Method for controlling torque of at least one wheel, use for calibrating a torque controller of a mobile platform, control device and / or computer program, and use for traction control of at least one wheel of a mobile platform

Abstract

A method for controlling torque of at least one wheel of a mobile platform, comprising the steps of: providing a current slip value of at least one of the wheels and a current wheel acceleration of at least one of the wheels as input values; providing a trained radial basis function network set up to determine at least one torque change as an output value for controlling the at least one wheel using the input values; and determining a current torque change for controlling the torque using the trained radial basis function network and the provided input values.

Description

[Technical field]

[0001] The present invention relates to a method for controlling the torque of at least one wheel of a mobile platform. [Background technology]

[0002] The vehicle dynamics control system influences the driving state of the vehicle depending on the current slip between the wheels and the road surface, for example in order to achieve the best possible force transmission between the wheels and the road surface or to achieve a stabilization of the driving state. The vehicle dynamics control system then intervenes in the vehicle's driving dynamics, for example by modifying the brake pressure or engine torque. Such corrective interventions can be triggered by the Electronic Stability Program (ESP), the Brake Slip Control (ABS) or the Traction Slip Control (ASR). Vehicle dynamics control systems must be calibrated at great cost for each new vehicle model to its specific requirements regarding comfort and braking behavior, and different road conditions must also be taken into account. Summary of the Invention

[0003] The present invention discloses a method for controlling the torque of at least one wheel of a mobile platform, a method for training a radial basis function network for estimating torque changes, a method for using the calibration method, a control device, a computer program and a method for using the control device, which at least partially solve the above-mentioned problems, according to the features of the independent claims. Preferred embodiments are the subject of the dependent claims and the following description.

[0004] Throughout this specification of the present invention, the order of each method step is described so that the method can be easily understood. However, one of ordinary skill in the art will recognize that many of the method steps can proceed in a different order and still lead to the same or equivalent results. In this sense, the order of each method step can be changed accordingly.

[0005] In one aspect, a method for controlling the torque of at least one wheel of a moving platform is proposed, comprising the following steps: In one step, a current slip value of at least one of the wheels and a current wheel acceleration of at least one of the wheels are provided as input values. In a next step, a trained radial basis function network is provided, which is set up to determine at least one torque change as an output value for controlling the at least one wheel using the input values. In a next step, a current torque change is determined for controlling the torque using the trained radial basis function network and the provided input values.

[0006] A Radial Basis Function network (RBF network) is a special type of layered artificial neural network that uses nonlinear radial basis functions as activation functions and has an input layer, an output layer and a layer of hidden neurons with RBF activation functions. The nonlinear RBF activation layer, which depends only on the distance from a central vector, can then be radially symmetric around this vector. That is, the radial basis functions depend only on the distance from a given center. The input values ​​for an RBF network can be modeled, for example, as vectors of real numbers. The output of an RBF network is typically a linear and / or sigmoidal combination of the input radial basis functions and the neuron parameters (weights w). The hidden layer may have a number of neurons, each with a central vector for each neuron. The linear output layer may have output neurons, which are functionally connected to each neuron through weights w. Typically, all inputs are connected to each hidden neuron. The norms of the distances of the input values ​​from the respective central vectors may be defined as Euclidean distances, and / or squared distances, and / or L0 norms, and / or L1 norms, and / or L2 norms, and may be defined as any function, for example linear, sigmoidal, quadratic, etc. Radial basis functions may be defined in particular as Gaussian distribution functions, and accordingly depend only on the distances of the neurons to the given centers. RBF networks may be trained by error feedback or gradient methods.

[0007] The wheel acceleration of a wheel corresponds to the angular acceleration of the wheel, or corresponds to the acceleration of a point on the circumference of the wheel.

[0008] Here, the torque change is a change in torque acting on at least one wheel to brake quasi-brake torque and / or a change in torque acting on at least one wheel to accelerate quasi-engine torque.

[0009] The indicator of control is interpreted broadly here and includes not only control in the narrow sense but also control of the torque of at least one wheel of the mobile platform.

[0010] The proposed method can simplify and automate the calibration of, for example, brake controllers, and by utilizing a trained radial basis function network, unlike deep neural networks, it provides a verifiable determination of the current torque change.

[0011] Notably, the method can be adapted to other complex controllers, thereby shortening the "time to market".

[0012] In this method, the advantage is realized that by determining the current change in torque by the trained radial basis function network, an absolutely small value for the change in torque has to be determined or estimated by the radial basis function network, thereby enabling a more accurate determination of the torque change. In addition, the determined current torque change allows a more stable control of the torque, in particular because unstable decisions can be more easily identified and filtered.

[0013] The proposed method has the advantage that it can be applied for slip control, for example in anti-lock systems. Alternatively or additionally, the proposed method can be applied for slip control in traction control systems (TCS).

[0014] Determining the current torque change by means of a trained radial basis function network has the advantage that it provides a robust and verifiable control of the torque of at least one wheel of a mobile platform, since the trained radial basis function network is based on a simple structure and can be analyzed with respect to functional relevance. In particular, the use of radial basis function networks for the determination of the current torque change allows the partitioning of the state space into regions according to the practical application of wheel torque control on different roadbeds, thus allowing verifiable state-action-assignments and thus a transparent representation of nonlinear relationships. Moreover, the use of a RBF architecture for the method is particularly flexible, allowing adaptation and extension of various parameters and other input quantities to be applied to almost any control problem.

[0015] Additionally, radial basis function networks have favorable statistical properties for controlling torque.

[0016] In one embodiment, the above described method proposes that the determined torque change is added to the current torque of the wheel to determine the current torque.

[0017] In one embodiment, the above described method proposes to determine the brake pressure and / or the brake pressure change, whereby the determined torque and / or the determined torque change, respectively, are calculated using a functional relationship between torque and brake pressure.

[0018] In one embodiment, it is proposed that the input values ​​provided for the trained radial basis function network additionally comprise a first sequence of previous values ​​of the wheel normal force and a second sequence of previous torque values ​​to determine the current torque change. The use of yet another input value format allows for an improved determination of the current torque change.

[0019] In one aspect, it is proposed that the input values ​​provided for the trained radial basis function network consist of a current slip value of the wheel, a current wheel acceleration of the wheel, a first sequence of previous values ​​of the normal force of the wheel, and a second sequence of previous torque values, and the current torque change is determined by the radial basis function network trained on these four input value formats.

[0020] These four input formats have the advantage that the overall situation of the mobile platform can be characterized particularly well in order to adapt the torque control, especially under different roadbed and environmental conditions and the resulting different friction values ​​and normal forces, by corresponding training of the radial basis function network with these four input formats. These four input formats allow different driving scenarios to be reflected in the state space of the radial basis function network, thus enabling reasonable state-action-assignments and correspondingly accurate determination of the current torque change by the proposed method. In this way, nonlinear relationships can be represented transparently.

[0021] In this case, the first sequence and the second sequence of previous values ​​can be targeted to corresponding values ​​in the form of input values ​​through the most recent or past time steps, and in particular, can be filtered through the most recent n time steps to make the torque control more robust and stable.

[0022] In one embodiment it is proposed that the input values ​​comprise the current friction coefficient of the wheel, and / or the current torque of the wheel, and / or the continuous average value of the torque of the wheel, and / or the current change in torque over time, and / or the gradient of the torque of the wheel, and / or the average value of the torque of the wheel, and / or the current torque of at least one other wheel of the moving platform, and / or the normal force of at least one other wheel of the platform, and / or the difference between the current slip and the target slip, and / or a dynamics value of the moving platform, and / or the current change in wheel acceleration over time, and / or the normal force of at least one other wheel of the moving platform, and / or the current slip value and at least one current wheel acceleration of at least one other wheel of the platform. The dynamics of the mobile platform may then be characterized by the yaw rate of the mobile platform and / or the acceleration of the mobile platform and / or the steering angle of the mobile platform.

[0023] In one aspect, it is proposed that the provided trained radial basis function network is set up to additionally determine, using input values, a change in engine torque as an output value for controlling the torque of at least one wheel, and the trained radial basis function network additionally determines a change in engine torque as an output value for controlling the torque of at least one wheel. By controlling both the driving torque and the braking torque using the engine torque and / or the brake torque, the dynamic behavior and / or traction of the mobile platform can be influenced within limits, and not only the best braking behavior but also comfort requirements related to the steering behavior and / or the braking behavior can be taken into account for the control. Such control of the torque changes, both for the braking torque and for the driving torque, applies in particular to the control or regulation of the torque with only the four input value formats as input quantities described above.

[0024] A method for training a radial basis function network to estimate torque changes of at least one wheel of a mobile platform is proposed, comprising the steps of: In one step, slip values ​​of the wheels are provided. In a next step, wheel accelerations of the wheels corresponding to the slip values ​​are provided. In a next step, input values ​​for a radial basis function network are formed by the slip values ​​and the respectively corresponding wheel accelerations. In a next step, target torque changes corresponding to the input values ​​are provided and assigned. In a next step, the radial basis function network is trained with a number of different input values ​​and respectively assigned target torque changes to estimate the torque change using the input values.

[0025] This method of training the radial basis function network can be performed online using a suitably set up mobile platform and / or offline using a suitably set up training device, for example using a brake test stand. In the case of online training, the torque control can also be adapted to changed conditions such as load and tire changes.

[0026] In the training method, the target torque change can be determined heuristically by the control by evaluating stored measured input values ​​and a controlled torque change, for example based on measurements on a brake test stand. According to the target torque change, the parameters of the radial basis function network can be adapted by gradient descent if there is a deviation from the best evaluated torque change.

[0027] For example, in a training method, the average brake slip s l is determined over the interval with time steps t to t+n, allowing a heuristic evaluation of the torque change estimated by the radial basis function network. In the next step, the average brake slip s l Goal slip sl soll is compared to. In the next step, the improved torque change a as the target torque change * At time step tm, Formula 1: a * =(a slsoll -a sl ) * It can be calculated by alpha, where a is the actual proposed torque change and alpha is a heuristic coefficient. In the next step, we update the parameters of the radial basis function network, e.g., the weights w at time step tm, using gradient descent and the error (a * -a)2 It can be implemented using: By using the average brake slip, the method allows the intrinsic system delays to be taken into account during training. Alternatively or additionally, the brake slip can be estimated by a Savitzky-Golay filter. Alternatively or additionally, such intrinsic system delays can be explicitly taken into account by a time shift interval. For the evaluation of the estimated torque change, the first or second time derivative of the target slip can also be used to determine the target torque change using the adapted equation 1. In this way, the gradient of the slip change can be derived from the interval with time steps t to t+n. Alternatively or additionally, the torque change proposed by the radial basis function network can be evaluated according to the above description with other input value formats of the radial basis function network and their changes over time.

[0028] In one embodiment, it is proposed that the method for training the radial basis function network is performed online by a correspondingly set-up control device of the mobile platform. The training of the radial basis function network performed online allows the torque controller of the mobile platform to be trained more quickly and / or allows the training to also take into account changes in the environment and / or changes in the wheels, for example after a wheel change.

[0029] In one embodiment, it is proposed that the method of training the radial basis function network is performed online, during the braking process. In other words, during the braking process, the parameters of the radial basis function network are adapted during the currently ongoing braking process to train the radial basis function network.

[0030] In one embodiment, it is proposed that the method for training the radial basis function network is performed offline, using a stored series of measurements of input and output values, in particular the torque change at at least one wheel, where other dynamic characteristics of the mobile platform can also be stored and used for offline training. Offline training has the advantage that a higher computational power is available for optimizing the radial basis function network.

[0031] In one aspect, it is proposed that the input values ​​provided for training the radial basis function network additionally comprise a first sequence of previous values ​​of the wheel normal force and a second sequence of previous torque values.

[0032] In one aspect, it is proposed that the input values ​​provided for training the radial basis function network consist of a current slip value of the wheel, a current wheel acceleration of the wheel, a first sequence of previous values ​​of the normal force of the wheel, and a second sequence of previous torque values.

[0033] In one aspect, it is proposed that each input value is additionally assigned a corresponding target change in engine torque, and a radial basis function network is trained with a large number of different input values, respectively assigned target torque changes, and assigned target changes in engine torque.

[0034] In one aspect, it is proposed that at least some of the different types of input values ​​described above are provided to a radial basis function network, and the radial basis function network is trained with these multiple input values ​​and the respectively assigned target torque changes and / or target changes in engine torque to estimate the torque change and / or engine torque change.

[0035] In one embodiment, it is proposed that the radial basis function network is configured based on expert knowledge and / or physical limits before training to achieve improved training. Besides the option of configuring the radial basis function network with random values ​​for training and / or uniformly defining neuron positions, the radial basis function network can be defined by expert knowledge to require less training and / or to produce better estimates. Expert knowledge can be incorporated, in particular with regard to the distribution of the number and positions of the neurons, by selecting the neuron positions according to the desired accuracy of the estimation depending on the physical constraints of the respective input values ​​and the desired granularity, in particular by selecting the neuron positions non-uniformly within the constraints, in particular by considering whether the damping torque and / or the driving torque should be controlled.

[0036] In one embodiment, it is proposed that the radial basis function network is configured based on expert knowledge by placing the centers of the radial basis function network in the state space and / or by weighting the centers of the radial basis function network.

[0037] A method is proposed in which a control signal is provided for controlling an at least semi-automated vehicle based on the torque and / or engine torque determined by one of the methods described above and / or a warning signal is provided for warning a vehicle occupant based on the determined torque and / or engine torque. Accordingly, for example, a control device of a mobile platform brakes the mobile platform by controlling the braking torque or braking force on at least one wheel, by transmitting a control signal to a brake actuator that can act on at least one wheel.

[0038] The concept of "based on" should be understood in a broad sense with respect to the requirement that a control signal is provided based on a determined torque and / or engine torque. This should be understood such that the determined torque and / or engine torque is invoked for each determination or calculation of the control signal, which does not exclude that other input quantities are also invoked for such a determination of the control signal. The same applies mutatis mutandis to the provision of a warning signal.

[0039] One usage of one of the methods described above for calibrating a torque controller of a mobile platform is proposed. At this time, the torque controller includes a controller for braking torque and / or engine torque.

[0040] A control device and / or a computer program set up to implement one of the methods described above is proposed.

[0041] With such a device, the corresponding method can be easily incorporated into various different systems.

[0042] A usage of the control device described above for traction control of at least one wheel of a mobile platform is proposed.

[0043] In another aspect, a computer program is described that includes commands instructing to implement one of the methods described above when executed by a computer. Such a computer program enables the application of the above method in various different systems.

[0044] A machine-readable storage medium storing the computer program described above is described.

[0045] The mobile platform may be an at least partially automated system that is mobile and / or may be a driver assist system. One example may be an at least partially automated vehicle or a vehicle with a driver assist system. That is, in this context, an at least partially automated system includes a mobile platform with respect to at least partially automated functionality, but a mobile platform includes vehicles and other mobile machines, including driver assist systems. Other examples of mobile platforms may be a driver assist system with multiple sensors, a mobile multi-sensor robot, such as a robotic vacuum cleaner or lawn mower, a multi-sensor surveillance system, a ship, an aircraft, a manufacturing machine, a personal assistant, an entrance control system. Each of these systems may be a fully or partially automated system.

[0046] An embodiment of the present invention is shown in relation to FIG. 1 and is described in detail below. [Brief description of the drawings]

[0047] [Figure 1] This is the structure of a radial basis function network. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0048] 1 shows a schematic outline of the structure of a radial basis function network 100, having an input layer 110a-d, a layer of hidden neurons 120a-f, RBF activation or transformation functions 130a-f, and an output layer 150, where different neurons 120a-f contribute to the output layer 150 according to weights w1 to w6 140a-f. Such a radial basis function network 100 can be used for torque control, since the radial basis function network 100 uses the input values ​​to estimate the torque change 160.

[0049] To control the torque of at least one wheel of the mobile platform, a current slip value of the wheel, a current wheel acceleration of the wheel, a first sequence of previous values ​​of the normal force of the wheel, and a second sequence of previous torque values ​​can be applied as input values ​​to an input layer 110a-d of a trained radial basis function network 100. The radial basis function network 100 trained with the input values ​​110a-d uses the applied inputs to estimate a current torque change 160 for controlling the torque. To do this, in one step, all input values ​​110a-d are normalized to the range from zero to one. The next step is to calculate the distance between the input signals and the neurons, where each neuron represents a specific point in the state space. To determine the distance, a Euclidean distance metric is applied. Alternatively or additionally, other distance metrics can be applied, for example the l1 distance metric. The output of each neuron is then the distance between its current state, i.e., the current input signal at time t, and the neuron's unique center. In the next step, this distance is transferred to the output layer 160 by a radial basis function f. This radial basis function f can be any function. For example, it can be a Gaussian distribution function that returns high values ​​when the input is close to zero and low values ​​when the input is high. The input is in this case the distance to the center. That is, the output of each neuron is transformed by the radial basis function and, according to a linear regression, a unique weight w n Multiplied with 140a-f and added together to estimate torque change 160. [Explanation of symbols]

[0050] 100 Radial Basis Function Networks 110a-d Input values 120a-f Hidden neurons 130a-f RBF activation functions or transformation functions 140a-f weight w n 150 output layers 160 Output Value

Claims

1. A method for controlling the torque of at least one wheel of a mobile platform, comprising the following steps executed by a control device and / or a computer program: a current slip value of at least one of the wheels and a current wheel acceleration value of at least one of the wheels are provided as input values ​​(110a-d); a trained radial basis function network (100) is provided, set up to determine, using said input values ​​(110a-d), at least one torque change as an output value (160) for control of at least one wheel; determining a current torque change for controlling a torque using the trained radial basis function network (100) and the provided input values; wherein the input values ​​(110a-d) provided to the trained radial basis function network (100) additionally comprise a first sequence of previous values ​​of wheel normal force and a second sequence of previous torque values ​​to determine a current torque change.

2. A method for controlling the torque of at least one wheel of a mobile platform, comprising the following steps executed by a control device and / or a computer program: a current slip value of at least one of the wheels and a current wheel acceleration value of at least one of the wheels are provided as input values ​​(110a-d); a trained radial basis function network (100) is provided, set up to determine, using said input values ​​(110a-d), at least one torque change as an output value (160) for control of at least one wheel; determining a current torque change for controlling a torque using the trained radial basis function network (100) and the provided input values; The input values ​​(110a-d) provided to the trained radial basis function network (100) consist of a current slip value of the wheel, a current wheel acceleration of the wheel, a first sequence of previous values ​​of the wheel normal force, and a second sequence of previous torque values, and a current torque change is determined by the radial basis function network (100) trained on these four forms of input values.

3. 3. The method according to claim 1 or 2, wherein said input values ​​(110a-d) comprise a current friction coefficient of the wheel, and / or a current torque of the wheel, and / or a continuous average value of the torque of the wheel, and / or a current change in torque over time, and / or a gradient of the torque of the wheel, and / or an average value of the torque of the wheel, and / or a current torque of at least one other wheel of the moving platform, and / or a normal force of at least one other wheel of the platform, and / or a difference between a current slip and a target slip, and / or a dynamics value of the moving platform, and / or a current change in wheel acceleration over time, and / or a normal force of at least one other wheel of the moving platform, and / or a current slip value and at least one current wheel acceleration of at least one other wheel of the platform.

4. 4. The method according to claim 1, wherein the trained radial basis function network (100) provided is set up to additionally determine, using the input values ​​(110a-d), a change in engine torque as an output value (160) for control of the torque of at least one wheel, and the trained radial basis function network (100) additionally determines a change in engine torque as an output value (160) for control of the torque of at least one wheel.

5. The method according to any one of claims 1 to 4, comprising: a wheel slip value is provided; and a wheel acceleration corresponding to the slip value is provided; and said slip values ​​and their corresponding wheel accelerations form input values ​​(110a-d) for said radial basis function network (100); a target torque change corresponding to the input value is provided and assigned; the radial basis function network (100) is trained with a number of different input values ​​(110a-d) and respective assigned target torque changes to estimate torque changes using the input values ​​(110a-d).

6. 6. The method of claim 5, wherein the input values ​​(110a-d) provided for training the radial basis function network (100) additionally comprise a first sequence of previous values ​​of wheel normal force and a second sequence of previous torque values.

7. A method for controlling the torque of at least one wheel of a mobile platform, comprising the following steps executed by a control device and / or a computer program: a current slip value of at least one of the wheels and a current wheel acceleration value of at least one of the wheels are provided as input values ​​(110a-d); a trained radial basis function network (100) is provided, set up to determine, using said input values ​​(110a-d), at least one torque change as an output value (160) for control of at least one wheel; determining a current torque change for controlling a torque using the trained radial basis function network (100) and the provided input values; wherein the input values ​​(110a-d) provided for training the radial basis function network (100) consist of a current slip value of the wheel, a current wheel acceleration of the wheel, a first sequence of previous values ​​of the normal force of the wheel, and a second sequence of previous torque values ​​of the wheel.

8. 8. The method of claim 7, further comprising: additionally assigning to each of the input values ​​a corresponding target change in engine torque, and training the radial basis function network with a number of different input values, the respective assigned target torque changes, and the assigned target changes in engine torque.

9. 9. The method according to claim 6, wherein at least a portion of the input values ​​of the format according to claim 3 are provided to the radial basis function network (100), and the radial basis function network (100) is trained with the multiple input values ​​(110a-d) and the respectively assigned target torque change and / or target change in engine torque to estimate the torque change and / or engine torque change.

10. 10. The method according to claim 6, wherein the radial basis function network (100) is configured prior to training based on physical limits to achieve improved training.

11. 11. The method of claim 10, wherein the radial basis function network (100) is configured based on the physical limits, by placing centers of the radial basis function network in a state space, and / or by weighting multiple centers of the radial basis function network (100).

12. Use of the method according to any one of claims 1 to 11 for calibrating a torque controller of a moving platform.

13. A control device set up to carry out a method according to any one of claims 1 to 11.

14. A computer program set up to carry out the method according to any one of claims 1 to 11.

15. Use of the control device as claimed in claim 13 for traction control of at least one wheel of a mobile platform.

Images