JP7748895B2 - Suspension Control Device - Google Patents

Description

This disclosure relates to a suspension control device that controls a vehicle suspension.

Conventional suspension control involves detecting or estimating the vehicle state and performing feedback control accordingly (see Patent Document 1). Feedback control uses, for example, the Skyhook control law or BLQ (Bi-linear Optimal Control). Patent Document 1 also discloses a method for achieving optimal control in real time without step-by-step optimization by having AI (artificial intelligence) learn the direct optimal control commands and vehicle state in advance, and then calculating commands using only the weighting coefficients of the learning results.

Japanese Patent Application Laid-Open No. 2021-109517

In the suspension control device disclosed in Patent Document 1, data measured during driving tests, etc., is learned by AI either theoretically or in real time, improving the accuracy of optimal control commands. In this case, the learning process is biased toward patterns that are frequently included, resulting in higher accuracy of the optimal control commands generated by the AI. The measured data includes a mixture of data on road surfaces with large inputs to the vehicle and data on flat road surfaces. On flat road surfaces, there is no need to actively control the suspension unless there is vehicle input such as roll. However, if all measured data is used for learning, there is a risk that the proportion of patterns that require active suspension control will be relatively low. In this case, there is an issue of reduced accuracy of the optimal control commands generated by the AI for patterns required for control.

The objective of one embodiment of the present invention is to provide a suspension control system that can effectively extract learning data and improve the accuracy of AI output.

One embodiment of the present invention is a suspension control device that is installed between a vehicle body and a wheel and controls a force generating mechanism that can adjust the force between the vehicle body and the wheel. The device has a vehicle state quantity acquisition unit that detects or estimates a state quantity of the vehicle, and an AI learning unit that learns a command value for the force generating mechanism based on the results acquired by the vehicle state quantity acquisition unit. The AI learning unit has a function of removing data whose absolute value does not exceed a threshold, based on the absolute value of a physical quantity that is highly correlated with whether or not the force generating mechanism needs to operate.

Another embodiment of the present invention is a suspension control device that is installed between a vehicle body and a wheel and controls a force generating mechanism that can adjust the force between the vehicle body and the wheel, and includes a vehicle state quantity learning unit that estimates a state quantity of the vehicle, and an AI learning unit that learns a command value for the force generating mechanism based on the estimation result of the vehicle state quantity learning unit, wherein the AI learning unit has a function of removing data whose absolute value does not exceed a threshold based on the absolute value of the state quantity that is highly correlated with whether or not the force generating mechanism needs to operate, and the vehicle state quantity learning unit learns the state quantity based on physical quantities related to the operation of the vehicle using the data removed by the AI learning unit.

According to one embodiment of the present invention, learning data can be effectively extracted, improving the accuracy of AI output.

FIG. 1 is a diagram schematically illustrating a suspension control device according to first to third embodiments. FIG. 10 is an explanatory diagram showing a procedure for learning the DNN of the controller. FIG. 2 is a block diagram illustrating an AI learning unit according to the first embodiment. FIG. 10 is a block diagram showing an AI learning unit according to a second embodiment. FIG. 10 is a block diagram showing an AI learning unit according to a third embodiment. FIG. 10 is a diagram schematically illustrating a suspension control device according to a fourth embodiment. FIG. 10 is a block diagram showing an AI learning unit according to a fourth embodiment.

The suspension control device according to an embodiment of the present invention will be described in detail below with reference to the accompanying drawings, taking as an example a case in which it is applied to a four-wheeled vehicle.

Figures 1 to 3 show a first embodiment of the present invention. In Figure 1, for example, left and right front wheels and left and right rear wheels (hereinafter collectively referred to as wheels 2) are provided on the underside of a vehicle body 1 that constitutes the body of a vehicle. These wheels 2 are configured to include tires 3. The tires 3 act as springs that absorb small irregularities in the road surface.

The suspension device 4 is installed between the vehicle body 1 and the wheels 2. The suspension device 4 is composed of a suspension spring 5 (hereinafter referred to as spring 5) and an adjustable damping shock absorber (hereinafter referred to as variable damper 6) installed in parallel with the spring 5 between the vehicle body 1 and the wheels 2. Note that Figure 1 schematically illustrates a case in which one set of suspension devices 4 is installed between the vehicle body 1 and the wheels 2. In the case of a four-wheeled vehicle, a total of four sets of suspension devices 4 are installed individually and independently between the four wheels 2 and the vehicle body 1.

Here, the variable damper 6 of the suspension device 4 is a force generation mechanism that generates an adjustable force between the vehicle body 1 side and the wheel 2 side. The variable damper 6 is configured using a hydraulic shock absorber with adjustable damping force. The variable damper 6 is equipped with a variable damping force actuator 7 consisting of a damping force adjustment valve or the like to continuously adjust the characteristics of the generated damping force (i.e., the damping force characteristics) from hard characteristics to soft characteristics. Note that the variable damping force actuator 7 does not necessarily have to be configured to continuously adjust the damping force characteristics; it may also be capable of adjusting the damping force in multiple stages, for example, two or more stages. The variable damper 6 may also be a pressure control type or a flow control type.

The sprung acceleration sensor 8 detects the vertical acceleration of the vehicle body 1 (sprung mass). The sprung acceleration sensor 8 is provided at any position on the vehicle body 1. For example, the sprung acceleration sensor 8 is attached to the vehicle body 1 at a position near the variable damper 6. The sprung acceleration sensor 8 detects vertical vibration acceleration on the vehicle body 1 side, which is the so-called sprung mass side, and outputs the detection signal to the electronic control unit 11 (hereinafter referred to as ECU 11).

The vehicle height sensor 9 detects the height of the vehicle body 1. Multiple vehicle height sensors 9 (e.g., four) are provided on the vehicle body 1, for example, on the sprung side, corresponding to each wheel 2. That is, each vehicle height sensor 9 detects the relative position (height position) of the vehicle body 1 with respect to each wheel 2 and outputs the detection signal to the ECU 11. The vehicle height sensor 9 and the sprung acceleration sensor 8 constitute a vehicle state quantity acquisition unit that detects vehicle state quantities. Note that the vehicle state quantities are not limited to the vertical acceleration of the vehicle body 1 and the height of the vehicle body 1. The vehicle state quantities may include, for example, a relative velocity obtained by differentiating the height (vehicle height) of the vehicle body 1, and a vertical velocity obtained by integrating the vertical acceleration of the vehicle body 1. In this case, in addition to the vehicle height sensor 9 and the sprung acceleration sensor 8, the vehicle state quantity acquisition unit includes a differentiator that differentiates the vehicle height, an integrator that integrates the vertical acceleration, and the like.

The road surface measurement sensor 10 constitutes a road surface profile acquisition unit that detects road surface profiles as road surface information. The road surface measurement sensor 10 is composed of, for example, multiple millimeter-wave radars. The road surface measurement sensor 10 measures and detects the road surface conditions ahead of the vehicle (specifically, including the distance and angle to the road surface to be detected, as well as the screen position and distance). The road surface measurement sensor 10 outputs a road surface profile based on the detected road surface values.

The road surface measurement sensor 10 may be, for example, a combination of millimeter-wave radar and a mono camera, or may be configured as a stereo camera including a pair of left and right image sensors (digital cameras, etc.), as described in JP 2011-138244 A, etc. The road surface measurement sensor 10 may also be configured as an ultrasonic distance sensor, etc.

The ECU 11 is a control device that controls vehicle behavior, including vehicle attitude control. The ECU 11 is mounted on the vehicle body 1. The ECU 11 is configured using, for example, a microcomputer. The ECU 11 has a memory 11A that can store data. The ECU 11 is equipped with a controller 12.

The input side of the ECU 11 is connected to the sprung acceleration sensor 8, vehicle height sensor 9, and road surface measurement sensor 10, and the output side is connected to the variable damping force actuator 7 of the variable damper 6. The ECU 11 outputs the road surface profile and vehicle state quantities to the controller 12 based on the vertical vibration acceleration detected by the sprung acceleration sensor 8, the vehicle height detected by the vehicle height sensor 9, and the road surface detected by the road surface measurement sensor 10. The controller 12 calculates the force to be generated by the variable damper 6 (force generation mechanism) of the suspension unit 4 based on the road surface profile and vehicle state quantities, and outputs a command signal to the variable damping force actuator 7 of the suspension unit 4.

The ECU 11 stores vehicle state variables and road surface input data in memory 11A over a period of several seconds, for example, when the vehicle has traveled approximately 10 to 20 meters. This allows the ECU 11 to generate time-series data of road surface inputs (road surface profile) and time-series data of vehicle state variables when the vehicle has traveled a specified distance. The controller 12 controls the variable damper 6 to adjust the damping force to be generated based on the road surface profile and the time-series data of vehicle state variables.

The controller 12 includes a trained deep neural network (DNN) 13 that constitutes the AI. The DNN 13 is part of the AI learning unit 23 and is configured as a multi-layer neural network, for example, with four or more layers. Each layer includes multiple neurons, and neurons in two adjacent layers are connected by weighting coefficients. The weighting coefficients are set through prior training. The controller 12 acquires time-series data of road surface input (road surface profile) and time-series data of vehicle state quantities based on the vertical vibration acceleration detected by the sprung acceleration sensor 8, the vehicle height detected by the vehicle height sensor 9, and the road surface detected by the road surface measurement sensor 10. The controller 12 outputs time-series data of optimal command values based on the time-series data of road surface input and the time-series data of vehicle state quantities. At this time, the latest optimal command value corresponds to the current optimal damping force command value. As a result, the controller 12 outputs the most appropriate damping force command value for the current vehicle and road surface. The damping force command value corresponds to the current value used to drive the damping force variable actuator 7.

Next, the learning method of the DNN 13 of the controller 12 will be explained with reference to the explanatory diagram shown in Figure 2. The DNN 13 is constructed by executing the following processes: (1) direct optimal control command value search, (2) command value learning, and (3) weighting coefficient download.

First, to perform a direct optimal control command value search, an analytical model 20 including a vehicle model 21 is constructed. The analytical model 20 constitutes a vehicle state quantity acquisition unit that estimates the vehicle state quantities. Figure 2 illustrates an example in which the vehicle model 21 is a single-wheel model. The vehicle model 21 may be, for example, a pair of two-wheel models (left and right), or a four-wheel model (full vehicle model). Road surface input and an optimal command value from the direct optimal control unit 22 are input to the vehicle model 21. The direct optimal control unit 22 finds the optimal command value according to the following procedure for direct optimal control command value search.

(1) Direct Optimal Control Command Value Search The direct optimal control unit 22 searches for an optimal command value by repeated calculations using an analytical model 20 including a vehicle model 21 in advance. The search for the optimal command value is formulated as the optimal control problem shown below, and is obtained numerically using an optimization method.

The motion of the target vehicle is expressed by the equation of state in Equation 1. Note that the dot in the equation represents the first derivative with respect to time t (d/dt).

Here, x is the state variable and u is the control input. The initial condition of the state equation is given by equation 2.

The equality constraints and inequality constraints imposed from the initial time _t0 to the terminal time _tf are expressed as in Equation 3 and Equation 4.

The optimal control problem is the problem of finding a control input u(t) that minimizes the evaluation function J shown in equation 5 while satisfying the state equation shown in equation 1, the initial condition shown in equation 2, and the constraints shown in equations 3 and 4.

Solving an optimal control problem with constraints like the one above is extremely difficult. For this reason, we use a direct optimization method that can easily handle constraints. This method converts the optimal control problem into a parameter optimization problem and uses an optimization method to obtain a solution.

In order to convert the optimal control problem into a parameter optimization problem, the period from the initial time _t0 to the terminal time _tf is divided into N intervals. If the terminal times of each interval are represented as _t1 , _t2 , ..., _tN , the relationship between them is as shown in Equation 6.

The continuous input u(t) is replaced by a discrete value u _i at the end time of each interval, as shown in Equation 7.

The state equation is numerically integrated from the initial condition _x0 for the inputs _u0 , _u1 , ..., _uN to find the state quantities _x1 , _x2 , ..., _xN at the end time of each interval. At this time, the input within each interval is found by linearly interpolating the input given at the end time of each interval. As a result of the above, the state quantities are determined for the inputs, and this expresses the evaluation function and constraint conditions. Therefore, the converted parameter optimization problem can be expressed as follows:

If the parameters to be optimized are collectively referred to as X, it becomes as shown in equation 8.

Therefore, the evaluation function shown in equation 5 can be expressed as in equation 9.

Furthermore, the constraints shown in equations 3 and 4 can be expressed as equations 10 and 11.

In this way, the optimal control problem described above can be converted into a parameter optimization problem expressed by equations 8 to 11.

The evaluation function J for formulating the problem of finding an optimal control command according to the road surface as an optimal control problem is defined as in Equation ₁₂ , so that the vertical acceleration Az is minimized to provide a comfortable ride and the control command value u is small. Here, _q1 and _q2 are weighting coefficients, which are set in advance based on, for example _, experimental results.

The direct optimization control unit 22 numerically analyzes the parameter optimization problem formulated in this way using an optimization method, and derives optimal command values for various road surfaces.

(2) Command Value Learning The AI learning unit 23 includes a DNN 24. The AI learning unit 23 outputs an optimal command value derived by direct optimal control command value search, and inputs the road surface profile and vehicle state quantities at that time, causing the DNN 24, which serves as artificial intelligence, to learn various road surface inputs and outputs. The DNN 24 is a deep neural network for learning, and has the same configuration as the on-board DNN 13. The DNN 24 receives time-series data of road surface inputs and time-series data of vehicle state quantities as road surface profiles. At this time, weighting coefficients between neurons in the DNN 24 are calculated using the time-series data of the optimal command value corresponding to the road surface input and the vehicle state quantities as training data. Note that data extracted by a learning data extraction unit 25 (described later) (time-series data of road surface inputs, time-series data of vehicle state quantities, and time-series data of the optimal command value) are used for learning the DNN 24.

(3) Weighting Coefficient Download The weighting coefficients of the DNN 24 learned by the command value learning are set in the DNN 13, which is the command value determination unit of the actual ECU 11. In this way, the DNN 13 of the controller 12 is configured.

(4) Calculation of Optimal Command Value A controller 12 including a DNN 13 is mounted on a vehicle. The input side of the controller 12 is connected to the sprung acceleration sensor 8, the vehicle height sensor 9, and the road surface measurement sensor 10. The output side of the controller 12 is connected to the variable damping force actuator 7 of the variable damper 6. The controller 12 acquires road surface inputs and vehicle state quantities based on the detection signals of the sprung acceleration sensor 8, the vehicle height sensor 9, and the road surface measurement sensor 10. The controller 12 inputs time-series data of the road surface inputs and time-series data of the vehicle state quantities as a road surface profile to the DNN 13. When the time-series data of the road surface inputs and the vehicle state quantities are input, the DNN 13 outputs a command value for the variable damper 6 that becomes an optimal command in accordance with the learning results.

In this way, the direct optimization control unit 22 derives direct optimal control commands under various conditions through offline numerical optimization. The artificial intelligence (DNN 24) learns the road surface profile, vehicle state variables, and optimal commands used in this process. As a result, direct optimal control can be achieved by the controller 12 (ECU 11) equipped with the DNN 13, without performing optimization for each step.

Next, the data extraction process performed by the learning data extraction unit 25 will be described. First, data on road surface inputs, vehicle state variables, and optimal control commands to be used for learning is accumulated. This data may be calculated from road surface input data using the analysis model 20 (vehicle model 21) and the direct optimal control unit 22, for example, or may be data acquired by sensors (sprung acceleration sensor 8, vehicle height sensor 9, road surface measurement sensor 10) while the vehicle is actually running. When data measured by sensors is used, the sensors constitute a vehicle state variable acquisition unit that detects the vehicle state variables.

From this data, the correlation is found between the sprung speed included in the vehicle state quantities and the control variable of the optimal control command, which is the learned value. A constant value is determined as the threshold for the absolute value of the sprung speed, within a range where the control variable does not increase. This constant value is a threshold for the amplitude of the sprung speed. This threshold is set from the upper limit side. When setting from the lower limit side, the threshold is determined so that data from when the vehicle is traveling straight on a flat road can be excluded.

Next, data (detected values) measured by the sprung acceleration sensor 8, vehicle height sensor 9, and road surface measurement sensor 10 during a vehicle driving test or the like are input to the AI learning unit 23. The AI learning unit 23 then directly calculates the optimal command value (control amount) using the optimal control unit 22.

Meanwhile, the learning data extraction unit 25 acquires the sprung velocity based on, for example, the sprung acceleration from the sprung acceleration sensor 8. The learning data extraction unit 25 uses the sprung velocity corresponding to the detected value from the sensor to determine whether to extract the acquired data as data for AI learning. Specifically, the learning data extraction unit 25 determines whether to extract the data as data for AI learning based on the sprung velocity as a physical quantity.

Sprung speed has a high correlation with the control variable (optimal command value). For this reason, if the absolute value of the sprung speed is equal to or greater than a predetermined constant value (threshold), it is determined that the data should be extracted and used for learning. For learning by the DNN24, data where the absolute value of the sprung speed is equal to or greater than the constant value (threshold) is used. On the other hand, for learning by the DNN24, data where the absolute value of the sprung speed is less than the constant value is not used.

In this case, the value of the sprung velocity (physical quantity) is a value calculated using the detection value of the sprung acceleration sensor 8. However, the value of the sprung velocity may also be measured directly using a sprung velocity sensor. The value of the sprung velocity may also be an estimated value estimated from road surface input using the analysis model 20 (vehicle model 21). In other words, the data of the vehicle state quantity before extraction is not limited to data acquired during a driving test, but may also be calculated by the analysis model 20 based on the road surface input.

Furthermore, the data extraction determination result by the learning data extraction unit 25 may be used not only to determine whether or not to use the data for learning by the DNN 24, but also to determine whether or not to calculate optimal command values (control variables) for the measured data directly using the optimal control unit 22. In this case, there is no need to calculate optimal command values for data that was not extracted, so the calculation of unnecessary optimal command values can be avoided, thereby shortening the learning time. Furthermore, the data extraction determination result by the learning data extraction unit 25 may be used to determine whether or not to save data being measured in real time, or whether or not to input it to the DNN when performing real-time learning of the DNN. Furthermore, when training a DNN offline using data collected online, the data extraction determination result by the learning data extraction unit 25 may be used to determine whether or not to send the measurement data to a server.

Thus, according to this embodiment, the suspension control device controls a variable damper 6 (force generating mechanism) that is installed between the vehicle body 1 and the wheel 2 and can adjust the force between the vehicle body 1 and the wheel 2. It includes a sprung acceleration sensor 8 and a vehicle height sensor 9 as a vehicle state quantity acquisition unit that detects vehicle state quantities, and an AI learning unit 23 that learns command values for the variable damper 6 based on the results acquired by the sprung acceleration sensor 8 and the vehicle height sensor 9. The AI learning unit 23 includes a learning data extraction unit 25 that removes data in which the sprung velocity does not exceed a threshold value, based on the absolute value of the sprung velocity (physical quantity) that is highly correlated with whether the variable damper 6 needs to operate.

As a result, the AI learning unit 23 sets a threshold for data extraction by focusing on physical quantities that can be measured by sensors or the like while driving, and excludes patterns from the learning data that do not require active control of the suspension device 4. As a result, the DNN 24 can learn by removing data that has a low correlation with the control amount of the optimal control command. As a result, learning data for the DNN 24 can be effectively extracted, and the estimation accuracy of the DNN 13 based on the learning results of the DNN 24 can be improved.

In addition, since sprung speed is the main input when calculating control commands, it has a strong correlation with the command value. However, there is a range of insensitivity to sprung speed, for example, due to filtering. In response to this, the learning data extraction unit 25 determines whether to extract data based on the magnitude of the amplitude of the sprung speed as the absolute value of the sprung speed. This makes it possible to remove data in the range of insensitivity to sprung speed. This makes it easy to extract data after determining the threshold value for sprung speed. As a result, it can be easily applied even when performing real-time learning on the DNN 13 using data acquired during driving.

In the first embodiment, the same DNN 13 is used for all four wheels. However, the present invention is not limited to this. Different DNN weights may be set for the front and rear wheels, and the variable dampers may be controlled independently for each of the four wheels.

Next, Figure 4 shows a second embodiment. A feature of the second embodiment is that when the rate of change in the relative velocity between the sprung and unsprung masses meets a predetermined condition, the AI learning unit performs learning using data in which the absolute value of the relative velocity exceeds a threshold, and removes other data. Note that in the second embodiment, the same components as those in the first embodiment described above are assigned the same reference numerals, and their description will be omitted.

The AI learning unit 31 trains the DNN 24 using data extracted by the learning data extraction unit 32 according to the second embodiment (time series data of road surface input, time series data of vehicle state quantities, and time series data of optimal command values). Therefore, the data extraction process performed by the learning data extraction unit 32 will be described below.

First, data on road surface inputs, vehicle state variables, and optimal control commands to be used for learning is accumulated. The learning data extraction unit 32 is equipped with a bandpass filter 33 (BPF). Of this data, the bandpass filter 33 passes signals in the frequency band to which control should be applied for the relative speed between sprung and unsprung parts, which is included in the vehicle state variables, and attenuates signals in other frequency bands. In this case, the passband of the bandpass filter 33 is set to a frequency band where control increases, such as around 1 Hz. The passband of the bandpass filter 33 is determined in advance by analyzing driving data and examining frequency bands where control variables increase.

The correlation between the relative speed signal that has passed through the band-pass filter 33 and the control variable of the optimal control command, which is the learning value, is calculated. A fixed value that serves as the threshold for the relative speed is determined within a range that does not increase the control variable. This fixed value is the threshold for the amplitude of the relative speed. The fixed value (threshold) may be determined to a value that allows data from when the vehicle is traveling straight on a flat road to be excluded.

Next, data (detected values) measured by the sprung acceleration sensor 8, vehicle height sensor 9, and road surface measurement sensor 10 during a vehicle driving test or the like are input to the AI learning unit 31. The AI learning unit 31 then directly calculates the optimal command value (control amount) using the optimal control unit 22.

Meanwhile, the learning data extraction unit 32 acquires the relative velocity based on, for example, the detection values of the sprung acceleration sensor 8 and the vehicle height sensor 9. The learning data extraction unit 32 uses the relative velocity corresponding to the detection value from the sensor to determine whether to extract the acquired data as data for AI learning. Specifically, the learning data extraction unit 32 extracts the signal in the passband of the bandpass filter 33 from the relative velocity as a physical quantity. Then, if the magnitude of the amplitude of the relative velocity that has passed through the bandpass filter 33 is equal to or greater than a predetermined constant value (threshold), the learning data extraction unit 32 determines to extract the data and use it for learning. For learning by the DNN 24, if the absolute value (magnitude) of the passband component of the relative velocity is equal to or greater than a predetermined value (threshold), the data at that time is used. On the other hand, other data is not used for learning by the DNN 24.

In this case, the value of the relative velocity (physical quantity) is calculated using the detection values of the sprung acceleration sensor 8 and the vehicle height sensor 9. Alternatively, the value of the relative velocity may be measured directly, for example, using a stroke sensor equipped with a variable damper. The value of the relative velocity may also be an estimated value estimated from road surface input using the analysis model 20 (vehicle model 21). In other words, the data of the vehicle state quantity before extraction is not limited to data acquired in a driving test, but may also be calculated by the analysis model 20 based on the road surface input. Furthermore, the determination result of data extraction by the learning data extraction unit 32 is not limited to whether or not to use the data for learning by the DNN 24, but may also be used to determine whether or not to calculate an optimal command value (control quantity) directly using the optimization control unit 22 for the measured data.

Thus, the second embodiment can achieve substantially the same effects as the first embodiment. In the case of relative velocity, the command value (control amount) increases or decreases depending on the frequency, which is the rate of change of the relative velocity. In contrast, in the second embodiment, the AI learning unit 31 performs learning using data in which the amplitude (absolute value) of the relative velocity exceeds a threshold when the frequency (rate of change) of the relative velocity satisfies a predetermined condition. Therefore, the DNN 24 can be trained using data in a range where the sensitivity of the relative velocity is high.

Next, Figure 5 shows a third embodiment. A feature of the third embodiment is that when the rate of change of sprung acceleration satisfies a predetermined condition, the AI learning unit performs learning using data in which the absolute value of sprung acceleration exceeds a threshold value, and removes other data. Note that in the third embodiment, the same components as those in the first embodiment described above are assigned the same reference numerals, and their description will be omitted.

The AI learning unit 41 trains the DNN 24 using data extracted by the learning data extraction unit 42 according to the third embodiment (time series data of road surface input, time series data of vehicle state quantities, and time series data of optimal command values). Therefore, the data extraction process performed by the learning data extraction unit 42 will be described below.

First, data on road surface inputs, vehicle state variables, and optimal control commands to be used for learning is accumulated. The learning data extraction unit 42 is equipped with a bandpass filter 43 (BPF). Of this data, the bandpass filter 43 passes signals in the frequency band to which control should be applied for the sprung acceleration included in the vehicle state variables, and attenuates signals in other frequency bands. At this time, the passband of the bandpass filter 43 is set to a frequency band where control increases, such as around 1 Hz. The passband of the bandpass filter 43 is determined in advance by analyzing driving data and examining frequency bands where control variables increase.

The correlation between the sprung acceleration signal that has passed through the band-pass filter 43 and the control variable of the optimal control command, which is the learning value, is calculated. A fixed value that serves as the threshold for sprung acceleration is determined within a range that does not increase the control variable. This fixed value is the threshold for the amplitude of the sprung acceleration. The fixed value (threshold) may be determined to a value that allows data from when the vehicle is traveling straight on a flat road to be excluded.

Next, data (detected values) measured by the sprung acceleration sensor 8, vehicle height sensor 9, and road surface measurement sensor 10 during a vehicle driving test or the like are input to the AI learning unit 41. The AI learning unit 41 directly calculates the optimal command value (control amount) using the optimal control unit 22.

Meanwhile, the learning data extraction unit 42 acquires the sprung acceleration, for example, as a detection value from the sprung acceleration sensor 8. The learning data extraction unit 42 uses the sprung acceleration corresponding to the detection value from the sensor to determine whether to extract the acquired data as data for AI learning. Specifically, the learning data extraction unit 42 extracts the signal in the passband of the bandpass filter 43 from the sprung acceleration as a physical quantity. Then, if the magnitude of the amplitude of the sprung acceleration that has passed through the bandpass filter 43 is equal to or greater than a predetermined constant value (threshold), the learning data extraction unit 42 determines to extract the data and use it for learning. For learning by the DNN 24, if the absolute value (magnitude) of the passband component of the sprung acceleration is equal to or greater than a predetermined value (threshold), the data at that time is used. On the other hand, other data is not used for learning by the DNN 24.

In this case, the value of the sprung acceleration (physical quantity) is the detected value of the sprung acceleration sensor 8. However, the value of the sprung acceleration may also be an estimated value estimated from road surface input using the analysis model 20 (vehicle model 21). In other words, the data on the vehicle state quantity before extraction is not limited to data acquired in a driving test, but may also be data calculated by the analysis model 20 based on the road surface input. Furthermore, the result of the data extraction determination by the learning data extraction unit 42 is not limited to whether or not to use the data for learning by the DNN 24, but may also be used to determine whether or not to calculate an optimal command value (control quantity) directly using the optimal control unit 22 for the measured data.

Thus, the third embodiment can achieve substantially the same effects as the first embodiment. In the case of sprung acceleration, the command value (control amount) increases or decreases depending on the frequency of the sprung acceleration. In contrast, in the third embodiment, the AI learning unit 41 performs learning using data in which the amplitude (absolute value) of the sprung acceleration exceeds a threshold when the frequency (rate of change) of the sprung acceleration satisfies a predetermined condition. Therefore, the DNN 24 can be trained using data in a range where the sensitivity of the sprung acceleration is high.

Next, Figures 6 and 7 show a fourth embodiment. The fourth embodiment is characterized in that the AI learning unit has a function of removing data whose absolute values do not exceed a threshold value based on the absolute values of vehicle state quantities that are highly correlated with whether or not the force generation mechanism needs to operate, and the vehicle state quantity learning unit uses the data removed by the AI learning unit to learn the vehicle state quantities based on physical quantities related to the operation of the vehicle. Note that in the fourth embodiment, the same components as in the first embodiment described above are designated by the same reference numerals, and their description will be omitted.

In the fourth embodiment, the ECU 51 is a control device that controls vehicle behavior, including vehicle attitude control. The ECU 51 is mounted on the vehicle body 1. The ECU 51 is configured in the same manner as the ECU 11 of the first embodiment. The ECU 51 is configured using, for example, a microcomputer, and has a memory 51A that can store data. The ECU 51 is equipped with a controller 52.

The input side of the ECU 51 is connected to the CAN 53 and the road surface measurement sensor 10, and the output side is connected to the variable damping force actuator 7 of the variable damper 6. The ECU 51 acquires physical quantities related to the vehicle's operation from the CAN 53 (Controller Area Network). Physical quantities related to the vehicle's operation include, for example, wheel speed and steering angle. The controller 52 estimates the vehicle's state quantities based on the physical quantities related to the vehicle's operation. The controller 52 calculates the force to be generated by the variable damper 6 (force generation mechanism) of the suspension unit 4 based on the vehicle's state quantities and the road surface detection values obtained by the road surface measurement sensor 10, and outputs a command signal to the variable damping force actuator 7 of the suspension unit 4.

The ECU 51 stores vehicle state variables and road surface input data in memory 51A over a period of several seconds, for example, when the vehicle has traveled approximately 10 to 20 meters. This allows the ECU 51 to generate time-series data of road surface inputs (road surface profile) and time-series data of vehicle state variables when the vehicle has traveled a specified distance. The controller 52 controls the variable damper 6 to adjust the damping force to be generated based on the road surface profile and the time-series data of vehicle state variables.

The controller 52 is equipped with a trained DNN 13 (deep neural network). In addition, the controller 52 is equipped with a vehicle state quantity learning unit 54 that estimates vehicle state quantities based on physical quantities related to the vehicle's operation. The vehicle state quantity learning unit 54 is equipped with a DNN and has learned the correlation between physical quantities related to the vehicle's operation and the vehicle's state quantities. As a result, the vehicle state quantity learning unit 54 estimates the vehicle's state quantities from the physical quantities related to the vehicle's operation obtained from the CAN 53.

The controller 52 acquires time-series data of road surface inputs (road surface profile) and time-series data of vehicle state quantities based on the estimated values of vehicle state quantities estimated by the vehicle state quantity learning unit 54 and the road surface detection values by the road surface measurement sensor 10. The DNN 13 of the controller 52 outputs time-series data of optimal command values based on the time-series data of road surface inputs and the time-series data of vehicle state quantities. This allows the controller 52 to output a damping force command value that is most appropriate for the current vehicle and road surface. The damping force command value corresponds to the current value for driving the damping force variable actuator 7.

The AI learning unit 55 trains the DNN 24 using data extracted by the learning data extraction unit 56 according to the fourth embodiment (time series data of road surface input, time series data of vehicle state quantities, and time series data of optimal command values). In addition, the vehicle state quantity learning unit 54 trains the DNN of the vehicle state quantity learning unit 54 using data extracted by the learning data extraction unit 56 according to the fourth embodiment (time series data of road surface input, time series data of vehicle state quantities). Here, the data extraction process performed by the learning data extraction unit 56 will be described.

First, data on road surface inputs, vehicle state variables, and optimal control commands to be used for learning are accumulated. This data may be calculated from road surface input data using the analytical model 20 (vehicle model 21) and direct optimal control unit 22, or it may be data acquired by actually driving the vehicle and using sensors (sprung acceleration sensor 8, vehicle height sensor 9, road surface measurement sensor 10). From this data, the correlation between the sprung velocity included in the vehicle state variables and the control variable of the optimal control command is determined. A constant value is determined as a threshold for the absolute value of the sprung velocity within a range that does not increase the control variable. This constant value is a threshold for the amplitude of the sprung velocity. This threshold is set from the upper limit of the threshold. When setting from the lower limit of the threshold, the threshold is determined so that data from driving straight on a flat road can be excluded.

Next, the wheel speed and steering angle data acquired from CAN 53, along with data (detected values) measured by sprung acceleration sensor 8, vehicle height sensor 9, and road surface measurement sensor 10 during vehicle driving tests, etc., are input to AI learning unit 55. Learning data extraction unit 56 of AI learning unit 55 acquires sprung velocity from, for example, the sprung acceleration from sprung acceleration sensor 8. Using the sprung velocity corresponding to the detected value from the sensor, learning data extraction unit 56 determines whether to extract the acquired data as data for AI learning. Specifically, the determination is made based on the sprung velocity as a physical quantity, whether to extract the data as data for AI learning.

Like the learning data extraction unit 25 in the first embodiment, the learning data extraction unit 56 determines to extract data and use it for learning when the magnitude of the amplitude of the sprung velocity, as the absolute value of the sprung velocity, is equal to or greater than a predetermined constant value (threshold value). The learning of the DNN 24 and the vehicle state quantity learning unit 54 uses data when the absolute value of the sprung velocity is equal to or greater than the constant value (threshold value). On the other hand, the learning of the DNN 24 and the vehicle state quantity learning unit 54 does not use data when the absolute value of the sprung velocity is less than the constant value.

The DNN 24 uses the data extracted by the learning data extraction unit 56 (time series data of road surface input, time series data of vehicle state quantities, and time series data of optimal command values) to learn the correlations between road surface inputs, vehicle state quantities, and control quantities. The vehicle state quantity learning unit 54 also uses the data extracted by the learning data extraction unit 56 (time series data of wheel speed and steering angle, time series data of road surface input, and time series data of vehicle state quantities) to learn the correlations between wheel speed and steering angle, road surface input, and vehicle state quantities.

Thus, the fourth embodiment can achieve substantially the same effects as the first embodiment. In the fourth embodiment, the AI learning unit 55 has a learning data extraction unit 56 that removes data whose absolute value does not exceed a threshold, based on the absolute value of a state quantity (sprung velocity) that is highly correlated with whether or not the variable damper 6 (force generation mechanism) needs to operate. In addition, the vehicle state quantity learning unit 54 uses the data removed by the learning data extraction unit 56 of the AI learning unit 55 to learn vehicle state quantities (sprung acceleration, vehicle body height, relative velocity, sprung velocity, etc.) based on physical quantities related to vehicle operation (wheel speed, steering angle, etc.).

The learning data extraction unit 56 sets an extraction threshold based on physical quantities that can be measured by sensors or the like while driving, and excludes patterns from the learning data that do not require active control of the suspension device 4. As a result, data that has a low correlation with the control quantity of the optimal control command can be removed, allowing the DNN of the vehicle state quantity learning unit 54 to be trained. As a result, learning data for the vehicle state quantity learning unit 54 can be effectively extracted, and the estimation accuracy of the vehicle state quantities based on the learning results of the vehicle state quantity learning unit 54 can be improved.

In the fourth embodiment, physical quantities related to vehicle operation (wheel speed, steering angle, etc.) are acquired from CAN 53. However, the present invention is not limited to this, and for example, physical quantities related to vehicle operation may be detected by a sensor or the like.

Furthermore, the learning data extraction unit 56 is configured similarly to the learning data extraction unit 25 in the first embodiment, but may also be configured similarly to the learning data extraction units 32 and 42 in the second and third embodiments.

In each of the above embodiments, the road surface profile acquisition unit detected the road surface profile using the road surface measurement sensor 10. However, the present invention is not limited to this. The road surface profile acquisition unit may acquire information from a server based on GPS data, for example, or may acquire information from other vehicles via vehicle-to-vehicle communication. The road surface profile acquisition unit may also estimate the road surface profile based on the vertical vibration acceleration detected by the sprung acceleration sensor 8 and the vehicle height detected by the vehicle height sensor 9. In this case, the road surface profile acquisition unit is composed of a calculation section within the ECU 11 in addition to various sensors.

In each of the above embodiments, the suspension control device has a road surface profile acquisition unit in addition to having a vehicle state quantity acquisition unit or a vehicle state quantity learning unit. The present invention is not limited to this, and the suspension control device does not need to have a road surface profile acquisition unit. In this case, the controller of the suspension control device adjusts the force generated by the force generation mechanism based on the results acquired only by the vehicle state quantity acquisition unit or the vehicle state quantity learning unit. The controller has an AI learning unit that learns a command value for the force generation mechanism based on the results acquired by the vehicle state quantity acquisition unit or the vehicle state quantity learning unit. The AI learning unit of the controller learns the command value obtained in advance using an optimization method to minimize a certain evaluation function and the results acquired by the vehicle state quantity acquisition unit or the vehicle state quantity learning unit.

In the above embodiments, the force generation mechanism is a variable damper 6 made of a semi-active damper. The present invention is not limited to this, and an active damper (either an electric actuator or a hydraulic actuator) may be used as the force generation mechanism. In the above embodiments, the force generation mechanism that generates an adjustable force between the vehicle body 1 side and the wheel 2 side is configured as a variable damper 6 made of a damping force adjustable hydraulic shock absorber. The present invention is not limited to this, and the force generation mechanism may be configured, for example, with air suspension, stabilizer (kinesus), electromagnetic suspension, etc. in addition to a hydraulic shock absorber.

In the above embodiments, a vehicle behavior device for use in a four-wheeled vehicle has been described as an example. However, the present invention is not limited to this and can also be applied to, for example, two-wheeled and three-wheeled vehicles, or work vehicles and transport vehicles such as trucks and buses.

The above embodiments are merely examples, and it goes without saying that partial substitution or combination of the configurations shown in different embodiments is possible.

1: Vehicle body, 2: Wheel, 3: Tire, 4: Suspension device, 5: Suspension spring, 6: Variable damper (force generation mechanism), 7: Variable damping force actuator, 8: Sprung acceleration sensor (vehicle state quantity acquisition unit), 9: Vehicle height sensor (vehicle state quantity acquisition unit), 10: Road surface measurement sensor, 11, 51: ECU, 12, 52: Controller, 13, 24: DNN, 20: Analysis model (vehicle state quantity acquisition unit), 21: Vehicle model, 22: Direct optimization control unit, 23, 31, 41, 55: AI learning unit, 25, 32, 42, 56: Learning data extraction unit, 33, 43: Bandpass filter, 54: Vehicle state quantity learning unit

Claims (5)

1. A suspension control device that controls a force generating mechanism that is interposed between a body and a wheel of a vehicle and is capable of adjusting a force between the body and the wheel,
a vehicle state quantity acquisition unit that detects or estimates a state quantity of the vehicle;
an AI learning unit that learns a command value for the force generating mechanism based on the acquisition result of the vehicle state quantity acquisition unit,
The AI learning unit is a suspension control device having a function of removing data whose absolute value does not exceed a threshold value based on the absolute value of a physical quantity that is highly correlated with whether or not the force generation mechanism needs to operate. The suspension control device described in claim 1, wherein the physical quantity is sprung velocity and the absolute value is the magnitude of the amplitude of the sprung velocity. the AI learning unit performs learning using data in which the absolute value of the physical quantity exceeds the threshold when the rate of change of the physical quantity satisfies a predetermined condition, and removes other data;
the physical quantity is the relative velocity between the sprung and unsprung masses,
2. The suspension control device according to claim 1, wherein the AI learning unit extracts data in which the frequency of the relative velocity, which is the rate of change, is within a predetermined range and the magnitude of the amplitude of the relative velocity, which is the absolute value, exceeds a predetermined value, which is the threshold value. the AI learning unit extracts data in which the absolute value of the physical quantity exceeds the threshold when the rate of change of the physical quantity satisfies a predetermined condition, and performs learning, and removes other data,
the physical quantity is a sprung acceleration,
The AI learning unit extracts data in which the frequency of the sprung acceleration, which is the rate of change, is within a predetermined range and the magnitude of the amplitude of the sprung acceleration, which is the absolute value, exceeds a predetermined value, which is the threshold value. A suspension control device as described in claim 1. 1. A suspension control device that controls a force generating mechanism that is interposed between a body and a wheel of a vehicle and is capable of adjusting a force between the body and the wheel,
a vehicle state quantity learning unit that estimates a state quantity of the vehicle;
an AI learning unit that learns a command value for the force generating mechanism based on an estimation result of the vehicle state quantity learning unit,
the AI learning unit has a function of removing data whose absolute value does not exceed a threshold value based on an absolute value of the state quantity that has a high correlation with whether or not the force generation mechanism needs to operate,
The vehicle state quantity learning unit is a suspension control device that learns the state quantity based on physical quantities related to the operation of the vehicle using the data removed by the AI learning unit.