METHOD AND DEVICE FOR USING AND PRODUCING MULTI-DIMENSIONAL CHARACTERISTIC MAPS FOR CONTROLLING AND REGULATING TECHNICAL DEVICES

Abstract

A computer-implemented method for operating a technical device with the aid of a multi-dimensional characteristic map. The characteristic map is defined by data points, to each of which a characteristic field value is assigned. For reading out the characteristic map, an output value is determined, as a function of an input variable point to be evaluated for the technical device, with the aid of one-dimensional basis functions, which are assigned to each dimension of a data point. The function values of the one-dimensional basis functions respectively have a monotone curve to a neighboring data point, which has the function value 0, and are outside of the neighboring data point 0. The technical device is operated as a function of the output value.

Claims

1-11. (canceled)

12. A computer-implemented method for operating a technical device with a control unit and using a multi-dimensional characteristic map stored in a characteristic map memory, the method comprising the following steps: providing, by the control unit, for an ascertainment of an operating parameter which represents a correction parameter using the characteristic map, an adaptation parameter or a function value of a function mapping a physical behavior and with which the technical device is operated, wherein the characteristic map is defined by data points, to each of which a characteristic field value is assigned; for reading out the characteristic map, determining an output value as a function of an input variable point to be evaluated for the technical device, using one-dimensional basis functions, which are assigned to each dimension of each data point of the characteristic map, wherein function values of the one-dimensional basis functions respectively have a monotone curve to a neighboring data point, at which the basis function has a function value 0, and are outside of a region between the data point and the neighboring data point 0, wherein, for the input variable point to be evaluated, the function values of the one-dimensional basis functions of the data points surrounding the input variable point with respect to every dimension are multiplied in order to determine the output value; and operating the technical device as a function of the output value.

13. The method as recited in claim 12, wherein the input variable point has more than two dimensions, and wherein for computation of the output value for the input variable point having more than two dimensions, multiplication results of function values of the one-dimensional basis functions are stored and used repeatedly.

14. The method as recited in claim 12, wherein the data points of the characteristic map form an unstructured lattice, which includes basis units as simplexes that connect a number of directly neighboring data points to one another, which is greater by 1 than the dimensionality of the characteristic map, wherein for computing the output value, a transformation of an n simplex surrounding the input variable point to an n+1 dimensional space is performed as a function of an input variable point and the simplex is transformed to a corresponding unit simplex, wherein the transformation is described by a multiplication with a (n+1)×(n+1) projection matrix, which results from projecting nodes of the simples, the output value resulting from the multiplication of the projection matrix with the input variable point complemented by a component having a value 1.

15. A system for operating a technical device with a control unit and using a multi-dimensional characteristic map stored in a characteristic map memory, the system comprising: the control unit; and the characteristic map memory in which the multi-dimensional characteristic map is stored; wherein the control unit provides, for ascertainment of an operating parameter which represents a correction parameter with the aid of a characteristic map, an adaptation parameter or a function value of a function mapping a physical behavior and with which the technical device is operated, wherein the characteristic map is defined by data points, to each of which a characteristic field value is assigned, wherein the system is configured to determine, for reading out the characteristic map, an output value, as a function of an input variable point to be evaluated for the technical device, using one-dimensional basis functions, which are assigned to each dimension of each data point of the characteristic map, wherein function values of the one-dimensional basis functions respectively have a monotone curve to a neighboring data point, at which the basis function has a function value 0, and are outside of a region between the data point and a neighboring data point 0, to multiply, for the input variable point, the function values of the one-dimensional basis functions of the data points surrounding the input variable point with respect to every dimension in order to determine the output, and to operate the technical device as a function of the output value.

16. A computer-implemented method for providing a multi-dimensional characteristic map for operating a technical device, wherein the characteristic map is defined by data points, to each of which a characteristic field value is assigned, the method comprising the following steps: determining an output value as a function of an input variable point to be evaluated for the technical device, using one-dimensional basis functions which are assigned to each dimension of each data point of the characteristic map, wherein function values of the one-dimensional basis functions respectively have a monotone curve to a neighboring data point, at which the basis function has a function value 0, and are outside of a region between the data point and neighboring data point 0; calibrating or adapting the characteristic map using one or multiple specified input variable points and respectively associated output values in that characteristic map values are adapted so as to minimize a total error between the output values at the input variable points and the output values of the characteristic map for the input variable points.

17. The method as recited in claim 16, wherein the data points of the characteristic map form an unstructured lattice, which includes basis units as simplexes that connect a number of directly neighboring data points to one another, which is greater by 1 than a dimensionality of the characteristic map, wherein the basis functions of the unstructured lattice are ascertained via the simplexes from selected data points, a density of the distribution of the selected data points being selected in such a way that an expected behavior of the output value can be mapped by linear interpolation between the data points.

18. A system for providing a multi-dimensional characteristic map for operating a technical device, wherein the characteristic map is defined by data points, to each of which a characteristic field value is assigned, wherein an output value is determined as a function of an input variable point to be evaluated for the technical device, using one-dimensional basis functions which are assigned to each dimension of each data point of the characteristic map, wherein function values of the one-dimensional basis functions respectively have a monotone curve to a neighboring data point, at which the basis function has a function value 0, and are outside of a region between the data point and a neighboring data point 0, the system being configured to calibrate or to adapt the characteristic map using one or multiple specified input variable points and respectively associated output values by adapting the characteristic map values so as to minimize a total error between the output values at the input variable points and the output values of the characteristic map for the input variable points.

19. A non-transitory machine-readable storage medium on which is stored a computer program for operating a technical device using a multi-dimensional characteristic map stored in a characteristic map memory, the computer program, when executed by a processing unit, causing the processing unit to perform the following steps: providing, for an ascertainment of an operating parameter which represents a correction parameter using the characteristic map, an adaptation parameter or a function value of a function mapping a physical behavior and with which the technical device is operated, wherein the characteristic map is defined by data points, to each of which a characteristic field value is assigned; for reading out the characteristic map, determining an output value as a function of an input variable point to be evaluated for the technical device, using one-dimensional basis functions, which are assigned to each dimension of each data point of the characteristic map, wherein function values of the one-dimensional basis functions respectively have a monotone curve to a neighboring data point, at which the basis function has a function value 0, and are outside of a region between the data point and the neighboring data point 0, wherein, for the input variable point to be evaluated, the function values of the one-dimensional basis functions of the data points surrounding the input variable point with respect to every dimension are multiplied in order to determine the output value; and operating the technical device as a function of the output value.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Specific example embodiments of the present invention are explained in greater detail below on the basis of the figures.

[0023] FIG. 1 shows a schematic illustration of a control device with access to a characteristic map memory for operating a technical device, according to an example embodiment of the present invention.

[0024] FIG. 2 shows a schematic illustration of a two-dimensional characteristic map, according to an example embodiment of the present invention.

[0025] FIG. 3 shows the curve of basis functions with respect to one dimension of the characteristic map, according to an example embodiment of the present invention.

[0026] FIG. 4 shows a tree structure for simplifying the computation of the function value of the multi-dimensional basis function, according to an example embodiment of the present invention.

[0027] FIG. 5 shows a schematic illustration of an unstructured characteristic map having arbitrarily distributed data points in two dimensions, according to an example embodiment of the present invention.

[0028] FIG. 6 shows an exemplary form of a data point lattice with local refinement, according to an example embodiment of the present invention.

[0029] FIG. 7 shows an illustration of linear basis functions of an unstructured two-dimensional characteristic map, according to an example embodiment of the present invention.

[0030] FIG. 8 shows an illustration of a triangle in barycentric coordinates formed by data points of the unstructured characteristic map, according to an example embodiment of the present invention.

[0031] FIG. 9 shows an illustration of the extrapolation in unstructured lattices, according to an example embodiment of to present invention.

DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS

[0032] FIG. 1 shows a block diagram for illustrating a system 1 for controlling a technical device 2 with a control unit 3. The control unit 3 is connected to a characteristic map memory 4, in which at least one characteristic map is stored in a parameterized manner.

[0033] For operating the technical device 2, control unit 3 provides for ascertaining an operating parameter B, which may represent a correction parameter, an adaptation parameter or a function value of a function mapping a physical behavior. For ascertaining the operating parameter B, the control unit 2 uses the characteristic map in the characteristic map memory 4 and operates the technical device 3 in accordance with the ascertained operating parameter B.

[0034] FIG. 2 shows an example for such a characteristic map including input variables x1, x2, which defined a lattice, and an output-side operating parameter as output variable y, the respective output values of which are symbolized by the filled in circles at the lattice intersections. The coordinates, to which respectively an output value of the output variable (of operating parameters to be ascertained) is assigned, correspond to the lattice intersections and are called data points.

[0035] For each input variable point, a multi-dimensional basis function is defined, which is a product of the individual basis functions. An output value of the output variable may thus be computed from a characteristic map as:

[00001]$f^{\prime }\left(\stackrel{.fwdarw.}{x}\right)=\underset{i}{.Math.}{b}_i\left(\stackrel{.fwdarw.}{x}\right)y_i$

[0036] where the index i takes into account each of the data points of the characteristic map lattice.

[0037] The basis functions b.sub.i are computed as products of the one-dimensional basis functions at the input value of the respective dimension of the input variable of the characteristic map.

[0038] For an individual dimension x, the basis functions, as shown in FIG. 3, correspond to the following definition:

[00002]$\underset{i}{.Math.}{b}_i^x\left(x\right)=1\forall x\in {\Omega }_x$$b_i^x\left(x_i\right)=1\forall i$$b_i^x\left(x_k\right)=0\forall i\ne k$

[0039] Accordingly, the multi-dimensional basis function is then ascertained by multiplication

b.sub.i()=b.sub.i.sup.x.sup.i(x.sub.1)b.sub.i.sup.x.sup.2(x.sub.2)

[0040] For training such a characteristic map, an output value of the output variable y=f′(x) is assigned to a data point. For this purpose, a learning algorithm receives an operating parameter to be learned at a specific data point x.sub.1, x.sub.2, . . . , whereby the latter can be used to improve or enter the existing learned values.

[0041] After a sufficient number of learning events, the characteristic map may indicate a correct output value of an output variable in accordance with a predefined input variable point (input variable vector). If the characteristic map is to exhibit a PT1 behavior, the output value output by the characteristic map will tend in the direction of the actual operating parameter to be learned, according to:

f′({right arrow over (x)}).fwdarw.({right arrow over (x)})

[0042] If an integrating behavior is to be stored, a discrete integral of the input variable point {right arrow over (x)} results as the output value of the characteristic map

[00003]$f^{\prime }\left(\stackrel{.fwdarw.}{x}\right)=\underset{\tau }{.Math.}f\left(\stackrel{.fwdarw.}{x},\tau \right)K$

[0043] where K is an integration speed parameter and τ corresponds to the preceding discrete time steps. No continuous function is available, however, as the output f′ of the characteristic map; rather, the output values for corresponding input variable points must be approximated on the basis of the characteristic map values at the data points (lattice intersections of the characteristic map or entries at the data points). Hence

[00004]$f^{\prime }\left(\stackrel{.fwdarw.}{x}\right)=\underset{i}{.Math.}{b}_i\left(\stackrel{.fwdarw.}{x}\right)y_i$

[0044] where b.sub.i({right arrow over (x)}) are basis functions and y.sub.i the corresponding discrete learned characteristic map values at the data points of the characteristic map.

[0045] During an online learning step, a measurement f({right arrow over (x)}) is evaluated. First, a residual error δ is computed, which represents the error of the currently learned value. An integrator behavior corresponds to δ=f({right arrow over (x)}). For a PT1 behavior, δ=f({right arrow over (x)})−f′({right arrow over (x)}), which corresponds to the difference between the characteristic map value of the characteristic map and the output value currently to be learned at the input variable point of the measurement.

[0046] Next, the learned characteristic map values y.sub.i at the data points are modified in such a way that f′({right arrow over (x)}) corresponds better to the correct output values defined above, i.e. the residual error is compensated. This is achieve in that the basis functions are used as weights for modifying the learned characteristic map values

y.sub.i.fwdarw.y.sub.i+Kb.sub.i′({right arrow over (x)})δ

[0047] where K represents a learning speed and to which K in

[00005]$f^{\prime }\left(\stackrel{.fwdarw.}{x}\right)=\underset{\tau }{.Math.}f\left(\stackrel{.fwdarw.}{x},\tau \right)K$

[0048] may correspond as the integration speed parameter.

[0049] During the offline learning, the learned characteristic map values y.sub.i are determined in such a way that the outputs f′({right arrow over (x)}) agree best with the output value of the characteristic map for the input variable point (evaluation point) .

[0050] This may be carried out by the method of the smallest squares in accordance with

[00006]$\left\{y_i\right\})=\arg \underset{y}{\min }{.Math.X.Math.\stackrel{.fwdarw.}{y}-\stackrel{.fwdarw.}{\beta }.Math.}^2$

[0051] where the matrix elements are given by

X.sub.ji=b.sub.i({right arrow over (x)}.sub.j)

[0052] The sum operation of the equation

[00007]$f^{\prime }\left(\stackrel{.fwdarw.}{x}\right)=\underset{i}{.Math.}{b}_i\left(\stackrel{.fwdarw.}{x}\right)y_i$

[0053] is performed in every line in the product x.Math.{right arrow over (y)}.

[0054] A learned characteristic map value y.sub.i thus exists for every basis function b.sub.i(). These basis functions b.sub.i() are selected in order to span a multi-dimensional volume Ω, in which a learning operation is to be performed.

[0055] As seen in FIG. 2, the basis functions are efficiently defined on a structured rectangular data point lattice, which is shown for a two-dimensional characteristic map in the input-side variables x1 and x2.

[0056] The lattice points are indicated by all combinations of the points {x.sub.1}, {x.sub.2}, i.e. all gray circles in FIG. 2. The rectangle formed in this manner, which is spanned by the data points in two dimensions (cuboid for more than two dimensions) defines the input variable range Ω.

[0057] A multidimensional basis function b.sub.i is defined for every lattice point {right arrow over (x)}.sub.1. The basis functions b.sub.i are computed as products of the one-dimensional basis functions according to every dimension of the input variables of the characteristic map. For an individual dimension x, the basis functions, as shown in FIG. 3, are defined as indicated above. Accordingly, the multi-dimensional basis function is then ascertained by multiplication

b.sub.i()=b.sub.i.sup.x.sup.1(x.sub.1)b.sub.i.sup.x.sup.2(x.sub.2)

[0058] The properties given in the above definition of the basis functions may then be expanded to the higher dimensionalities

[00008]$\underset{i}{.Math.}{b}_i^x\left(\stackrel{.fwdarw.}{x}\right)=1\forall x\in {\Omega }_x$$b_i^x\left({\stackrel{.fwdarw.}{x}}_i\right)=1\forall i$$b_i^x\left({\stackrel{.fwdarw.}{x}}_k\right)=0\forall i\ne k$

[0059] At a certain input variable point {right arrow over (x)} the basis functions, which correspond to 2.sup.N multi-dimensional data points, are unequal to 0, where N represents the number of dimensions. Thus, 2.sup.N learned values are accessed for the interpolation or modified by a learning step. The multi-dimensional data points {right arrow over (x)}.sub.1 comprise products of the one-dimensional basis functions. For every dimension, the one-dimensional basis functions of a low (index l) and an upper (index u) data point are taken into account, which enclose the input variable point {right arrow over (x)} to be evaluated. For example, in the case of three dimensions, the eight (2.sup.3) multi-dimensional basis functions correspond to the eight corners of the cuboid, which enclose the input variable point {right arrow over (x)}: [0060] b.sub.1=b.sub.x.sup.lb.sub.y.sup.lb.sub.z.sup.l, b.sub.2=b.sub.x.sup.lb.sub.y.sup.lb.sub.z.sup.u, b.sub.3=b.sub.x.sup.lb.sub.y.sup.ub.sub.z.sup.l, b.sub.4=b.sub.x.sup.lb.sub.y.sup.ub.sub.z.sup.u, [0061] b.sub.5=b.sub.x.sup.ub.sub.y.sup.lb.sub.z.sup.l, b.sub.6=b.sub.x.sup.ub.sub.y.sup.lb.sub.z.sup.u, b.sub.7=b.sub.x.sup.ub.sub.y.sup.ub.sub.z.sup.l, b.sub.8=b.sub.x.sup.ub.sub.y.sup.ub.sub.z.sup.u,

[0062] where the index “l” corresponds to the lower and the index “u” corresponds to the upper data point. Since products are formed multiple times in the computation of the multi-dimensional basis functions, it is possible to use a computation tree-based approach, as illustrated in FIG. 4, so as to be able to exclude double multiplications. Instead of the 2.sup.N (N−1) multiplications, which are given in the above equation, this makes it possible to reduce complexity to Σ.sub.i=2.sup.N2.sup.i multiplications, which is relevant especially for higher dimensionalities.

[0063] An extrapolation of the output value on input variable points to be evaluated that lie outside of the input variable space Ω is performed by projecting the input variable point onto a limit of the input variable space Ω. Since the input variable space Ω is always convex, this projection is unambiguous.

[0064] In contrast to the above-described exemplary embodiment, characteristic maps may also be unstructured, i.e. have no hypercuboid contour. This may be expedient if the value of the input variable point (evaluation point) to be learned exists only for a non-cuboid set of data points of the input variable space. In a cubically arranged lattice, it may otherwise happen that the output values to be learned for the input variable points are not distributed over the entire input variable space and that thus some output value are never updated or accessed. This results on the one hand in a waste of resources, since the unused learned operating parameters must be stored, and on the other hand the learned values are not extrapolated in these regions during the readout, since they are not located in the extrapolation region, i.e. are not located outside of the input variable space Ω. Instead, a learned setpoint value, such as e.g. zero, is output in these regions, exactly as in the corresponding extrapolation regions.

[0065] In addition, the resolution of the learned values cannot be selected as desired with the aid of the above-described routines. With the aid of rectangular lattices, the data points can only be refined in a dimension-wise manner. The refinement in one dimension will thus be applied to all combinations of the other dimensions, regardless of whether this is necessary or not. This results in a waste of resources, since high resolutions are unnecessarily introduced in operating areas where these are not necessary. Unnecessarily high resolutions may also result in lower performance and noise suppression, since measuring noise is falsely interpreted as spatial variation.

[0066] An approach for applying unstructured characteristic maps to the learning algorithm described above is described below. The data point lattices of the characteristic maps may be selected to describe arbitrary forms and resolutions with the aid of simplexes, i.e. 1-D line segments, 2-D triangles, 3-D tetrahedrons etc., as basis units. The approach may be applied to any desired number of dimensions. In the above-described learning and evaluation approach for a cuboid data point distribution, an input variable space Ω may be spanned by the data points of the characteristic map {right arrow over (x)}.sub.i. For each data point of the characteristic map, a value y.sub.i to be learned is stored. Learning and reading out are performed with the aid of basis functions b.sub.i(). The basis functions b.sub.i() are defined as indicated above.

[0067] Above, the data points were defined on a rectangular characteristic map lattice, which is defined by the individual data points for every dimension. For the application of the approach described above, the data points of unstructured characteristic maps are spanned by independent data points, as shown in FIG. 5 by way of example. Every data point {right arrow over (x)}.sub.i is described by a vector that is independent of all other data points. Lattice cells Ω.sub.k are defined as simplexes, which connect n+1 data points with one another. Such a data point lattice may take an arbitrary form and may be refined locally, as is illustrated by way of example in FIG. 6. The corresponding linear basis functions are graphically illustrated in FIG. 7 for two dimensions.

[0068] The computation of the linear basis functions of unstructured characteristic map lattices may be performed efficiently with the aid of barycentric coordinates. For this purpose, a transformation of an n-simplex into an n+1-dimensional space is performed, and the simplex is transformed onto a corresponding unit simplex. For example, a 2-D triangle may be transformed onto the unit 2-simplex in three dimensions, as shown in FIG. 8. For an arbitrary n-simplex Ω.sub.k, the transformation may be described by a multiplication with a (n+1)×(n+1) matrix.

{right arrow over (λ)}.sub.k=P.sub.k.Math.{right arrow over (x)}′

[0069] Here {right arrow over (x)}′ corresponds to a (n+1)-dimensional vector as a function of the n-dimensional vector, {right arrow over (x)}, to which a component with the value 1 is appended, e.g. (x1, x2, 1). The values of P.sub.k are obtained by projecting the nodes of the simplex e.g. for Ω.sub.1 in

[0070] FIG. 6

P.sub.1.Math.{right arrow over (x)}′.sub.1=(1,0,0),

P.sub.1.Math.{right arrow over (x)}′.sub.2=(0,1,0),

P.sub.1.Math.{right arrow over (x)}′.sub.3=(0,0,1),

i.e., the columns of the inverse matrix P.sup.−1.sub.1 correspond to the coordinates of the nodes of the simplex, to which 1 is appended.

[0071] The barycentric coordinates have the following advantages: [0072] Only when an input variable point {right arrow over (x)} lies within a simplex Ω.sub.k or at its limit, do all components of {right arrow over (λ)}.sub.k become greater than or equal to zero. This may be used for the efficient search for a simplex, in which an evaluation point {right arrow over (x)} is located. [0073] The sum of all components of each {right arrow over (λ)}.sub.k is always 1. [0074] If {right arrow over (x)}∈Ω.sub.k, then the components of the projected {right arrow over (λ)}.sub.k are equal to the values of the linear basis functions according to the corners of the simplex Ω.sub.k at the input variable point {right arrow over (x)}. Thus, one obtains the values of the basis functions directly through the transformation to the barycentric coordinates.

[0075] The basis functions in unstructured lattices may be ascertained via the simplexes from the selected data points. The data points are selected in such a way that firstly they cover the expected range of the input variable point and that secondly the density of their distribution is sufficiently high that the expected behavior of the output value may be mapped by linear interpolation between the data points.

[0076] An extrapolation from the unstructured characteristic map lattices cannot be performed in a simple manner, as described above, because the characteristic map lattice is not necessarily convex and therefore an unambiguous projection onto the limit does not always exist. Accordingly, for unstructured characteristic map lattices, it is proposed to carry out the following method in order to obtain continuous values for data points outside of the discretized input variable space Ω. This makes it possible further to avoid jumps in the output values for continuously changing input variable points.

[0077] The oriented edges L.sub.k form the limit of the input variable space Ω, with the outwardly directed normals {right arrow over (n)}.sub.k, as illustrated in FIG. 9. For a specific input variable point {right arrow over (x)}.Math.Ω to be evaluated, all edges L.sub.k,out, for which the input variable point x is outside of the edge, may be ascertained:

({right arrow over (x)}−{right arrow over (x)}.sub.k).Math.{right arrow over (n)}.sub.k>0

[0078] where {right arrow over (x)}.sub.k is a point on the edge L.sub.k, e.g. one of the limit nodes. For each of these edges, the edge point {right arrow over (x)}.sub.near on the edge L.sub.k is determined, which is closest to the input variable point {right arrow over (x)} to be evaluated. This point may be on the edge or on a limit node of the edge. The corresponding output value for the extrapolation is the interpolated value at position {right arrow over (x)}.sub.near, whereby a weighting, given by

[00009]${\omega }_k=\frac{\cos \phi }{d}$

[0079] is taken into account. Here, d is the Euclidean distance between {right arrow over (x)} and the edge point {right arrow over (x)}.sub.near, and δ is the angle between the normal {right arrow over (n)} and ({right arrow over (x)}−{right arrow over (x)}.sub.near) The extrapolated output value y′ may then be computed as

[00010]$y^{\prime }=\frac{{.Math.}_k^{L_{k,\mathrm{out}}}{y}_k{\omega }_k}{{.Math.}_k^{L_{k,\mathrm{out}}}{\omega }_k}$