposed estimation), and an example of transfer learning in this embodiment, and the ratios (%) of the other examples are shown, with the RMSE for Comparative Example 3 being set to 100%.

[0053] In Comparative Example 3, tire force is estimated using a calculation model trained with a first combination of summer tires with certain specifications, a vehicle with model name A1, and a summer season in region B1 where road surfaces are often dry. In Comparative Example 4, the calculation model trained in Comparative Example 3 is used as is for a second combination of winter tires, a vehicle with the same model name A1, and a winter season in region B2 where snow and ice are often present, to estimate tire force. In the transfer learning of this embodiment, a calculation model trained with the second combination is used, which is based on the calculation model trained with the first combination.

[0054] 7, the example of transfer learning in this embodiment has roughly the same RMSE as Comparative Example 3 (reference estimation) trained with the first combination, and it is clear that transfer learning improves the estimation accuracy of tire force. On the other hand, in Comparative Example 4, the calculation model trained with the first combination is directly used for the second combination to estimate tire force, and it is clear that the estimation accuracy is degraded compared to the example of transfer learning in this embodiment.

[0055] The tire physical information estimation system 100 includes a calculation model 32a having a feature extraction unit 52 and a full connection unit 54 that perform convolution operations, and estimates tire physical information such as tire force F generated by the motion of the tire 10. The feature extraction unit 52 performs convolution operations in which a filter is set for a first combination of the tire 10, the vehicle, and the road surface condition. The full connection unit 54 calculates the tire physical information by full connection operations learned for a second combination different from the first combination.

[0056] The tire physical information estimation system 100 estimates tire physical information for the second combination by transfer learning from a computational model trained using the first combination, thereby reducing the cost and time required for learning for the second combination and easily constructing a computational model to estimate physical information about the tire. Furthermore, the computational model generation system 110 generates a computational model for the second combination by transfer learning from a computational model trained using the first combination, thereby easily generating a computational model.

[0057] The tire physical information estimation system 100 can easily construct a calculation model 32a for the different tire by including a tire in the second combination that is different from the tire in the first combination. The tire physical information estimation system 100 can easily construct a calculation model 32a for the different road surface condition by including a road surface condition in the second combination that is different from the road surface condition in the first combination.

[0058] The tire physical information estimation system 100 can easily construct a calculation model 32a for a vehicle with a different vehicle name from the vehicle in the first combination by including the vehicle with the different vehicle name in the second combination.

[0059] The tire physical information estimation system 100 acquires acceleration data measured in the tire 10 using the data acquisition unit 31, inputs the acceleration data to the calculation model 32a, and outputs the tire force F. In this way, the tire physical information estimation system 100 can provide information necessary for analyzing behavior of the tire 10, such as slippage.

[0060] (Variation) Fig. 8 is a block diagram showing a functional configuration of a tire physical information estimation system 100 according to a modified example. In the modified example shown in Fig. 8, data to be input to the calculation model 32a is acquired from a vehicle control device 90. Note that both the data from the vehicle control device 90 and the data from the sensor 20 (see Fig. 2) can also be used as data to be input to the calculation model 32a.

[0061] The vehicle control device 90 acquires driving data such as the vehicle's driving speed, acceleration in three axial directions, and three-axial angular velocity from a digital tachometer or the like of the vehicle, as well as load data such as the vehicle weight and axle loads on the axles. The vehicle control device 90 outputs the driving data and load data to the tire physical information estimation device 30.

[0062] The tire physical information estimation device 30 estimates tire physical information such as the tire force F, the road friction coefficient, and moments around three axes acting on the tire 10 using a calculation model 32a based on data input from a vehicle control device 90. The calculation model 32a is constructed by performing learning in advance to estimate tire physical information based on data input from the vehicle control device 90, for example, through test driving using an actual vehicle.

[0063] In the above-described embodiment and modified examples, tire physical information estimated by the computational model 32a has been described as including tire force F, road friction coefficient, and moments acting on the tire 10 about three axes. However, it is also possible to estimate, for example, loosening of fastening parts such as wheel nuts used to mount the tire 10. Since vibrations due to loosening of fastening parts such as wheel nuts appear in the acceleration data measured on the tire 10, a computational model 32a that estimates loosening of fastening parts by comparison with other tire forces F is constructed and trained in a CNN format. The tire physical information estimation system 100 can estimate loosening of fastening parts of the tire 10 by performing calculations using the computational model 32a based on input data such as acceleration data acquired during actual vehicle travel.

[0064] 1, the sensor 20 may be, for example, a microphone provided on the tire 10 or in the vicinity of the tire 10. The calculation model 32a may estimate the tire physical information using audio data collected by a microphone or the like.

[0065] In the above-described embodiment and modified examples, the computational model 32a uses a CNN-type DenceNet model, but it is also possible to use a model structure such as a so-called LeNet model, ResNet model, MobileNet model, or PeleelNet model. Also, a model may be constructed by incorporating a modular structure such as a Residual Block into the computational model 32a.

[0066] Next, features of the tire physical information estimation system 100, the calculation model generation system 110, and the tire physical information estimation method according to the embodiment will be described. A tire physical information estimation system 100 according to the embodiment includes a physical information estimation unit 32 and a data acquisition unit 31. The physical information estimation unit 32 has a learning-type computational model 32a extending from an input layer 50 to an output layer 55, and estimates tire physical information resulting from the motion of the tire 10. The data acquisition unit 31 acquires input data for the input layer 50. The computational model 32a has a feature extraction unit 52 that performs a convolution operation in which a filter is set for a first combination of the tire 10, the vehicle, and road surface conditions, and a full connection unit 54 that has trained for a second combination different from the first combination. This enables the tire physical information estimation system 100 to simply construct the computational model 32a and estimate physical information related to the tire 10.

[0067] The second combination also includes tires different from the tires in the first combination, which allows the tire physical information estimation system 100 to easily construct a calculation model 32a for the different tires.

[0068] The second combination includes a road surface condition different from the road surface condition in the first combination, which allows the tire physical information estimation system 100 to easily construct a calculation model 32a for different road surface conditions.

[0069] The second combination also includes vehicles with names different from those of the vehicles in the first combination, which allows the tire physical information estimation system 100 to easily construct a calculation model 32a for vehicles with different names.

[0070] The data acquisition unit 31 also acquires acceleration data measured in the tire 10 . The computation model 32a receives acceleration data as input to the input layer 50 and outputs a tire force F from the output layer 55. This enables the tire physical information estimation system 100 to provide information necessary for analyzing behavior of the tire 10, such as slippage.

[0071] The computational model generation system 110 includes a physical information estimation unit 32, a data acquisition unit 31, and a learning processing unit 71. The physical information estimation unit 32 has a learning-type computational model 32a extending from an input layer 50 to an output layer 55, and estimates tire physical information generated by the motion of the tire 10. The data acquisition unit 31 acquires input data for the input layer 50. The learning processing unit 71 trains the computational model using tire physical information measured at the tire 10 as training data. The computational model 32a includes a feature extraction unit 52 that executes a convolution operation in which a filter is set for a first combination of the tire 10, the vehicle, and road surface conditions, and a full connection unit 54 that trains the computational model for a second combination different from the first combination. As a result, the computational model generation system 110 generates a computational model for the second combination by transfer learning from the computational model trained using the first combination, thereby enabling simple generation of a computational model.

[0072] The tire physical information estimation method includes a physical information estimation step and a data acquisition step. The physical information estimation step includes a learning-type computational model 32a extending from an input layer 50 to an output layer 55, and estimates tire physical information resulting from the motion of the tire 10. The data acquisition step acquires input data for the input layer 50. The computational model 32a includes a feature extraction unit 52 that performs a convolution operation with a filter set for a first combination of the tire 10, the vehicle, and road surface conditions, and a full connection unit 54 that has trained for a second combination different from the first combination. According to this tire physical information estimation method, the computational model 32a can be easily constructed to estimate physical information about the tire 10.

[0073] The present invention has been described above based on the embodiments. These embodiments are merely examples, and it will be understood by those skilled in the art that various modifications and changes are possible within the scope of the claims of the present invention, and that such modifications and changes also fall within the scope of the claims of the present invention. Therefore, the descriptions and drawings in this specification should be treated as illustrative rather than restrictive. [Explanation of symbols]

[0074] 10 tire, 32 physical information estimation unit, 32a calculation model, 50 input layer, 52 feature extraction unit, 55 output layer, 71 learning processing unit, 100 Tire physical information estimation system, 110 Computational model generation system.

Claims

1. a physical information estimation unit having a learning-type calculation model extending from an input layer to an output layer, and estimating tire physical information generated by tire motion; a data acquisition unit that acquires input data for the input layer, The calculation model includes a feature extraction unit that performs a convolution operation in which a trained filter is set for a first combination of a tire, a vehicle, and a road surface condition, and a full connection unit that is trained for a second combination different from the first combination, and the trained filter is reused and the full connection unit is trained for the second combination.

2. 2. The tire physical information estimation system according to claim 1, wherein the second combination includes tires different from the tires in the first combination.

3. 2. The tire physical information estimation system according to claim 1, wherein the second combination includes a road surface condition different from the road surface condition in the first combination.

4. 2. The tire physical information estimation system according to claim 1, wherein the second combination includes vehicles with vehicle names different from those of the vehicles in the first combination.

5. the data acquisition unit acquires acceleration data measured at a tire, The tire physical information estimation system according to claim 1 , wherein the calculation model receives the acceleration data in the input layer and outputs a tire force from the output layer.

6. a physical information estimation unit having a learning-type calculation model extending from an input layer to an output layer, and estimating tire physical information generated by tire motion; a data acquisition unit that acquires input data for the input layer; a learning processing unit that learns the calculation model using the tire physical information measured on the tire as training data, The computational model has a feature extraction unit that performs a convolution operation in which a trained filter is set for a first combination of tire, vehicle, and road surface condition, and a full connection unit that is trained for a second combination different from the first combination, and the trained filter is reused to train the full connection unit for the second combination.

7. a physical information estimation step having a learning-type computation model extending from an input layer to an output layer, and estimating tire physical information generated by tire motion; a data acquisition step of acquiring input data for the input layer, the computational model includes a feature extraction unit that performs a convolution operation in which a trained filter is set for a first combination of a tire, a vehicle, and a road surface condition, and a full connection unit that is trained for a second combination different from the first combination, and the trained filter is reused and the full connection unit is trained for the second combination.

Images