Method and device for designing a feedfor control system for a controlled variable of a machine

Abstract

The invention relates to a computer-implemented method for designing a feedforward control system for a controlled variable of a machine using a Local Model Network (LMN), the method comprising the steps: - Providing (S1) input data, in particular current and historical, to the controlled variable of the machine, as well as data of the controlled variable corresponding to the input data, in particular current and historical; - Training (S2) the Local Model Network based on the input data and the data of the controlled variable and converting (S3) the Local Model Network into a state space representation; - Applying (S4) a feedback linearization to the trained Local Model Network in the state-space representation to provide the feedforward control for the controlled variable; and - Providing (S5) the feedforward control for a controlled variable to control the machine's controlled variable.

Description

[0001] The invention relates to a computer-implemented method for designing a feedforward control for a controlled variable of a machine. The invention further relates to a device for designing a feedforward control for a controlled variable of a machine. State of the art

[0002] In recent years, interest in assistance and autonomous functions for work machines has increased significantly. This trend is being driven by the ongoing development of electro-hydraulic control systems, which are now standard in new work machines such as excavators.

[0003] These systems make it possible to reduce workload and increase productivity by taking responsibility for demanding and / or repetitive tasks that require high accuracy. A key factor in implementing these functions is precise trajectory tracking, allowing experienced operators to track trajectories quickly and precisely. These demanding requirements place similarly high demands on autonomous functions, pushing simple, empirically based control structures to their limits.

[0004] In the wake of these developments, learning-based artificial intelligence (AI) methods are gaining increasing importance. They aim to design control systems based on collected system data, thus eliminating the need for time-consuming and potentially inaccurate modeling processes. In known AI-based methods, the system controllers are usually trained directly on the basis of collected data by modeling an inverse system behavior. In the case of work machines, for example, the actual input joystick behavior is inversely learned based on a system output behavior, in particular the cylinder stroke speeds of the variable-displacement cylinders. AI methods are also particularly advantageous for applications in which systems are manufactured in small quantities and with great variety, such as mobile work machines in mining and construction.

[0005] Although artificial intelligence solutions are already known, there is still a need for development, especially to create an alternative to approaches that represent inverse system behavior.

[0006] It is an object of the invention to provide an improved method and / or an improved device.

[0007] The object is achieved by a method according to the features of patent claim 1. The object is achieved by a device according to the features of patent claim 10. Disclosure of the invention

[0008] According to a first aspect, a computer-implemented method for designing a feedforward control for a controlled variable of a machine using a Local Model Network (LMN) is proposed, the method comprising the steps: - Providing input data, in particular current and historical, for the controlled variable of the machine as well as data, in particular current and historical, of the controlled variable corresponding to the input data; - Training the Local Model Network based on the input data and the controlled variable data and converting the Local Model Network into a state space representation; - Applying a feedback linearization to the trained Local Model Network in the state-space representation to provide the feedforward control for the controlled variable; and - Providing feedforward control for a controlled variable to control the controlled variable of the machine.

[0009] It is understood that the steps according to the invention, as well as other optional steps, do not necessarily have to be performed in the order shown, but can also be performed in a different order. Furthermore, additional intermediate steps can be provided. The individual steps can also comprise one or more substeps without thereby departing from the scope of the method according to the invention.

[0010] According to a second aspect, a device is proposed for designing a feedforward control for a controlled variable of a machine using a Local Model Network (LMN), wherein the device comprises an evaluation and computing device which is designed to carry out the following steps: - Providing input data, in particular current and historical, for the controlled variable of the machine as well as data, in particular current and historical, of the controlled variable corresponding to the input data; - Training the Local Model Network based on the input data and the controlled variable data and converting the Local Model Network into a state space representation; - Applying a feedback linearization to the trained Local Model Network in the state-space representation to provide the feedforward control for the controlled variable; and - Providing feedforward control for a controlled variable to control the controlled variable of the machine.

[0011] The statements made for the method apply accordingly to the device. It is understood that linguistic modifications of procedurally formulated features can be reformulated for the device according to common linguistic practice, without such formulations having to be explicitly listed here.

[0012] The present design method pursues a different design strategy from the prior art, in which collected machine data is first used to design a system model, namely the LMN. The feedforward control is then determined based on the LMN. The method described herein, particularly a computer-implemented one, is based on mathematical methods for designing a feedforward control algorithm using local model networks (LMNs). Using the presented method, such feedforward control algorithms can be developed for a wide variety of applications.

[0013] Local Model Networks (LMNs) provide a data-driven modeling framework for nonlinear systems or machine control variables with limited inputs (e.g., fewer than 20-30 features). LMNs exhibit favorable extrapolation behavior and enable efficient online adaptation of local model parameters. The simple system structure of LMNs also enables better interpretability, which is an advantage in validating the trained models or the provided feedforward control. In this case, the LMN is used to train the system behavior or machine behavior as a function of the controlled variable based on the collected or provided (input and output) data.The feedforward controller for tracking the controlled variable (also called manipulated variable) is preferably generated automatically using a feedback linearization based on the data-based LMN.

[0014] A Local Model Network (LMN) is a type of adaptive or data-driven model used in control and prediction engineering. LMNs belong to the category of nonlinear modeling methods and work with a combination of several local models, each of which applies to a specific region of an input space. The overall output of the network is preferably formed by the weighted sum of the outputs of these local models, with the weights preferably depending on the input conditions.

[0015] In other words, Local Model Networks (LMNs) are a machine learning approach in which a model is split into several local models, each responsible for specific subsets of the data. These local models are then combined into a global model to make a prediction for new data.

[0016] Feedback linearization is a method for controlling nonlinear systems that aims to "linearize" the nonlinear behavior of a system through feedback. This means that, by applying a special feedback control, the originally nonlinear system is transformed into an equivalent, linear system. This simplifies the analysis and design of control systems because linear systems are easier to treat mathematically. Furthermore, it requires less computational power.

[0017] In a further aspect, it is proposed that a nonlinear behavior of the controlled variable is represented by the trained local model network as a combination of a plurality of local linear models, in particular each with model weights, each of which has a nonlinear weighting and / or a nonlinear validation.

[0018] In this case, the controlled variable-dependent system or machine behavior is identified using the LMN. The basic idea behind LMNs is to generate a model output ŷ as a weighted sum of local (individual) models ŷ i The weights are so-called validations or validity functions Φ i , and the local models ŷ i are usually of a linear type.

[0019] The model output ŷ is described by equation (1): y^=∑i=1Mγ^i(ulin)Φi(uval)=∑i=1M(wi0+wi1ulin1+wi2ulin2+⋯+wi,nulinn)Φi(uval_).

[0020] With this model construction of the LMN, it is possible to describe a nonlinear system, process, machine, and / or controlled variable behavior, while only linear equations are used at a local level.

[0021] Another aspect of LMN models is that both the input data for calculating the local models ŷ i as well as the input data for calculating the validations or validity functions Φ i can be treated individually. This allows a distinction to be made between the input data and val for the validity functions Φ i that influence the LMN model in a nonlinear way and the inputs u lin for the local models ŷ i which have a linear effect on the model output ŷ.

[0022] This is preferred when it comes to controlling dynamic processes and / or controlled variables of a machine, where, for example, delayed versions of physical inputs have a high correlation and thus only in the inputs and lin for the local models ŷ i are to be included, whereas in the input data u val for the validity functions Φ i should be disregarded.

[0023] As already shown in equation (1), the i-th local linear model (LLM) is described by equation (2): y^i(x_)=wi0+wi1ulin1+wi2ulin2+⋯+wi,nulinnx=ulin˜_Twi_ where w i = [w i0 w i1 ...w i,nx ] T the local model parameters, ie an offset and the gradients in each input dimension and ulin_˜=[1 ulin1…ulinnx]T are the input vector for the respective local model. The regressor “1” in ulin_˜ extends the input vector for estimating the offsets. This linear formulation has the advantage that standard (weighted) least squares methods can be applied for the estimation. Furthermore, this is particularly preferred in the context of the underlying application, as it enables the use of recursive algorithms for online parameter updates.

[0024] In a further aspect, it is proposed that a weighted mean is formed from the combination of the plurality of local linear models, each of which has a / the non-linear weighting and / or a / the non-linear validation, in order to represent the non-linear behavior of the controlled variable.

[0025] In a further aspect, it is proposed that the respective non-linear weighting and / or the non-linear validation is determined on the basis of a probability distribution function, in particular on the basis of a normalized Gaussian participant function.

[0026] The validity functions or validations preferably determine in which input region of the input space a local model is valid. Determining the position and shape of these validity functions, or so-called partitioning, is a complex nonlinear optimization problem and is preferably solved heuristically.

[0027] The validity functions or validations Φ i can be calculated, for example, as normalized Gaussian membership functions, as in equation (3): μi(uval_)=exp(−12((uval__1−ci1)2σi12+⋯+(uval__nz−ci,nz)2σi,nz2)) where σ ij the standard deviations and c ijare the centers of the individual local models. The membership functions are then preferably normalized to achieve a division of one. Therefore, all validity functions Φ i for each of the input data u val to one, as described by equation (4): Φi(uval)=μi(uval_)∑j=1Mμj(uval_),∑i=1MΦi(uval)=1

[0028] In a further aspect, it is proposed that the transfer of the local model network into a state space representation comprises a transfer of respective model parameters, in particular the model weights and the validations, to the plurality of local linear models into a state space, in particular each as difference equations.

[0029] LMNs can be used in the structure of a nonlinear autoregressive exogenous model (NARX) to represent dynamic systems. For a system, machine, or controlled variable with multiple inputs and one output, this can be written as shown in equation (5): y^(k+1)=fLMN(u(k−m1),...,u(k−m|M|)d(k−q1),...,d(k−q|Q|),y^(k−n1),....y^(k−n|N| );u(k−m˜1),...,u(k−m˜|M˜|),d(k−q˜1),...,d(k−q˜|Q˜|).y^(k−n˜1),...,y^(k−n˜|N˜|)) with f LMN (and lin ; u val ) as an LMN. The number of input data, especially delayed input data, is preferably a hyperparameter of the LMN and is determined during training. The parameters of this LMN can preferably be trained using the collected input and output data, which preferably have corresponding data pairings.

[0030] To apply feedback linearization to an LMN, it is preferable to transform the representation described in equation (5) into a state-space representation. The state-space representation preferably takes the form of a linear parametric variable (LPV) system, as shown in equation (6): xd(k+1)=Ad(Φk)xd(k)+Bd(Φk)u(k) +Dd(Φk)d(k)+fd(Φk)y^(k)=c⊤xd(k), where the disturbance d(k) preferably enters as a second input of the input data and can be considered as an input feature that influences the system behavior but cannot be controlled. The extended state vector has the form of equation (7): xd(k)=[[u(k−M+1) ⋯ u(k−1)]⊤[u(k−Q+1) ⋯ u(k−1)]⊤[y^(k−N+1) ⋯ y^(k)]⊤]

[0031] The system matrix A d (Φ k ), the input matrices B d (Φ k ) and D d (Φ k ) and the offset term f d (Φ k) are defined by equations (8): Ad(Φk)=∑i=1lΦi(k)Ad,iBd(Φk)=∑i=1lΦi(k)Bd,iDd(Φk)=∑i=1lΦi(k)Dd,ifd(Φk)=∑i=1lΦi(k)fd,i where l is the number of local models and where equation (9) defines: fd i=[0p−1×1 c(i)].

[0032] The upper displacement matrix and the submatrix of the identity are defined as equation (10): Ur=[[0r−1×1] 01×r], Ur∈ℝr×rI¯r=[0r−1×1 Ir−1], I¯r∈ℝr−1×r.

[0033] It should be noted that the last row of the system matrix A d,i which contains the linear model parameters described by equations (11): b(i)T=[bM(i),...,b2(i)]d(i)T=[bq(i),...,b2(i)]a(i)T=[bN(i),...,b1(i)] with bM(i) as a factor of the local model i for the input delayed by time steps m, dq(i) as a factor of the local model i for the disturbance delayed by q time steps and aN(i) as a factor for the feedback delayed by n time steps and c (i) as the offset term, and U as the upper diagonal matrix.

[0034] In a further aspect, it is proposed that by applying the feedback linearization to the trained local model network in the state space representation, a control law for the controlled variable is provided.

[0035] The control law resulting from the feedback linearization is given by equation (12): where δ is the relative degree of the system or machine corresponding to the control input, and δ d the relative degree of the system or machine corresponding to the disturbance input is defined. The vector x d,w (k) is the state vector x d(k), where the values ​​ŷ(k) are replaced by the desired output w(k). The above equation can be solved analytically if Φ k does not depend on the control input u, that is, if u is not a validation function. Otherwise, a numerical solution can be found.

[0036] Note that the control law in the above equation is not causal, as it requires, for example, future values ​​of the desired trajectory and the acting disturbance. A causal control law is obtained if these signals are replaced, for example, by estimates at the current time.

[0037] Preferably, the method further comprises a step of controlling the controlled variable of the machine using the provided pilot control, wherein the controlled variable is preferably designed as a movement variable, in particular as a speed and / or acceleration, of an adjusting cylinder of a mobile or stationary work machine.

[0038] In a further aspect, it is proposed to provide a controlled variable controller for a machine, in particular a working machine, comprising at least one processor which is designed to control a controlled variable of the machine on the basis of a feedforward control which is designed according to the method described above.

[0039] The controlled variable controller may also be included in a vehicle with an autonomous driving function and / or a robotic system and / or an industrial machine.

[0040] In a further aspect, it is proposed that the machine is a work machine, in particular an excavator, with at least one adjusting cylinder, wherein the input data comprises data about a joystick movement for controlling the at least one adjusting cylinder and preferably further disturbance variables, for example a load and / or a pressure on the at least one adjusting cylinder, wherein the controlled variable comprises a speed of the at least one adjusting cylinder, wherein the input data and the data of the controlled variable can each be detected by sensors of the work machine, and wherein the provided pilot control is designed to control, in particular to regulate, the speed of the at least one adjusting cylinder based on the joystick movement in pilot control mode. The work machine can be a stationary machine, e.g., a stationary industrial robot, etc., or a mobile work machine, e.g., an excavator, wheel loader, telehandler, etc., preferably a hydraulic cylinder of the working machine.

[0041] In the application example of a working machine with a tracking / control of the cylinder speed for providing an assistance function of the working machine, an acting system pressure can preferably be considered as a particularly measurable, but not controllable, disturbance variable.

[0042] A particularly relevant application of the present method or the resulting feedforward control as a software product is its use on, in particular, mobile, work machines, such as excavators, dump trucks, shovel loaders, etc. The feedforward control is used in particular for the (speed) control and / or regulation of hydraulic cylinders of such work machines. The feedforward control provided here can be distributed both as a standalone control algorithm or control software ("software-as-a-product") for such hydraulic cylinders and / or work machines. Furthermore, the feedforward control can also be distributed together with the respective hydraulic cylinder and / or the respective work machine.

[0043] In a further aspect, a computer program with program code is claimed for carrying out at least parts of the present method in one of its aspects when the computer program is executed on a computer. In other words, a computer program (product) comprising instructions that, when executed by a computer, cause the computer to carry out the method(s) of the method in one of its aspects.

[0044] In a further aspect, a computer-readable data carrier with program code of a computer program is proposed for executing at least parts of the present method in one of its aspects when the computer program is executed on a computer. In other words, the invention relates to a computer-readable (storage) medium comprising instructions that, when executed by a computer, cause the computer to execute the method(s) of the method in one of its aspects.

[0045] The described designs and further training courses can be combined as desired.

[0046] Further possible embodiments, further developments and implementations of the invention also include combinations of features of the invention described previously or below with regard to the embodiments that are not explicitly mentioned. Short description of the drawings

[0047] The accompanying drawings are intended to provide a further understanding of embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain principles and concepts of the invention.

[0048] Other embodiments and many of the aforementioned advantages will become apparent upon review of the drawings. The elements illustrated in the drawings are not necessarily drawn to scale.

[0049] They show: Fig. 1 is a schematic flow diagram of an embodiment of the present method; Fig. 2 is a schematic block diagram of an embodiment of the device; and Fig. 3 a schematic representation of control results generated with the method.

[0050] In the figures of the drawings, the same reference symbols designate the same or functionally equivalent elements, parts or components, unless otherwise stated.

[0051] Fig. 1 shows a schematic flow diagram of a method for designing a feedforward control for a controlled variable of a machine using a Local Model Network (LMN).

[0052] In any embodiment, the method can be carried out by a device 100, which for this purpose can comprise several components not shown in detail, for example, one or more provision devices and / or at least one evaluation and computing device. It is understood that the provision device can be designed jointly with the evaluation and computing device or can be different from it. Furthermore, the device 100, which can be part of a system, can comprise a storage device and / or an output device and / or a display device and / or an input device.

[0053] The computer-implemented method comprises at least the following steps: In a step S1, input data, in particular current and historical, are provided for the controlled variable of the machine as well as data, in particular current and historical data corresponding to the input data, of the controlled variable

[0054] In step S2, the local model network is trained based on the input data and the controlled variable data.

[0055] In step S3, the local model network is converted into a state space representation.

[0056] In a step S4, a feedback linearization is applied to the trained local model network in the state space representation to provide the feedforward control for the controlled variable.

[0057] In a step S5, the feedforward control for a controlled variable is provided to control the controlled variable of the machine.

[0058] Fig. Figure 2 shows a schematic block diagram of an embodiment of the device 100. The proposed method for designing feedforward controllers can be applied to all applications where LMNs can provide good model accuracy. For example, the design method can be applied to the problem of speed tracking of hydraulic cylinders in a work machine 200, such as a mobile excavator. Fig. 2 shows an exemplary control approach for this application. As in Fig. As shown in Figure 2, a local model network LMN 202 is first trained using collected machine data, i.e., input data 204 and output data 206, which are generated, for example, using random movements of the work machine 200. The input data 204 can, for example, include data about a joystick position. The output data 206 can include data about a speed of a hydraulic cylinder. The LMN 202 preferably describes the (speed) behavior (as the control variable) of respective hydraulic cylinders of the work machine 200. Subsequently, with the aid of the feedback linearization 208, a feedforward control 210, here for a speed controller 211, is designed for the respective hydraulic cylinders.In particular, together with a pose controller 212, which can be provided in an outer control loop, the speed controller 211 can replace a human operator 213 and ensure the tracking of a specific trajectory of a boom 214 of the work machine 200. The speed controller 211 can also receive sensor information 216 as feedback, which may be processed by a filter 218. Furthermore, a trajectory to be generated by the speed controller 211 can be specified by a trajectory generator 220. Furthermore, a difference generator 222 can be provided for speed control.

[0059] The design strategy was validated using vehicle measurements. Good tracking accuracy of all cylinders was achieved, as shown by Fig. 3. The designed / designed and measured speeds of an arm cylinder, a boom cylinder, and a bucket cylinder are shown during an evaluation cycle. The designed speeds were generated using the speed controller 211 derived from the LMN. Furthermore, pressure information was included as a disturbance in the input data.

Claims

[1] Computer-implemented method for designing a feedforward control for a controlled variable of a machine using a Local Model Network (LMN), the method comprising the steps: - Providing (S1) of, in particular current and historical, input data for the controlled variable of the machine as well as of, in particular current and historical data of the controlled variable corresponding to the input data; - Training (S2) the local model network based on the input data and the controlled variable data and converting (S3) the local model network into a state space representation; - Applying (S4) a feedback linearization to the trained Local Model Network in the state space representation to provide the feedforward control for the controlled variable; and - Providing (S5) the feedforward control for a controlled variable to control the controlled variable of the machine. [2] Method according to claim 1, wherein a non-linear behavior of the controlled variable is represented by the trained local model network as a combination of a plurality of local linear models, in particular each with model weights, each having a non-linear weighting and / or a non-linear validation. [3] Method according to claim 2, wherein a weighted mean is formed from the combination of the plurality of local linear models, each having a non-linear weighting and / or a non-linear validation, in order to represent the non-linear behavior of the controlled variable. [4] Method according to claim 2 or 3, wherein the respective non-linear weighting and / or the non-linear validation is determined on the basis of a probability distribution function, in particular on the basis of a normalized Gaussian participant function. [5] Method according to one of claims 2 to 4, wherein the conversion of the local model network into a state space representation comprises a conversion of respective model parameters, in particular the model weights and the validations, to the plurality of local linear models into a state space, in particular each as difference equations. [6] Method according to one of the preceding claims, wherein applying the feedback linearization to the trained local model network in the state space representation provides a control law for the controlled variable. [7] Control variable controller for a machine, in particular a working machine, with at least one processor which is designed to control a control variable of the machine on the basis of a feedforward control which is designed according to the method according to one of the preceding claims 1 to 6. [8] A computer program comprising program code for carrying out the method according to any one of claims 1 to 6 when the computer program is executed on a computer, or for providing a feedforward control provided according to the method according to any one of claims 1 to 6 in a machine-readable form when the computer program is executed on a computer. [9] A computer-readable data carrier comprising program code of a computer program for carrying out the method according to any one of claims 1 to 6 when the computer program is executed on a computer, or for providing a feedforward control provided according to the method according to any one of claims 1 to 6 in a machine-readable form when the computer program is executed on a computer. [10] Device (100) for designing a feedforward control for a controlled variable of a machine using a Local Model Network (LMN), wherein the device (100) has an evaluation and computing device which is designed to carry out the following steps: - Providing (S1) of, in particular current and historical, input data for the controlled variable of the machine as well as of, in particular current and historical data of the controlled variable corresponding to the input data; - Training (S2) the local model network based on the input data and the controlled variable data and converting (S3) the local model network into a state space representation; - Applying (S4) a feedback linearization to the trained Local Model Network in the state space representation to provide the feedforward control for the controlled variable; and - Providing (S5) the feedforward control for a controlled variable to control the controlled variable of the machine. [11] Device (100) according to claim 10, wherein the machine is a work machine, in particular an excavator, with at least one adjusting cylinder, wherein the input data comprises data about a joystick movement for controlling the at least one adjusting cylinder and preferably further disturbance variables, for example a load and / or a pressure on the at least one adjusting cylinder, wherein the controlled variable comprises a speed of the at least one adjusting cylinder, wherein the input data and the data of the controlled variable are each detectable by sensors of the work machine, and wherein the provided pilot control is designed to control the speed of the at least one adjusting cylinder on the basis of the joystick movement in pilot control mode.

Images