Suspension system

Abstract

A suspension system performs calculations using machine learning by mapping sensor information and suspension state information. The suspension system includes: a weighting parameter storage unit; a vehicle state estimation unit; a first vehicle behavior calculation unit that calculates a first physical value based on the estimation result; a second vehicle behavior calculation unit that calculates a second physical value based on the sensor information; an estimation accuracy verification unit that outputs an estimation accuracy of the suspension state by comparing the first and second physical values; and a driving data control unit that instructs weighting parameter learning based on an output result of the estimation accuracy.

Description

Suspension system

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

[0002] Suspension control systems control vehicle posture by changing the damping force of the suspension springs according to road conditions and driving conditions. The systems calculate control values ​​by detecting sensor data obtained from sensors that detect the piston speed and sprung vertical speed of the suspensions on each of the vehicle's four wheels. To reduce costs and labor required for adaptation, there is a technology that makes this sensor function sensorless by incorporating logic that estimates sensor values ​​from Controller Area Network (CAN) data and a neural network that realizes the estimation function.

[0003] However, roads vary widely, and it is not easy for the aforementioned sensor value estimation technology to replace sensor functions on all roads. In particular, when applying a neural network, it may be difficult to continuously replace sensors with a trained neural network that is installed at the time of shipment. For this reason, suspension control systems require a continuous learning system to improve road coverage.

[0004] Furthermore, in the field of image recognition, where machine learning is sometimes applied, it is easy for humans to identify images with low recognition rates, and additional learning can be carried out specifically for those images. However, in sensor value estimation technology that applies a neural network based on the sensorless concept, it is difficult to identify poor road conditions because there are no sensor values ​​that can serve as training data.

[0005] For example, Patent Document 1 listed below discloses a damper control system that can receive feedback data regarding vehicle behavior and execute a machine learning algorithm to control damper characteristics with high responsiveness and robustness.

[0006] Japanese Patent Application Laid-Open No. 2021-17168

[0007] Based on conventional technology, in order to realize a sensorless system in which the suspension state is estimated from information from other sensors without directly detecting it with a sensor, it is necessary to further extract weak points in estimating sensor values ​​and to perform continuous learning to improve the accuracy of the machine learning model. In light of this, an object of the present invention is to provide a sensorless suspension system that extracts weak points of a neural network and supports learning when the sensor function is replaced by a neural network.

[0008] A suspension system that performs machine learning calculations by associating sensor information obtained from a plurality of sensors provided on a vehicle with state information of a suspension provided on the vehicle, and includes: a weight parameter memory unit that stores weight parameters calculated by the machine learning; a vehicle state estimation unit that outputs an estimation result of the suspension state based on the sensor information and the weight parameters; a first vehicle behavior calculation unit that calculates a first physical value of a physical quantity related to the behavior of the vehicle based on the output estimation result; a second vehicle behavior calculation unit that calculates a second physical value of the physical quantity based on the sensor information; an estimation accuracy verification unit that outputs the estimation accuracy of the suspension state by the vehicle state estimation unit by comparing the first physical value with the second physical value; and a driving data management unit that instructs a learning management server to learn the weight parameters based on the output result of the estimation accuracy by the estimation accuracy verification unit.

[0009] When the suspension state is estimated from other sensor information rather than being directly detected by a sensor, a sensorless suspension system can be provided that extracts the weak points of the neural network and supports learning.

[0010] FIG. 1 is a functional block diagram of a suspension system according to a first embodiment of the present invention; FIG. 2 is a diagram showing the vehicle state estimation unit of FIG. 1 as a neural network configuration; FIG. 3 is a diagram showing the relationship between input and output data sets when learning a neural network; FIG. 4 is a flowchart related to vehicle-side vehicle state estimation and notification of need for learning; FIG. 5 is a flowchart related to vehicle-side weight data update; FIG. 6 is a flowchart related to server-side weight data learning; FIG. 7 is a flowchart related to server-side abnormality detection response;

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The following description and drawings are examples for explaining the present invention, and some omissions and simplifications have been made as appropriate for clarity of explanation. The present invention can be implemented in various other forms. Unless otherwise specified, each component may be singular or plural.

[0012] In order to facilitate understanding of the invention, the position, size, shape, range, etc. of each component shown in the drawings may not represent the actual position, size, shape, range, etc. Therefore, the present invention is not necessarily limited to the position, size, shape, range, etc. disclosed in the drawings.

[0013] (First embodiment of the present invention and overall configuration of the device) (Fig. 1) A suspension system that controls damping force to improve the ride comfort of an automobile, based on a sensorless system in which no suspension sensors are mounted on the vehicle, comprises a vehicle 101, a learning management server 102, and an internet environment 103. The vehicle 101 comprises a suspension ECU (Electronic Control Unit) 104, an ECU 105, sensors 106a...106b, a sensor ECU 107, a CAN 108, an active suspension 109 (hereinafter referred to as suspension 109), a display unit 110, and an interface (I / F) 111.

[0014] The suspension ECU 104 includes a vehicle state estimation unit 112, a weight data management unit 113, a weight parameter storage unit 114, a suspension control value calculation unit 115, a first pitch rate calculation unit 116, a second pitch rate calculation unit 117, an estimation accuracy verification unit 118, a traveling data management unit 119, and a traveling data storage unit 120.

[0015] The suspension 109 suppresses vibrations that are transmitted from the road surface via the tires and that are felt by the driver or passengers seated in the seat when the vehicle 101 is traveling. The damping force characteristics of the suspension 109 are controlled by the suspension ECU 104, and a suspension control value calculation unit 115 of the suspension ECU 104 derives a control value for the suspension 109.

[0016] The learning management server 102 is a server that manages the learning of the vehicle state estimation unit 112 included in the suspension ECU 104, and transmits information to the vehicle 101 via the Internet environment 103. The ECU 105 transmits the information of the learning management server 102 transmitted from the Internet environment 103 to the CAN 108. As a result, the suspension ECU 104 shares information with each functional unit via the CAN 108. The sensor ECU 107 acquires information from sensors 106a...106b and transmits it to the CAN 108, thereby sharing the sensor information with each functional unit of the vehicle 101.

[0017] The following describes each functional unit of the suspension ECU 104. The suspension ECU 104 acquires information from the CAN 108 via the interface 111 or outputs information to the CAN 108. The weight data management unit 113 determines calculation specifications to be used for estimation by the vehicle state estimation unit 112 by reading out weight parameters stored in a weight parameter storage unit 114 (weight parameter storage unit). The weight data management unit 113 also acquires learned weight parameters from the learning management server 102 via the CAN 108 and updates the weight parameters.

[0018] The vehicle state estimation unit 112 references sensor information acquired by the sensors 106a...106b as specific information from the data constantly transmitted within the CAN 108. The sensors 106a...106b detect the rotational speed (wheel speed) of one of the four wheels of the vehicle 101, as well as longitudinal acceleration, lateral acceleration, vertical acceleration, yaw rate, roll rate, pitch rate, etc., in the vehicle behavior system, and transmits the sensor information to the CAN 108. Note that while FIG. 1 illustrates the sensors 106a...106b as if there were only two types, it goes without saying that in reality, there are many sensors that measure the physical values ​​mentioned above. Furthermore, if the vehicle 101 on which the vehicle state estimation unit 112 is installed is a learning vehicle, it is equipped with an acceleration sensor for the suspension, and the data from this sensor is also transmitted to the CAN 108.

[0019] The vehicle state estimation unit 112 acquires sensor information from each of the sensors 106 a...106 b. The vehicle state estimation unit 112 also acquires the values ​​of weight parameters read by the weight data management unit 113 from the weight parameter storage unit 114 as data related to the vehicle state during traveling. The vehicle state estimation unit 112 outputs estimation results of physical quantities related to suspension control for each of the four wheels based on the acquired sensor information and the acquired weight parameters.

[0020] The suspension control value calculation unit 115 calculates a control value for controlling the damping force of the suspension 109 based on the physical quantity estimated by the vehicle state estimation unit 112. By calculating the control value using estimation in this way, it becomes unnecessary to install an acceleration sensor for measuring the sprung vertical velocity and the piston stroke velocity as in the conventional system, which contributes to cost reduction.

[0021] Next, verification of the estimation accuracy of the vehicle state estimation unit 112 will be described. First, the vehicle state estimation unit 112 transfers the estimated sprung vertical velocities of the suspensions of the four wheels (state information of the suspensions 109) to the first pitch rate calculation unit 116. The first pitch rate calculation unit 116 calculates a pitch rate (first physical value) based on the transferred sprung vertical velocities for the four wheels. Meanwhile, wheel speed data (sensor information) of the four wheels included in the CAN 108 acquired from multiple sensors 106a...106b provided on the vehicle 101 is transferred to the second pitch rate calculation unit 117. The second pitch rate calculation unit 117 calculates a pitch rate (second physical value) based on the transferred wheel speed data of the four wheels.

[0022] The present invention focuses on the fact that the pitch rate can be calculated using either the sprung vertical velocity or the wheel speeds of the four wheels, and aims to verify the estimation accuracy of the first pitch rate derived from the estimation result using the second pitch rate as a model. By comparing the two pitch rates derived by calculation based on the sprung vertical velocity and the wheel speeds of the four wheels, if the two pitch rates are similar values, it is determined that the estimation accuracy of the sprung vertical velocity estimated by the vehicle state estimation unit 112 is good.

[0023] The estimation accuracy verification unit 118 compares the two pitch rates (first physical value and second physical value) calculated by the first pitch rate calculation unit 116 and the second pitch rate calculation unit 117. If the difference between the two pitch rates is smaller than a predetermined threshold, the estimation accuracy verification unit 118 determines that the estimation result output by the vehicle state estimation unit 112 is acceptable. On the other hand, if the difference between the two pitch rates is larger than the predetermined threshold, the estimation accuracy verification unit 118 determines that the estimation result output by the vehicle state estimation unit 112 is unacceptable, and identifies a state in which the output accuracy of the vehicle state estimation unit 112 is poor. If the estimation accuracy verification unit 118 identifies a state in which the output accuracy of the vehicle state estimation unit 112 is poor, it assigns a flag indicating a problem to the time-series data set input to the vehicle state estimation unit 112 via the CAN 108.

[0024] In this way, information on the difference between the first physical value and the second physical value is output as the estimation accuracy for the state of the suspension 109, which is the estimation result output by the vehicle state estimation unit 112. Note that the first physical value and the second physical value are not limited to the pitch rate, and may be other vehicle behavior data that can be calculated based on multiple pieces of information or data, such as the roll rate.

[0025] In this way, by comparing the pitch rate calculated from the sprung vertical velocity estimated by the neural network (the pitch rate calculated by the first pitch rate calculation unit 116) with the pitch rate calculated from the wheel speed information from the CAN 108 that does not go through the neural network (the pitch rate calculated by the second pitch rate calculation unit 117), it is determined whether the estimation accuracy of the sprung vertical velocity by the vehicle state estimation unit 112 is low.

[0026] The weight parameters used in the neural network of the vehicle state estimation unit 112 are parameters obtained by learning training data created from a verification vehicle, etc., equipped with a suspension acceleration sensor, separate from the vehicle 101.

[0027] The estimation accuracy verification unit 118 transfers the dataset flagged as problematic to the traveling data management unit 119. The traveling data management unit 119 stores the transferred dataset in the traveling data storage unit 120. Through this process, datasets that the vehicle state estimation unit 112 deems inefficient are accumulated in the traveling data storage unit 120. The traveling data management unit 119 instructs the learning management server 102 to update the weight parameters for the traveling environment of the vehicle 101 based on the accumulated inefficient datasets. The learning management server 102 performs learning to update the weight parameters in accordance with the instructions of the traveling data management unit 119 made based on the output of the estimation accuracy verification unit 118. In this way, the learning management server 102 learns the weight parameters based on the output result of the estimation accuracy verification unit 118 in the suspension ECU 104 mounted on the vehicle 101.

[0028] (FIG. 2A) The neural network of vehicle state estimation unit 112 is configured as a three-layer hierarchical neural network in which elements of an input layer (number of elements i) 201, a hidden layer (number of elements j) 202, and an output layer (number of elements K) 203 are hierarchically connected. Input layer element group 201 of the neural network includes input element group 201a for wheel speed time series and input element group 201b for vehicle behavior time series, and the state of suspension 109 is estimated using these two element groups.

[0029] The number of elements in the hidden layer element group 202 is generally determined by the number of elements in the input layer element group 201 and the output layer element group 203, and is set to a number that maximizes the accuracy of state estimation by the neural network. The output layer element group 203 outputs an estimated instantaneous value of the piston velocity or the sprung vertical velocity of the vehicle suspension 109. Note that while Figure 2A shows a multilayer perceptron as an example of a neural network, a recurrent neural network, which is advantageous for learning time-series data, may also be used.

[0030] Each element of the input layer element group 201 and each element of the hidden layer element group 202 are connected by weights W1ij (i = 1 to I, j = 1 to J), and each element of the hidden layer element group 202 and the output layer element 203 are connected by weights W2jk (j = 1 to J, k = 1). Information on these weights (weight parameters) is expressed as a determinant of the weights W1ij and W2jk. The weight parameters are calculated in advance by machine learning and stored in the weight parameter storage unit 114, and the vehicle state estimation unit 112 performs calculations using a neural network by acquiring the weight parameters from the weight parameter storage unit 114 via the weight parameter management unit 113. Note that, in this example, the simplest hidden layer element group 202 is a neural network with one layer of all-element connections, but this is not limiting.

[0031] (Fig. 2B) The upper diagram shows an example of wheel speed from the information in CAN 108. As described above, neural network learning is performed by driving a verification vehicle equipped with an acceleration sensor that measures piston speed and sprung vertical speed for controlling suspension 109, and is performed using time-series data 204 of the data contained in the CAN of the verification vehicle and sensor data (time-series data of an on-board sensor) 205.

[0032] Specifically, discrete values ​​obtained by sampling data included in a window 206 from the time-series data 204 are set in the input layer element group 201 ( FIG. 2A ) of the neural network. Then, corresponding instantaneous sensor data values ​​207 are set in the output layer element 203 ( FIG. 2A ) of the neural network. Regarding the relationship between the window 206 and the instantaneous sensor data values ​​207, if the width of the window 206 is, for example, 1 second, the instantaneous sensor data values ​​207 are estimated from the driving history from 1 second prior. Therefore, if the sampling interval is, for example, 20 m / s, there will be 50 sampling points within the window 206 (= 1 second ÷ 20 m / s), and the combination of these 50 data groups and the instantaneous sensor data values ​​207 is defined as a training dataset. Then, by sliding the window 206 and the instantaneous sensor data values ​​207 in the time direction every 20 m / s, the dataset is multiplied by n. This allows neural networks to be trained using a large number of datasets.

[0033] (FIG. 3A) The following describes the estimation of the vehicle state and the notification of the need to learn in the vehicle 101. First, in step S101, the vehicle state estimation unit 112 determines whether the vehicle 101 is moving. If the vehicle 101 is moving, the process proceeds to steps S102 and S103, and if the vehicle 101 is not moving, the flow ends.

[0034] In step S102, the second pitch rate calculation unit 117 calculates the pitch rate from the wheel speed. Meanwhile, in step S103, the vehicle state estimation unit 112 estimates the sprung vertical velocity. In step S104, the first pitch rate calculation unit 116 calculates the pitch rate based on the sprung vertical velocity estimated in step S103, and the process proceeds to step S105.

[0035] In step S105, the estimation accuracy verification unit 118 calculates the estimation accuracy (error) based on the two pitch rates calculated in steps S102 and S104. In step S106, as an example of a predetermined threshold, if the estimation accuracy (error) is greater than 1 degree, the process proceeds to step S106, and if it is less than 1 degree, the process returns to step S101. Note that, although an error of 1 degree is used as the threshold for determining the estimation accuracy using the pitch rate, the value is not limited to this.

[0036] In step S107, the estimation accuracy verification unit 118 assigns a learning required flag to data that is determined to have poor estimation accuracy, with an estimation accuracy (error) greater than 1 degree. In step S108, the traveling data management unit 119 writes the traveling data that has been assigned a learning required flag to the traveling data storage unit 120. In step S109, the traveling data management unit 119 counts the number of learning required flags.

[0037] In step S110, the traveling data management unit 119 determines whether the number of learning required flags stored in the traveling data storage unit 120 has exceeded 1,000 counts. This 1,000 counts is the condition for transferring the data accumulated in the traveling data storage unit 120 to the learning management server 102. If the number of learning required flags exceeds 1,000 counts, the process proceeds to step S111, and the traveling data to which the learning required flags have been assigned is transferred to the learning management server 102, which is the traveling data server. Note that, although the standard for the number of learning required flags as a condition for transferring data to the learning management server 102 is set to 1,000, this value is not limitative.

[0038] (FIG. 3B) Data update of weight parameters in the vehicle 101 will now be described. In step S201, the weight data management unit 113 determines whether weight parameters have been received from the learning management server 102. If weight parameters have been received, the process proceeds to step S202; if not, step S201 is repeated.

[0039] In step S202, the weight data management unit 113 stores the acquired weight parameter data in the weight parameter storage unit 114. In step S203, the weight data management unit 113 determines whether or not storage of the weight parameters has been completed in the storage unit 114. If storage of the weight parameters has been completed, the process proceeds to step S204; if not, step S203 is repeated.

[0040] In step S204, the weight data management unit 113 switches the weight parameters, and the flow from step S201 is repeated again, and the flow proceeds to step S205. In step S205, the vehicle state estimation unit 112 estimates the vehicle state based on the weight parameters and outputs the estimation result.

[0041] In step S206, the estimation accuracy verification unit 118 determines whether there is an abnormality in the estimation output of the vehicle state estimation unit 112. If there is an abnormality, the process proceeds to step S207; if there is no abnormality, the process repeats from step S201. In step S207, the estimation accuracy verification unit 118 assigns a learning required flag to the driving data containing the abnormality. In step S208, the estimation accuracy verification unit 118 notifies the developer of the vehicle state estimation unit 112 who owns a terminal having a terminal screen 407 illustrated in FIG. 4B (described later), an inspector who performs maintenance inspections, etc., of the abnormality in the estimation output, and then repeats the process from step S201.

[0042] (FIG. 3C) The learning of weight parameters in the learning management server 102 will be described. In step S301, the learning management server 102 receives driving data from the vehicle 101. Then, in step S302, the learning management server 102 performs neural network learning. Next, in step S303, the learning management server 102 stores the weight parameters. Next, in step S304, the learning management server 102 transfers the weight parameters to the weight data management unit 113, and the flow ends.

[0043] (FIG. 3D) This is a flow chart showing how the learning management server 102 responds to the detection of an abnormality in the estimation of the vehicle state. In step S401, the learning management server 102 determines whether or not there is an abnormality notification regarding the estimation of the vehicle state from the estimation accuracy verification unit 118. If there is an abnormality notification, the process proceeds to step S402; if not, step S401 is repeated. In step S402, the learning management server 102 verifies the output of the vehicle state estimation unit 112.

[0044] In step S403, the learning management server 102 determines whether or not an abnormality has occurred in the output of the vehicle state estimation unit 112. If it is determined that an abnormality has occurred, the process proceeds to step S404, and if it is determined that an abnormality has not occurred, the process proceeds to step S406.

[0045] In step S404, the latest weight parameter data stored in the learning management server 102 is saved. Subsequently, in step S405, learning is performed again in the learning management server 102, and the flow of step S402 is repeated. In step S406, since it is determined in step S403 that no abnormality has occurred, the weight parameters are transferred from the learning management server 102 to the vehicle 101. Note that, although the re-learning has been described as saving the latest weight parameters and learning with new initial values, additional learning using the latest weight parameters as initial values ​​may also be performed.

[0046] (FIG. 4A) In a measurement device screen 401 for a driving test that visualizes data related to CAN-based data, vehicle behavior estimation results, and calculation result data for vehicle behavior physical values, the upper section displays multiple types of data 402, 403 (CAN-based data) referenced from CAN 108, the middle section displays vehicle behavior estimation results 404 output by vehicle state estimation unit 112, and the lower section displays time-series data 405 for vehicle behavior physical values. Also, a file output button 406 is provided for outputting and storing a data set as text data in a CSV file for input to a neural network, for example. The button 406 may be a physical button provided on the device, may be selected by touch operation, or may be selected by a pointing device or the like. Furthermore, the button 406 may be selected by touch operation with a finger in combination with a touch sensor function.

[0047] (FIG. 4B) Pitch rate 408 calculated by first pitch rate calculation unit 116 based on the output of the neural network, which is the output of vehicle state estimation unit 112, and pitch rate 409 calculated by second pitch rate calculation unit 117, which is not based on the output of the neural network, which is the wheel speeds of the four wheels, are presented side by side on terminal screen 407. Furthermore, accuracy determination result 410, which is displayed as a high level when the difference between the two pitch rates is equal to or greater than a predetermined value, and a low level when the difference between the two pitch rates is less than the predetermined value, is presented side by side with the two pitch rates.

[0048] (FIG. 4C) FIG. 4C is a diagram showing a case where the results of vehicle behavior estimation can be confirmed by the user through display unit 110, where the user is, for example, a driver. The display content of instrument panel 411 includes vehicle behavior estimation results 412. For example, lamp 413 is a lamp that lights up when accuracy determination result 410 is at a high level and the error is large. Display meter 414 shows a graph indicating the distance during which accuracy determination result 410 (FIG. 4B) is at a high level, or the proportion of time during which accuracy determination result 410 is at a high level, relative to the travel distance or travel time.

[0049] As a result, the estimation accuracy output by the estimation accuracy verification unit 118 is presented via the display unit 110 to users such as the developer of the vehicle state estimation unit 112, inspectors who carry terminals and perform maintenance inspections, and the driver of the vehicle 101, making it possible to check the accuracy of the suspension system to which the neural network is applied and to recognize low-accuracy states.

[0050] Second Embodiment (FIG. 5) The second embodiment focuses on the fact that the reliability of wheel speed-derived pitch rate calculations varies depending on weather and road surface conditions. Functional units that have been modified or added to the first embodiment include a weather information server 503, a GPS sensor 504, a stereo camera 505, and a wheel speed reliability determination unit 506. The wheel speed reliability determination unit 506 determines the reliability of the output of the estimation accuracy verification unit 118 based on information external to the vehicle 101, such as information from the weather information server 503 or the on-board stereo camera 505 and GPS sensor 504. In this way, by focusing on the fact that the reliability of the pitch rate changes depending on the road surface condition and by referring to the road surface condition, the aim is to extract weak points of the neural network by reflecting the reliability determination results.

[0051] (FIG. 6) The wheel speed reliability determination unit 506 has a road surface μ estimation unit 601 and a wheel speed reliability calculation unit 602. The wheel speed reliability determination unit 506 refers to information such as information from the stereo camera 505, information from the GPS sensor 504, information from the weather information server 503, and information on the condition of the tires mounted on the vehicle 101. The wheel speed reliability determination unit 506 refers to these information and inputs them to the road surface μ estimation unit 601.

[0052] The road surface μ estimating unit 601 has a function of detecting a state in which the road surface on which the vehicle is generally driven is prone to slipping, such as a state in which puddles have formed due to rainfall, or a state in which the road surface μ is low due to snow accumulation or ice, based on information from the stereo camera 505, for example. In this regard, the same detection can be performed even if the stereo camera 505 is not used, and therefore the road surface conditions may be estimated by querying the weather information server 503 using the position information of the vehicle obtained from the GPS sensor 504.

[0053] In addition, information on constantly determined factors other than the factors that change depending on the surrounding conditions as described above, such as whether the road is sandy or has a coarse-grained dirt surface, can be managed in association with a map, and whether the road surface is prone to slipping or wheel spinning can be detected based on information from the GPS sensor 504.

[0054] The wheel speed reliability determination unit 506 calculates the reliability of the wheel speed information based on the road surface information output by the road surface μ estimation unit 601, information about the vehicle 101, and information about the condition of the tires mounted on the vehicle 101 (for example, grip force), and outputs the calculation result to the estimation accuracy verification unit 118. Generally, road surface μ is about 0.8 for dry paved roads, in the range of 0.4 to 0.6 for wet paved roads, in the range of 0.2 to 0.5 for snow-covered roads, and in the range of 0.1 to 0.2 for icy roads. The total friction coefficient μ' is estimated from this road surface μ and the condition of the tires, and the physical quantity at which slippage begins is calculated by multiplying it by the weight N of the vehicle 101.

[0055] The estimation accuracy verification unit 118 (FIG. 5) verifies the estimation accuracy based on the reliability of the wheel speed information calculated by the wheel speed reliability determination unit 506, the pitch rate calculated by the first pitch rate calculation unit 116, and the pitch rate calculated by the second pitch rate calculation unit 117. If the reliability of the wheel speed information determined by the wheel speed reliability determination unit 506 is high, the processing proceeds in the same manner as in the first embodiment. However, if the reliability of the wheel speed information is low, the processing continues to estimate the vehicle state using a neural network or switches to another alternative method.

[0056] The road surface μ described above is an example, and if the tire conditions cannot be obtained, the calculation of μ' can be simplified. Furthermore, the weight N of the vehicle 101 may be calculated with the weight of the passengers and cargo taken into account in addition to the weight of the vehicle itself to improve accuracy.

[0057] (Third embodiment) (FIG. 7) Compared to the first embodiment, the suspension ECU 104 according to this embodiment newly includes a selector 703 and an estimated value verification / correction unit 705. Furthermore, vehicle state estimation units 112a to 112d are provided for each of the four wheels, and are linked to one another. The vehicle state estimation units 112a to 112d are provided as FL (front left) 112a, FR (front right) 112b, RL (rear left) 112c, and RR (rear right) 112d, respectively.

[0058] The selector 703 selects three estimation results from the four vehicle state estimation units 112a to 112d, and creates different combinations. The first pitch rate calculation unit 116 calculates four types of pitch rates (first physical values) based on the three estimation results selected by the selector 703. This is because, if the vehicle 101 is a rigid body, the first physical value (pitch rate) can be calculated from the estimation results (sprung vertical velocities) for three wheels, and therefore four types of first physical values ​​can be calculated by selecting the three estimation results.

[0059] The reason for this calculation is that if, for example, the sprung vertical velocity of only one wheel is low accuracy in the comparison made by the estimated value verification / correction unit 705, the pitch rate calculated using the sprung vertical velocities of the other three wheels excluding that one wheel is likely to approximately match the pitch rate derived from the wheel speed, and the three pitch rates including that one wheel are likely to not match the pitch rate derived from the wheel speed.In this way, the estimated value verification / correction unit 705 determines that the sprung vertical velocity estimated values ​​of the three wheels used in calculating the pitch rate that matches the wheel speed-derived pitch rate are highly accurate, transfers the estimated values ​​of the sprung vertical velocities of the corresponding three wheels to the suspension control value calculation unit 115, and excludes the remaining wheel.

[0060] Furthermore, as described above, since the pitch rate can be calculated with data for three wheels, the estimated value verification / correction unit 705 uses the correct pitch rate derived from the wheel speed and the sprung vertical speeds for the three wheels to back-calculate the sprung vertical speed of the one wheel determined to have low accuracy. The estimated value verification / correction unit 705 transfers the back-calculated low-accuracy sprung vertical speed of the one wheel to the suspension control value calculation unit 115. In other words, the estimated value verification / correction unit 705, which verifies the estimation accuracy, verifies the four types of estimation results and derives the estimation result by back-calculation.

[0061] In this way, if a high-precision pitch rate can be obtained for even one system, the estimated results of the sprung vertical velocities for the three systems used to derive it can be determined to be high-precision, and the estimated result of the sprung vertical velocity for the remaining system can be determined to be low-precision. If the sprung vertical velocity is low-precision, it can be used as the estimated result by back-calculating using the pitch rate calculated by the second pitch rate calculation unit 117 and the sprung vertical velocities for the three systems, and can be treated as training data for the neural network that realizes the vehicle state estimation unit 112 implemented in the learning management server 102.

[0062] When all four difference values ​​based on the comparison between each of the four first physical values ​​and the second physical value are smaller than a predetermined threshold, the estimation value verification / correction unit 705 adopts the four estimation results output from the four vehicle state estimation units 112 a to 112 d as the state of the suspension 109, and transfers the estimation results to the suspension control value calculation unit 115 (steps S10 and S18 in FIG. 9 , which will be described later).

[0063] The estimated value verification / correction unit 705 transfers the input data set of one of the vehicle state estimation units 112a to 112d that output the sprung vertical velocity that is determined to be low accuracy, and the sprung vertical velocity calculated by the estimated value verification / correction unit 705 to the driving data management unit 119, and the driving data management unit 119 stores the input data set and the sprung vertical velocity in the driving data storage unit 120.

[0064] When a certain number (for example, 1,000) of data sets are stored in the driving data storage unit 120, the driving data management unit 119 transfers the data sets to the learning management server 102 via the interface 111 using that count as a transfer condition.

[0065] This allows the learning management server 102 to accumulate data sets that the vehicle state estimation unit 112 is not good at, and by performing additional learning using the accumulated data, the accuracy of the vehicle state estimation units 112a to 112d can be improved.

[0066] (Fig. 8) The selector 703 will be described. In Fig. 8, the selector 703 has four phases based on a vehicle state estimation period 801, which is the operation period of the vehicle state estimation unit 112, and shows how to select three signals from an FL selection signal 802, an FR selection signal 803, an RL selection signal 804, and an RR selection signal 805. For example, if the period 801 is 20 m / sec, then the combination of three wheels to be selected will be different among the four portions of 20 m / sec.

[0067] 8, in the first period, the FL selection signal 802, the FR selection signal 803, and the RL selection signal 804 are all high, so that three types of FL, FR, and RL are selected. In the second period, the FR selection signal 803, the RL selection signal 804, and the RR selection signal 805 are all high, so that three types of FR, RL, and RR are selected. In the third period, the FL selection signal 802, the RL selection signal 804, and the RR selection signal 805 are all high, so that three types of FL, RL, and RR are selected. In the fourth period, the FL selection signal 802, the FR selection signal 803, and the RR selection signal 805 are all high, so that three types of FL, FR, and RR are selected.

[0068] In FIG. 8, the estimated value verification / correction unit 705 has one system, and the control signal is obtained when serial processing is performed. If the estimated value verification / correction unit 705 has four systems, parallel processing is generally performed.

[0069] (FIG. 9) A flowchart of the third embodiment will be described. In step S1, the counter value of the traveling data management unit 119 is set to 0 (reset). In step S2, the first pitch rate calculation unit 116 performs pitch rate calculation A using a sensor estimate calculated using the sprung vertical velocity of the wheels FL, FR, and RL. In step S3, the first pitch rate calculation unit 116 performs pitch rate calculation B using a sensor estimate calculated using the sprung vertical velocity of the wheels FR, RL, and RR. In step S4, the first pitch rate calculation unit 116 performs pitch rate calculation C using a sensor estimate calculated using the sprung vertical velocity of the wheels RL, RR, and FL. In step S5, the first pitch rate calculation unit 116 performs pitch rate calculation D using a sensor estimate calculated using the sprung vertical velocity of the wheels RR, FL, and FR.

[0070] In step S6, estimated value verification / correction unit 705 calculates the result obtained by subtracting the calculation result derived from the wheel speed calculated by second pitch rate calculation unit 117 from calculation result A calculated in step S2. In step S7, estimated value verification / correction unit 705 calculates the result obtained by subtracting the calculation result derived from the wheel speed calculated by second pitch rate calculation unit 117 from calculation result B calculated in step S3. In step S8, estimated value verification / correction unit 705 calculates the result obtained by subtracting the calculation result derived from the wheel speed calculated by second pitch rate calculation unit 117 from calculation result C calculated in step S3. In step S9, estimated value verification / correction unit 705 calculates the result obtained by subtracting the calculation result derived from the wheel speed calculated by second pitch rate calculation unit 117 from calculation result D calculated in step S5.

[0071] In step S10, the estimation value verification / correction unit 705 determines whether the difference between the four results calculated in steps S6 to S9 is smaller than a predetermined threshold. If the difference between the four calculated results is small, the process proceeds to step S18, where the estimation value verification / correction unit 705 adopts the four estimation results. Thereafter, the process returns to step S2 to continue determining whether the four estimation results can continue to be used (whether the difference is smaller than the predetermined threshold).

[0072] If the difference between the four calculated results is large in step S10, the process proceeds to step S11. In step S11, the estimate verification / correction unit 705 determines whether the difference between the calculation results calculated in step S6 is smaller than a predetermined value. If the difference is not smaller than a predetermined threshold, the estimate verification / correction unit 705 determines whether the difference between the calculation results calculated in step S7 is smaller than a predetermined threshold in step S12. If the difference is not smaller than a predetermined value, the estimate verification / correction unit 705 determines whether the difference between the calculation results calculated in step S8 is smaller than a predetermined threshold in step S13. If the difference is not smaller than the predetermined threshold, the estimate verification / correction unit 705 determines whether the difference between the calculation results D calculated in step S9 is smaller than a predetermined threshold in step S14. If the difference is not smaller than the predetermined threshold in step S14, the process returns to step S2.

[0073] In step S15, if the difference calculated in step S6 is smaller than a predetermined threshold value in step S11, the sensor estimated value calculated using the sprung vertical velocities of the wheels FL, FR, and RL in pitch rate calculation A in step S2 is adopted. In step S16, the sprung vertical velocity of the remaining wheel RR is calculated based on this, and in step S17, the calculated sprung vertical velocity of the wheel RR is stored in a data set. Note that steps S18 to S20, which are the calculation flow for the wheel FL, steps S21 to S23, which are the calculation flow for the wheel FR, and steps S24 to S26, which are the calculation flow for the wheel RL, are similar to the flow from step S15 to step S17, and therefore description thereof will be omitted.

[0074] In step S27, the traveling data storage unit 120 performs counter processing. If in step S28 the counter processing number in step S27 exceeds 1000, in step S29 the traveling data storage unit 120 sends an updated data set to the learning management server 102, and in step S30 the data set in the traveling data storage unit 120 is erased, and the process returns to step S1 and repeats the above-mentioned flow. If in step S28 the counter processing number does not exceed 1000, the process returns to step S2 and repeats the above-mentioned flow.

[0075] (Fourth embodiment) (FIG. 10) The fourth embodiment has a configuration in which relearning (parameter updating) can be performed in a virtual environment only if there are only good / bad flags for the driving data. In the above-mentioned embodiment, the estimation accuracy of the sprung vertical velocity and piston velocity output by the vehicle state estimation unit 112 while driving is improved, whereas the fourth embodiment reproduces a driving environment in a virtual environment and improves the estimation accuracy of the sprung vertical velocity and piston velocity in the virtual environment.

[0076] Specifically, the accuracy state is combined with road information (including road profiles) acquired by existing sensors such as the GPS sensor 504 and the stereo camera 604, and transferred to the learning management server 102, where it is reproduced in a simulation to create training data, thereby carrying out learning.

[0077] This embodiment is characterized by the addition of a vehicle behavior simulator 1103 to the learning management server 102 in the second embodiment, and further includes a map information server 1104 used to reproduce realistic terrain and road shapes in a virtual environment. The map information handled by the map information server 1104 may be, for example, the open-source OSM (Open Street Map), or other paid or free map data. This map data may include elevation / gradient information, and as data collection progresses, information such as road profiles may be applied.

[0078] Furthermore, adding information from the weather information server 503 to the map information server 1104 improves the reproducibility of the real environment in the virtual environment. Therefore, the learning management server 102 learns weighting parameters based on second sensor information acquired by a second sensor (weather information server 503, map information server 1104) mounted on the vehicle 101 and virtual environment information learned outside the vehicle 101. In this way, by utilizing the vehicle behavior simulator 1103 of the learning management server 102 in an advanced virtual environment, it becomes possible to improve the estimation accuracy of the sprung vertical velocity and piston velocity output by the vehicle state estimation unit 112, even if measurements are not necessarily taken of an actual vehicle.

[0079] According to the embodiment of the present invention described above, the following advantageous effects are achieved.

[0080] (1) A suspension system that performs machine learning calculations by associating sensor information acquired from a plurality of sensors 106 provided on a vehicle 101 with state information on a suspension 109 provided on the vehicle 101 includes: a weight parameter storage unit 114 that stores weight parameters calculated by the machine learning; and a vehicle state estimation unit 112 that outputs an estimation result of the state of the suspension 109 based on the sensor information and the weight parameters. The suspension system also includes: a first vehicle behavior calculation unit 116 that calculates a first physical value of a physical quantity related to the behavior of the vehicle 101 based on the output estimation result; a second vehicle behavior calculation unit 117 that calculates a second physical value of the physical quantity based on the sensor information; and an estimation accuracy verification unit 118 that compares the first physical value with the second physical value to output the estimation accuracy of the state of the suspension 109 by the vehicle state estimation unit 112. The suspension system also includes a traveling data management unit 119 that instructs a learning management server 102 to learn the weight parameters based on the output result of the estimation accuracy by the estimation accuracy verification unit 118. By doing this, it is possible to extract weak points of the neural network in a vehicle that is not equipped with a suspension sensor, and to support learning.

[0081] (2) The vehicle 101 has four vehicle state estimation units 112 linked to the four wheels of the vehicle 101, respectively, and a selector 703 that selects three of the four estimation results output from the four vehicle state estimation units 112. In this way, the vehicle state estimation unit 112 can estimate the state of the suspension 109 using only the estimation results for the three wheels.

[0082] (3) The first vehicle behavior calculation unit 116 calculates four types of first physical values ​​from the three estimation results selected by the selector 703, and the estimation accuracy verification unit 118 adopts the four estimation results output from the four vehicle state estimation units 112 as the state of the suspension if four difference values ​​based on comparison of each of the four types of first physical values ​​with the second physical value are smaller than a predetermined threshold value. The estimation accuracy verification unit 118 adopts the three estimation results selected by the selector 703 for calculating the first physical value as the state of the suspension 109 if the difference value based on comparison of one of the four calculated first physical values ​​with the second physical value is smaller than the predetermined threshold value. The first vehicle behavior calculation unit 116 recalculates the four types of first physical values ​​from the three estimation results selected by the selector 703 if the difference value based on comparison of the four calculated first physical values ​​with the second physical value is larger than the predetermined threshold value. This makes it possible to determine the estimation accuracy taking into account low-accuracy estimation results of the vehicle state.

[0083] (4) The estimation accuracy verification unit 118 calculates the estimation results not selected by the selector 703 using the three estimation results selected by the selector 703 and the second physical value. In this way, it is possible to calculate the first physical value for the estimation results not selected by the selector 703.

[0084] (5) The estimation accuracy output by the estimation accuracy verification unit 118 is presented to the user via the display unit 110. In this way, the estimation accuracy of the state of the suspension 109 can be visualized.

[0085] (6) The first physical value and the second physical value are the pitch rate or the roll rate of the vehicle 101. In this way, the first physical value and the second physical value can be calculated based on a plurality of vehicle behavior data.

[0086] (7) The learning management server 102 learns the weighting parameters based on second sensor information acquired by a second sensor mounted on the vehicle 101 and virtual environment information learned outside the vehicle 101. In this way, the estimation accuracy of the sprung vertical velocity and piston velocity in the virtual environment is improved.

[0087] (8) The learning management server 102 learns the weight parameters based on a computer installed in the vehicle 101. In this way, the weight parameters can be updated without installing the learning management server 102 on the vehicle 101 side.

[0088] The present invention is not limited to the above-described embodiments, and various modifications and combinations of other configurations are possible without departing from the spirit of the present invention. Furthermore, the present invention is not limited to those having all of the configurations described in the above-described embodiments, and includes those in which some of the configurations are omitted.

[0089] REFERENCE SIGNS LIST 101 Vehicle 102 Learning management server 103 Network 104 Suspension ECU 105 Vehicle ECU 106 Sensor 107 Sensor ECU 108 CAN 109 Active suspension 110 Display unit 111 Interface (I / F) 112 Vehicle state estimation unit 113 Weight data management unit 114 Weight parameter storage unit (memory unit) 115 Suspension control value calculation unit 116 First pitch rate calculation unit 117 Second pitch rate calculation unit 118 Estimation accuracy verification unit 119 Traveling data management unit 120 Traveling data storage unit 201 Neural network input layer element group 202 Neural network hidden layer element group 203 Neural network output layer element 204 Time series data 205 Sensor data 206 Window 207 Sensor data instantaneous value 401 Measuring device screen 402 Wheel speed data (CAN system data) 403 Longitudinal acceleration data (CAN system data) 404 Piston speed data (vehicle state estimation value) 405 Pitch rate data (vehicle behavior physical value) 406 File output button 407 Terminal screen 408 Pitch rate derived from neural network 409 Pitch rate derived from non-neural network 410 Accuracy judgment result 411 Instrument panel 412 Vehicle behavior estimation result 413 Judgment result display lamp 414 Odometer 503 Weather information server 504 GPS sensor 505 Stereo camera 506 Wheel speed reliability judgment unit 601 Road surface μ estimation unit 703 Selector 705 Estimated value verification / correction unit 801 Vehicle state estimation period 802 to 805 Selection signal 1103 Vehicle behavior simulator 1104 Map information server

Claims

1. A suspension system that performs machine learning calculations by correlating sensor information obtained from multiple sensors equipped on a vehicle with state information of a suspension equipped on the vehicle, the suspension system comprising: a weight parameter storage unit that stores weight parameters calculated by the machine learning; a vehicle state estimation unit that outputs an estimation result of the suspension state based on the sensor information and the weight parameters; a first vehicle behavior calculation unit that calculates a first physical value of a physical quantity related to the behavior of the vehicle based on the output estimation result; a second vehicle behavior calculation unit that calculates a second physical value of the physical quantity based on the sensor information; an estimation accuracy verification unit that compares the first physical value with the second physical value to output the estimation accuracy of the suspension state by the vehicle state estimation unit; and a driving data management unit that instructs a learning management server to learn the weight parameters based on the output result of the estimation accuracy by the estimation accuracy verification unit.

2. A suspension system according to claim 1, wherein the vehicle has four vehicle state estimating units linked to four wheels of the vehicle, respectively, and the suspension system is provided with a selector that selects three of the four estimation results output from the four vehicle state estimating units.

3. A suspension system according to claim 2, wherein the first vehicle behavior calculation unit calculates four types of first physical values ​​from the three estimation results selected by the selector; the estimation accuracy verification unit, if four difference values ​​based on comparison between each of the four types of first physical values ​​and the second physical value are smaller than a predetermined threshold, adopts the four estimation results output from the four vehicle state estimation units as the state of the suspension; the estimation accuracy verification unit, if the difference value based on comparison between the second physical value and any one of the four calculated first physical values ​​is smaller than a predetermined threshold, adopts the three estimation results selected by the selector for calculating the first physical value as the state of the suspension; and the first vehicle behavior calculation unit, if the difference value based on comparison between the second physical value and any one of the calculated four types of first physical values ​​is larger than a predetermined threshold, recalculates the four types of first physical values ​​from the three estimation results selected by the selector.

4. A suspension system according to claim 3, wherein the estimation accuracy verification unit calculates the estimation results not selected by the selector using the three estimation results selected by the selector and the second physical value.

5. A suspension system according to claim 1, wherein the estimation accuracy output by the estimation accuracy verification unit is presented to a user via a display unit.

6. A suspension system according to claim 1, wherein the first physical value and the second physical value are pitch rate or roll rate of the vehicle.

7. A suspension system according to claim 1, wherein the learning management server learns the weight parameters based on second sensor information acquired by a second sensor mounted on the vehicle and virtual environment information learned outside the vehicle.

8. A suspension system according to claim 4, wherein the learning management server learns the weight parameters based on a computer installed in the vehicle.

Images