Suspension control method and suspension control system

Abstract

The purpose of the present invention is to provide a suspension control device capable of performing optimum suspension control in vehicles of various types and specifications. A suspension control device according to the present invention acquires a parameter indicating correspondence between behavior information of a vehicle and sensor data of a suspension, estimates the sensor data on the basis of the parameter and the behavior information of the vehicle, and controls the suspension using the sensor data. The suspension control device is characterized by including an applicability determination unit 305 that compares a newly acquired new parameter with an acquired old parameter to determine whether the new parameter is applicable to the vehicle.

Description

Suspension control device and suspension control system

[0001] The present invention relates to a suspension control device and a suspension control system.

[0002] There is known a suspension control device that controls the attitude of a vehicle by changing the firmness and characteristics of the suspension according to road surface conditions and driving conditions. There is a type of suspension control device that detects sensor data from sensors that detect the piston speed and sprung mass speed of the vehicle's suspension and controls the suspension.

[0003] And in Patent Document 1, there is described a technique of a suspension control device that estimates sensor data related to a suspension based on vehicle behavior information and parameters determined by machine learning, and controls the suspension without using sensors based on the estimated sensor data.

[0004] Japanese Unexamined Patent Application Publication No. 2019-189228

[0005] According to the technique of Patent Document 1, optimal suspension control is performed if the parameters determined by machine learning are parameters determined based on vehicle behavior information and sensor data of a vehicle that matches the vehicle type and specifications.

[0006] However, when the parameters are not those determined based on vehicle behavior information and sensor data of a vehicle that matches the vehicle type and specifications, but are parameters estimated based on vehicle behavior information and sensor data of a vehicle with a different vehicle type and specifications, optimal suspension control is not always performed in that vehicle, and there is a possibility that it becomes difficult to improve the performance of the suspension.

[0007] The present invention has been made in view of the above problems, and an object thereof is to provide a suspension control device capable of performing optimal suspension control for each of various vehicle types and specifications.

[0008] The suspension control device of the present invention, which solves the above problems, is a suspension control device that acquires parameters indicating the correspondence between vehicle behavior information and suspension sensor data, estimates the sensor data based on the parameters and vehicle behavior information, and controls the suspension using the sensor data, and is characterized by having an applicability determination unit that compares newly acquired new parameters with previously acquired old parameters to determine whether the new parameters are applicable to the vehicle.

[0009] According to the present invention, it is possible to provide a suspension control device that enables optimal suspension control for various types and specifications of vehicles.

[0010] Further features related to the present invention will become apparent from the description herein and the accompanying drawings. Problems, configurations, and effects not described above will be revealed by the following description of embodiments.

[0011] A functional block diagram of the suspension control system in the first embodiment. A diagram showing an example of the vehicle state estimation unit of an AI-applied vehicle. A graph showing an example of time-series data of physical values ​​arriving via CAN and time-series data of estimated instantaneous values ​​of on-board sensor data. A diagram showing the block configuration of the data management unit included in the AI-applied vehicle. A diagram showing the block configuration of the management unit in the server of the first embodiment. A diagram showing an example of a driving log recorded in the log data storage unit of the first embodiment. A flowchart of the weight data distribution process in the server of the first embodiment. A flowchart of the applicability determination process in the AI-applied vehicle of the first embodiment. A functional block diagram of the suspension control system in the second embodiment. A flowchart of the ride comfort determination process in the second embodiment. A functional block diagram showing the suspension control system in the third embodiment.

[0012] Next, an embodiment of a suspension control system to which the suspension control device of the present invention is applied will be described in detail with reference to the drawings.

[0013] [First Embodiment] Figure 1 is a diagram showing the configuration of a suspension control system in the first embodiment. The suspension control system 100 includes a mass-produced vehicle 101, a learning management server 102, and an AI-applied vehicle 103. Figure 1 shows a system configuration in which there is one mass-produced vehicle 101, one learning management server 102, and one AI-applied vehicle 103, but a configuration in which there are multiple units of at least one of them is also possible.

[0014] The mass-produced vehicle 101 and the AI-applied vehicle 103 are four-wheeled automobiles, including, for example, gasoline engine vehicles, diesel engine vehicles, hybrid vehicles, electric vehicles, and fuel cell vehicles, and are equipped with an active suspension (hereinafter simply referred to as suspension) that controls the vehicle's posture by changing spring stiffness, damper characteristics, etc., according to road conditions and driving conditions.

[0015] The mass-produced vehicle 101, the learning management server 102, and the AI-applied vehicle 103 each have an ECU (Electronic Control Unit) as a control device. The ECU is mainly composed of a microcomputer consisting of a CPU, ROM, RAM, etc. The CPU is a computer that executes various programs installed in the storage device. The memory functions as the main memory unit that stores various programs, data, etc. necessary for the CPU to execute the various programs stored in the storage device. The storage device stores various programs and data. The various internal functions described later are realized by the CPU loading the programs stored in the storage device into the memory and executing them.

[0016] [Regarding mass-produced vehicles] The control device of the mass-produced vehicle 101 has as internal functions a vehicle state setting / detection unit 104, a vehicle state measurement unit 105, a suspension control unit 106, a driving data management unit 107, and a transmission unit 108.

[0017] The mass-produced vehicle 101 is equipped with suspension-specific sensors such as acceleration sensors to detect sensor data such as the vertical speed of the sprung mass (hereinafter referred to as sprung mass speed) and the relative speed of the piston stroke (hereinafter referred to as piston speed), and controls the suspension based on that sensor data. The mass-produced vehicle 101 is equipped with a dedicated sensor for controlling the damping force of the suspension, which is the vehicle state measurement unit 105 in Figure 1, and the vehicle state measurement unit 105 transfers physical values ​​such as sprung mass speed and piston speed acquired by the vehicle state measurement unit 105 to the suspension control unit 106. The suspension control unit 106 calculates command values ​​for suspension control based on the physical values, and although not shown in this figure, it determines the damping force of the suspension installed on each of the four wheels and controls the suspension.

[0018] The sensor data acquired by the vehicle condition measurement unit 105 may include not only data relating to the behavior of the mass-produced vehicle 101, but also sensor data from the chassis of the mass-produced vehicle 101, as well as other sensor data such as vehicle speed, GPS, and camera data. In particular, it is preferable that the sensor data from the chassis includes all or part of the acceleration data relating to the path from the road surface stimulus input via the tires to the occupant. The vehicle condition measurement unit 105 can acquire as sensor data, for example, acceleration of the lower and upper springs that constitute the suspension, the area around the seat where the occupant sits, the steering tie rod, the steering, etc.

[0019] The vehicle state setting / detection unit 104 sets, determines, and derives the vehicle state of the mass-produced vehicle 101. The driving data management unit 107 collects a dataset consisting of information on the vehicle state of the mass-produced vehicle 101, vehicle behavior information transmitted over the CAN while the mass-produced vehicle 101 is in motion, and physical values ​​measured by the vehicle state measurement unit 105, i.e., sprung mass velocity, piston velocity, etc. The transmission unit 108 transfers the dataset collected by the driving data management unit 107 to the learning management server 102. The vehicle behavior information transmitted over the CAN includes information such as the pitch rate, roll rate, and yaw rate of the mass-produced vehicle 101.

[0020] [About the AI-equipped vehicle] The AI-equipped vehicle 103 is equipped with a sensorless suspension that does not require dedicated suspension sensors. The AI-equipped vehicle 103 estimates sensor data related to the suspension based on vehicle behavior information and weight parameters determined by machine learning, and controls the suspension based on the estimated sensor data. The control device of the AI-equipped vehicle 103 has as internal functions a vehicle state setting / detection unit 114, a transmission / reception unit 115, a data management unit 116, a weight data storage unit 117, a vehicle state estimation unit 118, and a suspension control unit 119.

[0021] The AI-equipped vehicle 103 downloads learned weight parameters from the learning management server 102 that are determined to be suitable for the vehicle's state, and uses this parameter set to estimate physical values ​​used by the suspension control unit 119, such as sprung mass velocity and piston velocity. The suspension control unit 119 controls the suspension using the physical values ​​estimated by the dedicated sensorless vehicle state estimation unit 118. By downloading new weight parameters from the learning management server 102, the AI-equipped vehicle 103 can improve the performance of the vehicle state estimation unit 118, which replaces the sensor function, even after shipment.

[0022] In the AI-equipped vehicle 103, instead of directly applying the new parameter set obtained from the learning management server 102, the data management unit 116 within the vehicle performs a process to determine whether the acquired parameters can be applied after obtaining the parameter set or before applying the parameters, and performs suspension control using the parameters that are determined to be applicable.

[0023] The AI-equipped vehicle 103 is equipped with a vehicle state estimation unit 118 that replaces dedicated sensors for controlling the damping force of the suspension. The vehicle state estimation unit 118 calculates physical values, such as sprung speed and piston speed, and transmits them to the suspension control unit 119. The suspension control unit 119 then calculates command values ​​for suspension control based on these physical values ​​and, although not shown in this figure, determines the damping force of the suspensions installed on each of the four wheels of the AI-equipped vehicle 103 and performs suspension control.

[0024] The vehicle state estimation unit 118 acquires vehicle behavior information, such as wheel speed during driving, longitudinal acceleration, lateral acceleration, and yaw rate, from the communication data transmitted over the CAN (Control Area Network) within the vehicle. Then, based on this vehicle behavior information and the weight parameters in the weight data storage unit 117, it derives the aforementioned physical values, such as sprung mass velocity and piston speed.

[0025] In this embodiment, CAN is used as an example of an in-vehicle network, but other in-vehicle networks may be used. For example, CAN FD (CAN with Flexible Data Rate), FlexRay (registered trademark), and In-Vehicle Ethernet (Ethernet (registered trademark)) can be used in the same manner.

[0026] The vehicle state estimation unit 118 is composed of, for example, a neural network, and the weight parameters used in the calculation of the neural network are stored in the weight data storage unit 117, and the data management unit 116 reads the weight parameters from the weight data storage unit 117 and transfers them to the vehicle state estimation unit 118.

[0027] The learning management server 102 collects sensor data and vehicle behavior information from the mass-produced vehicle 101 and manages it according to the vehicle state. The learning management server 102 uses the managed dataset to perform correlation learning between the vehicle behavior information and sensor data and derives the weight parameters of the neural network to be implemented in the AI-applied vehicle 103. The learning management server 102 has internal functions including a transmission / reception unit 109, a management unit 110, a log data storage unit 111, a learning unit 112, and a weight data storage unit 113.

[0028] The weight parameters are derived by the learning management server 102. Although not shown in this figure, the learning management server 102 acquires a dataset consisting of vehicle behavior information and sensor data transmitted over the CAN bus by a test vehicle equipped with dedicated sensors, similar to the mass-produced vehicle 101. The acquired dataset is stored in the log data storage unit 111 via the management unit 110.

[0029] The management unit 110 transfers the dataset stored in the log data storage unit 111 to the learning unit 112. The learning unit 112 learns the correlation between vehicle behavior information and sensor data using a neural network and obtains weight parameters. When vehicle behavior information transmitted over the CAN is input to the neural network to which the weight parameters obtained through learning are applied, physical values ​​that were previously measured by dedicated sensors, specifically sprung mass velocity and piston speed, can be obtained.

[0030] Before shipment, the obtained weight parameters may be directly written to the ECU mounted on the AI-applied vehicle 103 or to the memory on the circuit board. However, the weight parameters are distributed from the transmission / reception unit 109 of the learning management server 102 to the weight data storage unit 117 via the transmission / reception unit 115 and data management unit 116 of the AI-applied vehicle 103.

[0031] The above describes the pre-shipment process. From here on, we move on to the process that is unique to this embodiment, which aims to improve the estimation performance of the vehicle state estimation unit 118, including the mass-produced vehicle 101, assuming the post-shipment period.

[0032] Conventionally, weight parameters were derived using test vehicles before shipment. In this embodiment, however, additional learning is performed using a dataset acquired from a mass-produced vehicle 101 to improve the estimation performance of the vehicle state estimation unit 118. Specifically, the driving data management unit 107 collects a dataset consisting of vehicle behavior information transmitted over the CAN by the mass-produced vehicle 101 while it is running, and physical values ​​such as sprung mass velocity and piston velocity measured by the vehicle state measurement unit 105, and transfers it to the learning management server 102 via the transmission unit 108.

[0033] In the suspension control system 100 of this embodiment, vehicle conditions such as the vehicle type, vehicle class, onboard components, and, in simpler terms, the condition of the tires and whether or not genuine parts are being used are managed. These vehicle conditions include vehicle information regarding the path that road surface stimuli, transmitted via the tires, take to reach the occupants.

[0034] To achieve this management, the mass-produced vehicle 101 is equipped with a vehicle status setting / detection unit 104. The vehicle status setting / detection unit 104 sets things that do not change after shipment, such as the vehicle segment, vehicle type, and year of manufacture, and things that have been intentionally changed after shipment, such as the replacement of onboard parts or the replacement of tires, and detects things that change unintentionally after shipment, such as deterioration.

[0035] For detecting deterioration, if a sensor is installed to detect the condition of the target part, the determination can be made based on the sensor's output value, or it can be derived by measuring the time elapsed since the target part was installed. The vehicle condition set, determined, or derived by the vehicle condition setting / detection unit 104 is transferred to the driving data management unit 107, and combined with the aforementioned data set, is transmitted to the learning management server 102.

[0036] As previously described, the management unit 110 included in the learning management server 102 acquires a dataset with vehicle status information arriving from the mass-produced vehicle 101 via the transmission / reception unit 109 and stores the acquired dataset in the log data storage unit 111, but it manages the dataset separately for each vehicle status. Furthermore, the neural network created in the learning unit 112 is also trained with the dataset corresponding to each vehicle status, and the obtained weight parameters are also managed separately for each vehicle status, including in the weight data storage unit 113.

[0037] The vehicle status setting / detection unit 114 included in the AI-applied vehicle 103, like the vehicle status setting / detection unit 104 in the mass-produced vehicle 101, sets things that do not change after shipment, such as vehicle segment, vehicle type, and year of manufacture, things that have been intentionally changed after shipment, such as replacement of onboard parts or tire changes, and detects things that change unintentionally after shipment, such as deterioration, and records and manages the results as the vehicle status. The recorded vehicle status is then supplied as information to the learning management server 102 in advance.

[0038] When the learning management server 102 receives an update to the weight data for each vehicle state managed by the management unit 110, it distributes the weight parameters to the AI-treated vehicle 103 whose vehicle state matches.

[0039] Figures 2a and 2b illustrate the implementation of vehicle state estimation using a neural network. Figure 2a shows a neural network; although the wiring resembles a simple perceptron, it is a recurrent neural network (RNN) suitable for learning time-series data, illustrating the input and output when the vehicle state estimation unit is implemented using a neural network. Figure 2b shows the relationship between time-series data from the CAN bus and time-series data from the on-board sensors.

[0040] In Figure 2a, 201 represents the input layer elements of the neural network, 202 represents the hidden layer elements of the neural network, and 203 represents the output layer elements of the neural network.

[0041] The input layer element group 201 is composed of an input layer element group 201a for time-series data of a plurality of physical values arriving via CAN, such as time-series data of wheel speed, and an input layer element group 201b for time-series data of vehicle behavior. The estimated instantaneous value of the sprung mass speed or piston speed of the suspension is output to the output layer element group 203.

[0042] The number of elements in the hidden layer element group 202 is generally determined from the number of elements in the input layer element group 201 and the output layer element 203, and is set to the number that maximizes the accuracy of vehicle state estimation by the neural network. Note that the output layer element 203 is determined by the output specification of vehicle state estimation. Note that the CAN-derived time-series data set in the input layer element group 201 is not limited to two types, and may be three or more types.

[0043] Also, in FIG. 2a, one output element is shown so that it becomes either the sprung mass speed or the piston speed, but it is not limited thereto. For example, two output elements may be provided for the sprung mass speed and the piston speed. Since the sprung mass speed and the piston speed are required for each of the four wheels, the network configuration may be summarized with one type of physical value, for example, the sprung mass speed, and four output elements may be provided for the four wheels.

[0044] 204 shown in FIG. 2b is the waveform of the state physical value of the wheel speed, 205 is the waveform of the state physical value of the piston speed, 206 is a window, and 207 is an estimated value (estimated instantaneous value). The vehicle state estimation unit 118 samples the discrete values of the data included in the window 206 from the time-series data such as the wheel speed, longitudinal and lateral accelerations, and yaw rate transmitted on the CAN, and inputs them to the input layer element group 201 of the neural network. The neural network outputs an estimated value 207 corresponding to the data included in the window 206. That is, when training the neural network, a mass-produced vehicle 101 equipped with a dedicated sensor is utilized, and the CAN data 204 (time-series data) during running and the physical value 205 output by the sensor are acquired as teacher data.

[0045] During actual learning, time-series data 204 within window 206 is set in the input layer element group 201 of the neural network, and the physical value (estimated value 207) corresponding to window 206 is set in output layer element 203. Then, learning is performed using the error backpropagation method or the like to determine the weight parameters.

[0046] To further explain the relationship between window 206, estimated value 207, and input layer element group 201, when the width of window 206 is made easy to understand as 1 second, the estimated value 207 is specified to be estimated from the driving history 1 second earlier.

[0047] Therefore, if the sampling interval of time-series data 204 is, for example, 10 milliseconds, the number of sampling points within window 206 is 100 (= 1 second ÷ 10 milliseconds). Further, since it is set in input layer element group 201 according to the types of data used for estimation, for example, if there are two types of data, considering the bias term, the actual number of elements in input layer element group 201 is 201 (= 100 elements / type × 2 types + bias term).

[0048] FIG. 3 is a diagram showing the block configuration of data management unit 116 included in AI application vehicle 103 of the present embodiment.

[0049] 301 is a control unit, 302 is buffer 1, 303 is buffer 2, 304 is switch SW1, 305 is an applicability determination unit, and 306 is switch SW2.

[0050] Data management unit 116 has a pair of buffers 302 and 303. Normally, first, it is assumed that the weight parameters are stored in one of the buffers 302, and the weight parameters stored in that buffer 302 are reflected in vehicle state estimation unit 118 via switch 306, vehicle state estimation is performed by AI, and damping force control of the suspension is performed by suspension control unit 119.

[0051] Also, in AI application vehicle 103, it is assumed that the vehicle setting state is detected by vehicle state setting / detection unit 114, and the information on the vehicle setting state is transferred to learning management server 102 via transmission / reception unit 115.

[0052] Then, when the learning management server 102 updates the weight parameters that are considered optimal for the AI-applied vehicle 103, the learning management server 102 transmits the weight parameters to the control unit 301 via the transmitting / receiving unit 115 and stores them temporarily in the weight data storage unit 117, which is a storage unit.

[0053] New weight parameters (new parameters) arriving from the learning management server 102 are transferred from the weight data storage unit 117 to the other buffer 303 via the control unit 301 and switch 304. The applicability determination unit 305 then reads the weight parameters from buffer 302 and buffer 303 respectively and compares them to determine whether the new weight parameters are applicable to the AI-applied vehicle 103.

[0054] There are various ways to make this determination, but one first step is to check whether the number of weight parameters is appropriate. For example, one could compare it to the number of weight parameters stored in buffer 302 being applied in the subsequent estimation model and see if there is any increase or decrease.

[0055] Alternatively, the applicability determination unit 305 may be configured to perform neural network calculations and matrix operations, and the determination may be made based on whether the calculation result is normal when time-series data related to vehicle behavior arriving from the CAN is input and weight parameters stored in the buffer 303 are selected.

[0056] In this case, determining whether the calculation result is normal involves, for example, checking whether the calculation result is fluctuating between the upper and lower limits of the expected output value, or in other words, whether it is fixed at either the upper or lower limit. Alternatively, the output of the subsequent vehicle state estimation unit 118 may be transferred to the applicability determination unit 305, and the transferred estimated value may be compared with the calculation result when the weight parameters stored in the aforementioned buffer 303 are selected.

[0057] Since the calculations performed for the same purpose are being conducted, it is unlikely that the estimated values ​​before updating the weight parameters and the calculation results after applying the updated weight parameters will differ significantly. Therefore, it is acceptable to set an arbitrary threshold and determine that there is no problem if the difference between the two values ​​is smaller than the threshold.

[0058] In this manner, or by any other method that can achieve a similar objective, the applicability determination unit 305 determines whether the weight parameters received from the learning management server 102 are suitable for application and transmits the determination result to the control unit 301.

[0059] If applicable, the control unit 301 switches the switch 306 to transfer the weight parameters stored in the other buffer 303 to the subsequent vehicle state estimation unit 118. It goes without saying that the weight parameters should not be updated while the vehicle state estimation unit 118 is performing calculations, and the switch 306 should be operated so that the weight parameters are updated when calculations are not being performed.

[0060] Figure 4 is a diagram showing the block configuration of the management unit in the server of the first embodiment. In Figure 4, reference numeral 401 denotes the data distribution unit, 402 denotes the group management table, 403 denotes map data, 404 denotes the distribution data group, 405 denotes distribution data 1, 406 denotes distribution data 2, 407 denotes distribution data N, 408 denotes the learning necessity determination unit, and 409 denotes the data management unit.

[0061] The data management unit 409 manages N types of vehicle / model settings, and within this range, it assumes a variety of AI-applied vehicles 103 and corresponding mass-produced vehicles 101. Driving data arriving from multiple mass-produced vehicles 101 is input to the data distribution unit 401 via the transmission / reception unit 109. The data distribution unit 401 refers to the group management table 402 to determine which of the N types of vehicle / model settings it corresponds to, and stores it in one of the corresponding distribution data 405 to 407 included in the distribution data group 404.

[0062] The learning necessity determination unit 408 checks for the presence or absence of driving history by referring to the map data 403 for each distribution data unit, and if a driving log for which there has been no prior driving experience is added, it determines that learning using time-series data for that period is necessary. The time-series data of vehicle behavior and vehicle condition measurements for the target period are transferred as driving logs to the log data storage unit 111.

[0063] Although not shown in the diagram, the log data storage unit 111 stores driving logs separately for each of the N types of vehicle / vehicle settings. The data management unit 409 transfers the driving logs to the learning unit 112 once a certain amount of data has been accumulated for each of the N types of vehicle / vehicle settings. As described above, the learning unit 112 learns the correlation between the time-series data of vehicle behavior included in the driving logs and the measured vehicle state values, and derives weight parameters. The weight parameters obtained through learning are transferred to the weight data storage unit 113 via the data management unit 409, and the weight parameters are further distributed to the corresponding AI-applied vehicle 103 via the data management unit 409.

[0064] Furthermore, in order to further improve the judgment made by the learning necessity determination unit 408, the learning necessity determination unit 408 acquires weight parameters corresponding to the distribution data stored in the weight data storage unit 113 via the data management unit 409. Then, by comparing the vehicle state estimate obtained by neural network calculation or matrix operation using the acquired weight parameters with the vehicle state measurement values ​​received from the mass-produced vehicle 101, the system determines that learning is necessary if the difference between the two values ​​is greater than a preset threshold, and that learning is not necessary if the difference is small, thereby directly determining the quality of the weight parameters.

[0065] Figure 5 shows an example of a driving log recorded in the log data storage unit 111 of the first embodiment. 501, 502, and 503 show how the log is managed according to the differences in vehicle type / vehicle settings. Although only three are shown in Figure 5, there are assumed to be N types. The driving log for each vehicle type / vehicle setting is managed by combining all or part of the time-series data of physical values ​​related to vehicle behavior, such as vehicle speed, steering angle, longitudinal acceleration, and lateral acceleration, with time-series data of vehicle behavior measurements, such as sprung speed and piston speed, all within the range of the same timestamp.

[0066] Figure 6 is a flowchart of the weight data distribution process in the server of the first embodiment. In Figure 6, reference numeral 601 denotes the process of receiving vehicle type / vehicle information / logs, 602 denotes the data distribution process, 603 denotes the distribution data storage process, 604 denotes the distribution data / past data reading process, 605 denotes the learning necessity determination process, 606 denotes the learning data storage process, 607 denotes the data backup process, 608 denotes the learning start determination process, 609 denotes the learning process, 610 denotes the weight data storage process, and 611 denotes the weight data distribution process.

[0067] First, it is assumed that information regarding the mass-produced vehicles 101 and the AI-applied vehicles 103, which are under the management of the learning management server 102, is stored in the group management table 402.

[0068] The vehicle type / vehicle information / log receiver 601 receives, in particular, group information from the group management table 402, that is, driving logs linked to the vehicle type / vehicle, specifically vehicle behavior information and vehicle behavior measurement results used for the suspension, with regard to mass-produced vehicles 101.

[0069] The data distribution 602 distributes the driving logs received by the vehicle type / vehicle information / log reception 601 to the corresponding group's storage area by referring to the group information in the group management table 402. The distributed data storage 603 stores the driving logs distributed to the corresponding group's storage area in the weight data storage unit 113. The post-distribution data / past data reading 604 reads the weight parameters corresponding to the group information and the driving logs stored for each group from the weight data storage unit 113.

[0070] The learning necessity determination 605 calculates the estimated vehicle state using the neural network with the weight parameters read by the sorted data / past data readout 604 and the driving log. It also compares the vehicle behavior measurement results included in the driving log with the aforementioned calculation results to calculate the quality of the vehicle state estimate by the neural network. This quality calculation is based on the difference between the measurement result and the estimation result, the so-called error. If the error is above a certain value, it is determined that learning is necessary to improve accuracy (Yes). If the error is below a certain value, it is determined that learning is not necessary (No), and the current weight parameters can be used for continued operation.

[0071] Then, the learning data storage 606 adds only the driving logs that it determines are necessary for learning to the log data storage unit 111, and any other driving logs that are not deemed necessary for learning are saved by the data saving 607.

[0072] The learning start determination 608 is a process designed to prevent learning from starting until a certain amount of additional data has been accumulated. This is because if learning were to start every time additional driving logs are added, the learning management server 102's computer would have to keep running continuously. For example, a rule may be defined to start learning once one day's worth of additional driving logs have been stored.

[0073] Training 609 is the training of a neural network that realizes vehicle state estimation, and is to be carried out using well-known methods such as backpropagation.

[0074] The weight data storage 610 stores the new weight parameters obtained in the learning 609 in the weight data storage unit 113. The weight data storage unit 113 is designed to retain at least two generations of weight parameters, and to allow users to revert to the previous generation of weight parameters as needed. Finally, the weight data distribution 611 distributes the weight parameters to the corresponding group of AI-applied vehicles 103.

[0075] Figure 7 is a flowchart of the applicability determination process in the AI-applied vehicle of the first embodiment. In Figure 7 and its description, the numbers before and inside the parentheses switch depending on the state.

[0076] In Figure 7, reference numeral 701 denotes the process of detecting the vehicle type / vehicle setting, 702 denotes the process of declaring the vehicle type / vehicle setting, 703 denotes the process of reading weight data, 704 denotes the process of transferring weight data to buffer 1, 705 denotes the process of estimating the driving and vehicle state, 706 denotes the process of determining the end of driving, 707 denotes the process of determining whether a call has been made, 708 denotes the process of receiving weight data, 709 denotes the process of transferring weight data to buffer 2, 710 denotes the process of determining whether the data is applicable, 711 denotes the process of selecting weight data from buffer 2, 712 denotes the process of clearing buffer 1, 713 denotes the process of writing buffer 2 data to storage, 714 denotes the process of selecting weight data from buffer 1, and 715 denotes the process of clearing buffer 2.

[0077] First, the ignition is turned on, and the various systems and drivetrain of the car transition to an activated state. During this time, the vehicle type / vehicle setting detection 701 detects the vehicle setting state of the AI-applied vehicle 103, and the vehicle type / vehicle setting declaration 702 declares the vehicle setting state of the AI-applied vehicle 103 to the learning management server 102, which is the management source. Furthermore, the weight data reading 703 reads the weight parameters from the storage (weight data storage unit 117) before the start of driving, and the weight data transfer to buffer 1(2) 704 stores the weight parameters in buffer 1(2) which is referenced by the vehicle state estimation unit 118. This completes the preparation for driving, and driving operations such as accelerator operation become possible from the driving & vehicle state estimation 705 onwards.

[0078] The driving completion determination 706 is a function that determines whether the driving process has ended and the ignition has been turned off. If it is determined to be completed (Yes), the flowchart ends. If it is determined to continue (No), the process of improving the estimation accuracy of the vehicle state estimation unit 118 continues in communication with the learning management server 102.

[0079] The call status determination 707 is a process that checks whether or not a notification for weight parameter updates has been received from the learning management server 102. Until an update notification is received, the system will continue to estimate driving and vehicle state using the current weight parameters.

[0080] If an update notification arrives (Yes), the weight data receiver 708 receives the weight parameters distributed from the learning management server 102, and the weight data transfer 709 to buffer 2 transfers the data to buffer 2(1), which is one of the two buffers and is not referenced by the vehicle state estimation unit 118.

[0081] Next, the applicability determination 710 determines whether the transferred weight parameters are appropriate, for example, by checking for consistency in the number of weight parameters, data corruption, or differences in the data format itself. Based on this determination result, subsequent processing is carried out. If the applicability determination 710 does not find any problems with the weight parameters received from the learning management server 102 (Yes), the weight data selection 711 of buffer 2(1) switches the second switch 306 to switch the reference destination of the vehicle state estimation unit 118 from buffer 1(2) to buffer 2(1).

[0082] Then, buffer 1(2), which will no longer be used, is cleared of its data by buffer 1(2) clear 712 in preparation for the arrival of the next weight parameter. Also, if a problem is found with the weight parameter that has arrived from the learning management server 102 (No), the second switch 306 is not switched by the weight data selection 714 of buffer 1(2), and the problematic weight parameter is deleted by buffer 2(1) clear 715.

[0083] Although the deletion of weight parameters was explained in buffer 1(2) clear 712, it is also acceptable to retain the values ​​and overwrite them at the next update timing. Furthermore, it is assumed that the weight data storage unit 117 holds weight parameters for two or more generations from the most recent, and that weight parameters can be set in buffers 302 and 303 as needed.

[0084] According to the suspension control system 100 of this embodiment, each time the applicability determination unit 305 of the data management unit 116 of the AI-applied vehicle 103 acquires a new weight parameter, it compares the newly acquired weight parameter (new parameter) with the old parameter already acquired in the weight data storage unit 117 (parameter storage unit) to determine whether the new parameter is applicable to the AI-applied vehicle 103. This prevents, for example, the application of a weight parameter that is not suitable for the AI-applied vehicle 103, and enables optimal suspension control.

[0085] [Second Embodiment] The second embodiment, like the first embodiment, is a dedicated sensorless suspension system to which AI is applied, and is designed to improve the performance of the vehicle state estimation unit 118, which replaces the sensor function even after shipment. In particular, in vehicles to which AI is applied, a ride comfort determination unit is installed to determine the quality of the weight parameters of the vehicle state estimation unit 118 based on the ride comfort.

[0086] In other words, if the ride comfort is good, the substitute performance of the sensor function in the vehicle state estimation unit 118 is also good. Conversely, if the ride comfort is poor, for example, if there is a lot of vertical vibration, it is determined that the substitute performance of the sensor function in the vehicle state estimation unit 118 may have been poor. Based on this determination, a key feature is that warnings and learning requests can be presented to the learning management server 102.

[0087] Details of the second embodiment will be described using Figures 8 and 9. Note that components similar to those in the first embodiment are denoted by the same reference numerals, and their detailed descriptions will be omitted.

[0088] Figure 8 is a functional block diagram of the suspension control system in the second embodiment. 801 is an AI-applied vehicle, 802 is a ride comfort determination unit, and 803 is a data management unit. A distinctive feature of this embodiment is the ride comfort determination unit 802, which determines ride comfort based on vehicle behavior information transmitted over the CAN.

[0089] The ride comfort determination unit 802 is located within the AI-applied vehicle 801 and refers to all or some of the time-series data of the aforementioned physical quantities of vehicle behavior arriving from the CAN, such as longitudinal acceleration, lateral acceleration, pitch rate, roll rate, and yaw rate, and focuses on their values ​​and their derivatives, i.e., the amount of change.

[0090] The thresholds used for comparison will be set separately. If the value is greater than the threshold, it will be judged that the ride comfort is likely to be poor, and if the value is less than the threshold, it will be judged that the ride comfort is good.

[0091] The ride comfort determination unit 802 can also use vehicle behavior information as input to estimate ride comfort using a neural network and use it for ride comfort determination. The vehicle behavior information transmitted over the CAN includes acceleration information from an acceleration sensor, image information from a camera, vehicle speed from a vehicle speed sensor, angular velocity information from a gyroscope, etc. It is preferable that acceleration sensors be mounted in multiple locations on the vehicle, and acceleration information from each part is obtained from acceleration sensors mounted in each part that makes up the vehicle's chassis, such as the unsprung mass and sprung mass that constitute the vehicle's suspension, the area around the seat where the occupants sit, and the steering wheel that the driver holds.

[0092] The ride comfort determination unit 802 has an evaluation circuit that calculates evaluation values ​​for ride comfort. The evaluation circuit calculates evaluation values ​​for ride comfort such as flatness, harshness, stroke, softness, and texture. The weight parameters used by the evaluation circuit when calculating the evaluation values ​​are stored in the weight data storage unit 117. If the evaluation circuit is implemented using a hierarchical neural network, the weight data storage unit 117 stores information on the inter-element connection weights Wij in each neural network.

[0093] The evaluation circuit uses a hierarchical neural network that takes time-series data of sensor information as input and outputs evaluation values ​​for sensory indicators. The evaluation circuit can be constructed as a three-layer hierarchical neural network in which elements of the input layer, hidden layer, and output layer are hierarchically connected. One element representing a bias term is set in both the input layer and the hidden layer. Each element of the input layer is connected to each element of the hidden layer by weight W1ij, and each element of the hidden layer is connected to each element of the output layer by weight W2ij (see Figure 2a).

[0094] Time-series data is data obtained by dividing the sensor information data output from each sensor in a sensor group at predetermined intervals into fixed-period segments and processing them as a single data sequence. An evaluation circuit can be constructed by building a hierarchical neural network by learning the correlation between a large amount of time-series data and the evaluation values ​​provided by expert drivers using machine learning. For machine learning of neural networks, techniques such as the commonly known backpropagation method can be used.

[0095] Figure 9 is a flowchart of the ride comfort determination process in the second embodiment. In Figure 9, reference numeral 901 denotes the process of determining whether the vehicle is in motion, 902 denotes the process of measuring vehicle behavior, 903 denotes the process of determining ride comfort, 904 denotes the process of assigning a flag that needs to be checked, 905 denotes the process of writing the driving log, 906 denotes the process of counting up the number of flags, 907 denotes the process of determining completion, 908 denotes the process of issuing a warning / learning request, and 909 denotes the process of deleting the log.

[0096] The ride comfort determination unit 802 may judge ride comfort based on instantaneous values, but here we will explain it assuming that it judges the overall ride comfort over a certain period of time or longer.

[0097] First, when the ignition is turned on, the car's various systems and drivetrain are activated, and the driving judgment 901 determines whether or not vehicle speed has been generated when the accelerator pedal is pressed. As mentioned above, the vehicle behavior measurement 902 acquires the physical values ​​of the vehicle behavior information arriving from the CAN, and the ride comfort judgment 903 determines the quality of the ride comfort based on the magnitude and amount of change of the physical values.

[0098] If the ride comfort judgment 903 determines that the ride comfort is poor (No), it is assumed that the vehicle state estimation assumed in this embodiment may have been inaccurate. In this case, a flag is added in the confirmation flag assignment 904, and the driving log writing 905 defines the target period and a certain period before and after it, for example, a 1-second driving log, as a single driving scene and stores it in the vehicle's storage device.

[0099] Then, the flag count up 906 counts the number of flags to determine the number of driving scenes in which ride comfort deteriorated. The termination judgment 907, for example, if the number of flags exceeds a certain number, determines that it is necessary to improve the accuracy of the vehicle state estimation function and terminates the series of ride comfort judgments.

[0100] If the ride comfort judgment 903 determines that the ride comfort is good (Yes), it is determined that the vehicle state estimation was highly accurate, and the log deletion 909 deletes the driving log corresponding to that driving scene. Once the ride comfort judgment is complete, the warning / learning request 908 sends a warning or a request for additional learning command to the learning management server along with the driving scene data.

[0101] In this explanation, the ride comfort determination unit 802 is assumed to be included in the AI-applied vehicle 801. However, the measurement results of the vehicle behavior measurement unit 902 may be sent to the learning management server as needed, and subsequent processing may be performed by the management unit 110 on the learning management server.

[0102] [Third Embodiment] The third embodiment, like the second embodiment, determines the quality of the weight parameters based on ride comfort, but is characterized by its ability to determine the quality of the weight parameters in a more sophisticated way by using information such as an onboard camera to grasp the road surface profile of the road in front of the vehicle, that is, the surface condition of the road.

[0103] More specifically, if the road surface is extremely rough, even if the damping force of the active suspension is properly controlled, the ride quality may be poor, meaning there may be significant vertical vibration. In such cases, the system determines that there was no problem with the substitute performance of the sensor function in the vehicle condition estimation unit 118. In this way, the system is characterized by its ability to capture ride quality more accurately and to present warnings and learning requests to the learning management server 102.

[0104] Details of the third embodiment will be described using Figure 10. Figure 10 is a functional block diagram showing the suspension control system in the third embodiment. Note that components similar to those in the second embodiment are denoted by the same reference numerals, and their detailed descriptions are omitted. In Figure 10, reference numeral 1001 represents the AI-applied vehicle, 1002 represents the road surface profile extraction unit, and 1003 represents the ride comfort determination unit.

[0105] A distinctive feature of this embodiment is the road surface profile derivation unit 1002. The road surface profile derivation unit 1002 acquires road surface information, i.e., unevenness information, in front of the AI-applied vehicle 1001 from images obtained by a vehicle-mounted camera system capable of acquiring 3D data, such as a stereo camera or a mono camera that derives depth information in the time direction.

[0106] If the elevation difference of the road surface irregularities is greater than a certain level, as mentioned above, even if the active suspension is properly controlled, it is quite possible that the ride comfort will be somewhat poor.

[0107] Based on the above, ride comfort is determined based on the determination result of the height difference of the unevenness of the road surface. In other words, in the simplest form, the ride comfort determination unit 1003 is activated only if the height difference of the unevenness of the road surface is less than a certain level, and based on the determination result, a warning or learning request is sent to the learning management server 102. If the height difference of the unevenness of the road surface is greater than a certain level, the ride comfort determination unit 1003 is not activated, and the AI-applied vehicle 1001 continues to drive. The rest is the same as in the second embodiment of the present invention, so the explanation is omitted.

[0108] Although embodiments of the present invention have been described in detail above, the present invention is not limited to the embodiments described above, and various design modifications can be made without departing from the spirit of the invention as described in the claims. For example, the embodiments described above are described in detail in order to explain the present invention in an easy-to-understand manner, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add a configuration of another embodiment to the configuration of one embodiment. Moreover, it is possible to add, delete, or replace a part of the configuration of each embodiment with other configurations.

[0109] 100...Suspension control system, 101...Mass-produced vehicle, 102...Learning management server, 103, 801, 1001...AI-applied vehicle, 104...Vehicle state setting / detection unit, 105...Vehicle state measurement unit, 106...Suspension control unit, 107...Driving data management unit, 110...Management unit, 111...Log data storage unit, 112...Learning unit, 113...Weight data storage unit, 114...Vehicle state setting / detection unit, 116...Data management unit, 117...Weight data storage unit, 118...Vehicle state estimation unit, 119...Suspension control unit, 301...Control unit, 302...Buffer 1, 303...Buffer 2, 304...Switch SW1, 305...Applicability determination unit, 306...Switch SW2, 408...Learning necessity determination unit, 802...Ride comfort determination unit, 1002...Road surface profile derivation unit, 1003...Ride comfort determination unit

Claims

1. A suspension control device that acquires parameters indicating the correspondence between vehicle behavior information and suspension sensor data, estimates the sensor data based on the parameters and vehicle behavior information, and controls the suspension using the sensor data, characterized in that it has an applicability determination unit that compares newly acquired parameters with previously acquired old parameters to determine whether the new parameters are applicable to the vehicle.

2. The suspension control device according to claim 1, characterized in that the applicability determination unit compares the number of parameters of the old parameters with the number of parameters of the new parameters and determines applicability based on whether there is an increase or decrease.

3. The suspension control device according to claim 1, characterized in that the applicability determination unit determines that the device is applicable when the difference between the estimated value of the sensor data estimated using the new parameters and the estimated value of the sensor data estimated using the old parameters is smaller than a threshold.

4. The suspension control device according to claim 1, characterized in that it has a ride comfort determination unit that determines whether or not the ride comfort of the vehicle is good based on changes in the vehicle's behavior.

5. The suspension control device according to claim 4, characterized in that the ride comfort determination unit estimates the ride comfort of the vehicle using a hierarchical neural network constructed by learning the correlation between time-series data of the vehicle's behavior and evaluation values ​​provided by an expert driver through machine learning.

6. The suspension control device according to claim 5, comprising a road surface profile derivation unit that derives a road surface profile in front of the vehicle using information from an onboard camera, wherein the ride comfort determination unit performs a ride comfort determination of the vehicle only when it determines from the road surface profile that the height difference of the unevenness of the road surface is less than a certain amount.

7. A suspension control system comprising: a suspension control device that acquires parameters indicating the correspondence between vehicle behavior information and suspension sensor data, estimates the sensor data based on the parameters and vehicle behavior information, and controls the suspension using the sensor data; and a learning management server that provides the parameters to the suspension control device, wherein the suspension control device has an applicability determination unit that compares newly acquired new parameters with previously acquired old parameters to determine whether the new parameters are applicable to the vehicle; and the learning management server has a management unit that acquires vehicle behavior information and suspension sensor data, and a learning unit that additionally learns the parameters using the vehicle behavior information and suspension sensor data acquired by the management unit.

Images