MACHINE IMPLEMENTED METHOD FOR OBTAINING DATA FROM A NON-LINEAR DYNAMIC ESTATE SYSTEM DURING A TEST

专利摘要:
In a machine-implemented method for obtaining data from a non-linear dynamic real-time system during a test run, for example, from an internal combustion engine, propulsion unit, or parts thereof, generates a train of dynamic excitation signals for at least one measurement channel according to previously generated trial design for the test run; and the system output of at least one output channel is measured. In order to enable the rapid and accurate generation of the experimental plans for the global measurement, modeling and optimization of a non-linear dynamic real system, it is proposed that the sequence of dynamic excitation signals be generated by the method of generating an experimental design with a sequence of excitation signals, with output data by feeding the sequence of excitation signals into a model for the real system, the model comprising non-linear dynamic models, and determining a criterion for the information content of excitation signals of the entire experimental planning sequence, and in a following step the entirety of the sequence of Excitation signals is changed, whereby new output data is obtained by feeding the changed sequence of excitation signals into the model for the real system, the criterion for the information content of excitation signals is redetermined, and these S Repeat until the criterion has reached its optimum value, the last generated sequence of excitation signals is used as a test design for the test run of the real system.
公开号:AT511577A2
申请号:T804/2011
申请日:2011-05-31
公开日:2012-12-15
发明作者:Markus Dipl Ing Stadlbauer；Christoph Dipl Ing Dr Techn Hametner；Stefan Dr Jakubek；Thomas Dr Ing Winsel
申请人:Avl List Gmbh；
IPC主号:

专利说明:

AV-3415 AT · * ·] ··· ·· *; · * ·· ··· · · · · · · · · > ·, II I * ·· ♦♦ · * «< ·. ·· ...... I ..
A machine implemented method for obtaining data from a non-linear dynamic real-time system during a test run
A machine implemented method for obtaining data from a non-linear dynamic real-time system during a test run, for example from an internal combustion engine, a power plant or parts thereof, according to the preamble of claim 1.
In the automotive field, there is an ever-increasing need for efficient and accurate models as the calibration of the engine control system becomes more complex and continually expensive due to increasingly stringent regulations. The basic requirements for good models are good measurement data and an appropriately selected Mcssplan. This increases the number of measurements and consequently the measurement period is longer. However, since time at the test stand is very expensive, there is a need for effective design plans that minimize the number of measurement points and cover the test period as effectively as possible while at the same time not reducing the quality of the models trained to use that data. These models are then used to optimize and calibrate the structures of the engine control units (ECU), or to make decisions about components.
Optimal Design of Experiments (OED) optimizes the informational content of excitation signals, which is used to determine the parameters of a model correctly with the least possible effort, as described in L. Pronzato. Optimal experimental design and some related control problems " Automatica 44 (2): 303-325, 2008, is explained. Dynamic excitation signals are characterized by their spatial distribution and their holding times. For identification of linear dynamic systems, pseudo binary noise signals (PRBS) are commonly used, see G. C. Goodwin and R. L. Payne "Dynamic System Identification:"
Experiment Design and Data Analysis'4 Academic Press [nc., New York, 1977. For non-linear dynamic systems, amplitude-modulated pseudo-binary noise (APRBS) signals are established as excitation signals to track nonlinear process identifiers; See, for example, O. Nelle's "Nonlinear System Identification: From Classical Approaches to Neural Networks and Fuzzy Models"; Springer, Berlin, 2001. In contrast to these very general methodologies for experimental designs, Model Based Design (DoE) is more specific to the process to be identified by using a model of a previous process, or at least a model structure, to obtain the information obtained from experiments maximize.
Real processes are subject to restrictions, which in principle
I REPLACED «♦ • · 2. **» * * * • «*« «« «* * * * · · · · · ·
AV-3415 AT outputs. For example, the modified variables and the control variables must not be outside the drivable range to provide specified operating conditions or to prevent damage to the system. To take into account the initial limitations of the system, the experimental design requires a model that predicts the initial dynamics. Model-based experiment design can also be used for on-line experiment design, continuously adapting the model to incoming data and generating experimental design sequentially for a given number of future system inputs. Such a procedure is commonly known as online or adaptive design and is depicted in FIG. Explanations can be found in Online Dynamic Black Box Modeling and Adaptive Experiment Design in Combustion Engine Calibration, Munich, Germany, 2010, and in Laszlo Gerncser, Hakan Hjalmarsson and Jonas Martcnsson .. Identification of arx Systems with nonstationary inputs - asymptotic analysis with application to adaptive input design " Automatica 45: 623-633, March 2009.
The purpose of the method presented here is to provide designs which are tuned to typical applications in the development of engines or powertrains and to the calibration of the ECUs or TCUs. The method is designed to allow fast and accurate generation of design plans for global measurements, modeling and optimization of a non-linear dynamic real-time system, e.g. As an internal combustion engine, a prime mover or subsystems thereof, as well as the global optimization thereof, while Verkrsuchslimits and additional criteria are taken into account.
To fulfill this purpose, the method described above is characterized by the characterizing part of claim 1. Preferred embodiments of this basic concept are given in the dependent claims.
The main advantages of model-based design versus traditional design techniques are that the model is used to optimize the design of experiments so that they are most effective and allow for the inclusion of various constraints, including system exit limitations, as well as online test design.
The invention is described in more detail in the following specification based on preferred examples and with reference to the figures of the enclosed drawing.
Fig. 2 shows a two-layer perceptron with three inputs, two neurons in the hidden layer and an output, Fig. 3 shows a structure of an artificial neural network in output error configuration (NNOE) with a individual input u and death time d, FIG. 4 does not show a | REPLACED |
AV-3415 AT linear constrained optimization with inequality constraint for a two-dimensional input signal u, Fig. 5 illustrates the excitation signal u and the model output y with the output limit exceedance y (k) > 6 shows the structure of the Wiener model, FIG. 7 is a representation of the excitation signal u, the model output y and the input _urate with the associated constraints (shown as dashed lines) for the initial experimental design (I) First Valid Design (II) (all restrictions are kept for the first time) and the Final Design (III), where the initial design exceeds the drivable range (shown as dashed line) of the model output, and Figure 8 shows the iterative magnification of log (det (I)) including the three stages from Fig. 7: Initial Design (I), First Valid Design (II) and Final Design (III).
The following specification explains how to work out an optimal experimental design batch method compliant with the input and output constraints, using a dynamic multilayer perceptron (MLP) for the representation of the non-linear realm system under observation, i. H. the internal combustion engine, the drive unit or the like, which is to be analyzed in a test run, for example on a test stand. The method aims to optimize the Fisher information matrix of an MLP and thus can be applied to a large class of dynamic systems. There are several methods in the literature for optimal experimental design based on artificial neural networks employing candidate sets for the optimization of excitation signals. In DA Cohn's "Neural network exploration using optimal experiment desigri1 Neural Networks 9 (6): 1071-1083, 1996, the optimal design of the experiment is applied to learning problems by allowing the lemer at each time step to create a new system input from a candidate set select. When selecting the new input, either static or dynamic input constraints are taken into account by minimizing the expected value of the mean squared error of the lemer. It is further known that the criterion of D-optimality is applied to the reduction of training data for global dynamic models based on dynamic artificial neural networks. Also, local sequential D-optimal design of the experiment has been proposed and formulated as a model predictive control problem via optimization via a set of allowable system inputs based on the temporal evolution of a given APRBS. The use of MLP networks for model-based experimental design is inspired by the above-mentioned publications and combined with a gradient technique to enhance the excitation signals. In contrast to most approaches in the prior art, the optimization of the design criterion is carried out analytically in this document, so that
«« 4. · ·
AV-3415AT no candidate set is required. The simultaneous optimization of the temporal and spatial development of the request signal and the compliance with the entry and exit restrictions are proposed. In this context, dynamic model-based design planning must take into account the inherent effect of each excitation signal on all future system outputs, so that its impact on the Fisher information matrix and compliance with output constraints can become very demanding.
As a nonlinear model structure, MLP networks are selected as a non-linear model structure because they belong to the class of universal approximators. Other nonlinear model structures that may be used in accordance with the present invention are local modal networks (LMN) or Takagi Sugeno fuzzy models.
In particular, MLP networks with multiple inputs and one output (MISO) are used, as shown in FIG. The free model parameters are given by the input weights W and the output weights! 1, which can be combined to form the parameter vector Θ.
Wio ...
W =:
(9
The model output y is calculated as a weighted sum of nonlinear sigma activation functions fi in the hidden layer. There are ηφ model inputs <pi that make up the regression vector φ.
The approximation capability of the MLP network is determined by the number of neurons in the hidden layer nh. In the nomenclature of vector matrices, the entrance to the hidden layer h and the exit of the hidden layer o are given as follows: | NACHGEREICH &quot; * · • ·
AV-34I5 AT «* • · Λ = Η &quot;
ο (Λ) ~ ft (hi), ./×6 (Α «λ). Ο
The output of the MLP network is a non-linear function g (<p, 0), which depends on the regression vector φ and the parameter vector Θ.
jj = g {tfi, 0) = wT i =! 1-1
Both the inputs to the neurons in the hidden layer and the hidden layer outputs contain the offset terms WjO and tol0. It is assumed that the system output of the actual process y (k) is given by the model y (k, 0) and a Gaussian error e (k) with a mean of zero and variance σ2: y {k) = 5 (v »(Fr), 0) + c (k) = y {k, Θ) + e (k) (7)
So far, it has not been specified whether the MLP network is used for static or dynamic systems. In the following, dynamic MLP networks in output error configuration (NNOE) are observed. Then, in k-th observation, the regression vector q> (k, 0) consists of past network outputs y (k-i) with i = 1 ... n and past system inputs u (kd-j + l) with j = 1 ... m and dead time d (see Fig. 3). φ (^ θ) = $ (£ -1,0) *. .y (fc-n, 0) * ..... i (k-d-m) ... tin, (fc-if), .. (£ -rf-m)) (8)
It is assumed that the excitation signal U consists of N observations and n different inputs. The measured system output y is given as follows: V = [ut ... «nj
Ul (l) ... Unu (l) 'vi {N) ... tinu (AT) (9)
The measured system output y is given as follows:
REPLACED * * *
AV-3415 AT r «o) (lo)
The network parameters 0 must be adapted to the input and output data of the experiment. The training of the network weights is usually done with a standard Levenberg-Marquardt algorithm, which aims to minimize a quadratic cost function VN (0) based on the prediction error c (k):
CH)
The optimal model-based experimental design aims to maximize the information content of the experiments. For this purpose, an optimality criterion, which is usually derived from the Fisher information matrix, is optimized. In this context, the process model in the form of an MLP network is needed for the calculation of the Fisher information matrix. From a statistical point of view, the Fisher information format I makes a statement about the information content of the data in the form of a covariance matrix of the estimated parameters. Common design criteria for model-based experimental design are the trace of Γ1 (A-optimality). the determinant of I (D-optimality) and the smallest eigenvalue of 1 (E-optimality).
Model-based design can be applied if a model of a system already exists and only some parts of the system have been changed, so that a similar behavior can be expected if the system is operated under changed environmental conditions and a model of a similar system is available.
The Fisher matrix is a statistical measure of the information set of the underlying data, and its inverse gives the covariance matrix of the estimated model parameters Θ. This matrix is the derivative of the model output with respect to the model parameters:
I Current model-based trial design criteria use a scalar value of the Fisher matrix or its inversion as the objective function. This requires a measure of the information content to optimize the stimulus signals. Since the Fisher information matrix provides statistical information about the information content of the underlying
SUBSEQUENT
AV-3415 AT
Data or the covariance matrix of the estimated parameters, it is a basis for common design criteria such as A, D and E optimality.
The Fisher information matrix depends on the parameter sensitivity vector | / (k), which describes the functional dependence of the model output on the model parameters: (12) «dead *, *), *) de
The Fisher information matrix comprises the vectors of the parameter sensitivity of all observations k = 1 ... N according to (13) ΐ (Φ) = 4φτφ. <τ where the parameter sensitivity matrix Ψ combines the parameter sensitivity of all observations: Φ =
(M)
The parameter sensitivity is determined separately for the MLP output weights v | / (o (k) and input weights TWik):
A (* J
Bw [o (h (k)) (15) ψω (ί :) e Hl '™ ·'! »1 (16) Φκ (*>« = = 6ά8 (0 '(Λ (*))) ώ [ ΐ <p (k, »)} (17) Φ" · (*) (18)
The vector of initial weight without bias term is expressed by u:
Cf - fu Ii
T (19) (20) (21) t - o '{/ * (fc)) = 1 - o {h (k)) o o (h {k) REPLACED | av-3415at • 1 * »· * · *** ..! * f · »· t φ t
Here, o '(h (k)) indicates a column vector whose i-th element is divided by the derivative of the i-th output of the hidden layer with respect to the i-th input into the hidden layer at the k-th In (17), diag (o '(h (k)) indicates a diagonal matrix whose inputs are the elements of the vector o' (h (k)), and o denotes the Hadamard product. Commonly used optimality criteria based on the Fisher information matrix are the A, D, and E optimalities, and the A optimality aims to minimize the sum of the parameter variances, so the associated design criterion is based on the track of I '. 1: (22)
D-optimality uses the determinant of I, which, in contrast to A-optimality, is more sensitive to individual parameter covariances because the determinant is similar to the product of the eigenvalues. In addition, the D-optimality is invariant under any non-singular reparameterization, which is not dependent on the experiment. ZD (*) = det (r {«)) (23)
In E-optimality, the smallest eigenvalue ληιϊη of the Fisher information matrix is maximized. Ζ «(Φ) = λ ™™ (Ζ (Φ)} (24)
The exclusion criterion J can only be influenced by changing the system inputs ui (k). To optimize the design criteria, a candidate set of allowed inputs is usually generated from which certain inputs are selected to optimize the design criteria.
The proposed method for model-based experimental design implements the minimization of (22) and the maximization of (23) and (24) by analytically calculating the optimized excitation signals taking into account the input and output constraints.
Generally, the enhancement of the excitation signals is aimed at optimizing the design criterion while meeting the constraints. Mathematically, the optimization problem, taking into account input, input rate and output constraints, is expressed as follows:
FOLLOW-UP (25) A-oprimality: J, min D and E-ptimality: Jd ^ Je - * max (26) »&lt; (*)
Jnpat constraim umi «&lt; «, - (*) &lt; iw * (27)
Rate constramr. Διtmin &lt; u * (£ 4 * 1) - u, (ik) &lt; Au luma · (28)
Oolpiit coastnint: ymin &lt; ¢ (¾) &lt; (29)
The enhancement of the design criterion J represents a non-linear optimization task for which different optimization methods are available. For improving the design criterion of linear dynamic systems, a gradient method is known. For use in the present invention, there is an iterative gradient method for optimizing the excitation signals of non-linear dynamic systems. The proposed method for the analytical calculation of the optimized excitation signals is carried out in two steps. First, the design criterion gradient is determined with respect to the dynamic system inputs, and second, the system inputs are recursively updated and the input and output constraints are considered.
The optimization is based on the calculation of the gradient of the objective function represented by the design criterion related to the dynamic system inputs. At each iteration, the excitation signal is updated so that the design criterion is improved while adhering to the constraints.
The derivative of the design criterion J with respect to the i-th system input uifk) is calculated by using the chain rule in three steps: Α / (Φ) _ γ 'dtn (k) ή- * <tJ (») άψ (1 ) ώφτ (1) dui (k) (30)
Ad (i): First, the derivative of the design criterion with respect to the parameter sensitivity vector is needed for the 1-th observation ψ (Γ). For A-optimality and D-optimality, the result is expressed as follows:
| SUBSEQUENT
AV-3415 AT ♦ «= -2β {/) ΦΐΦΓ * Γ * αΐ) άφ (ί) = 2e (!) JD (®) 4 '[®T * i-1 02) s (l) = [Ο. .. 1 ... Oje®1 ** 03) where s (l) represents the single input vector, which is 1 at the 1-th position and 0 everywhere else. The E-optimality requires the calculation of the derivation of the smallest eigenvalue of the Fisher matrix ληύη in relation to the vector of the parameter sensitivity ψ (1), which leads to the following result: = 04)
Here xmin indicates the eigenvector of the smallest eigenvalue λιηίη: 05)
In (31) and (32) shows the inversion of the Fisher matrix. Consequently, the Fisher matrix must be regular to be reversible. It has already been shown that a singular Fisher infonnation matrix based on an MLP network can be made regularly by removing redundant neurons.
Ad (ii): The derivative of the vector of the parameter sensitivity for the output weights ψω (1) with respect to the regression vector φ (1, θ) is given as follows: 0 MW) o
Here, the input weights without bias terms are indicated by ~ W: w = 'Wu ... wv w ^ ηϊ> η <<07) | REPLACED |
And the derivative of the parameter sensitivity matrix for the input weights PW (I) with respect to the i-th component of the regression vector φί (Ι, θ) is determined as follows: (39) where o &quot; (h (l)) is a column vector whose i-th input is given by the second derivative of the output of the hidden layer with respect to the i-th input into the hidden layer in the k-th observation:
O
aVMfcmdh (k) J (40) o &quot; (M0) = -MMO) ° o '(h (i)) (4i)
Here, ei indicates the direction vector in the i-th component of φ (1, θ), and s (i) is again the single input vector. et = j0 ... 1 ... 0] € 3 * (42) s (i) = [0 ... 1 ... Oj.eR1 * &quot; * (43)
Ad (iii): For dynamic autoregressive systems, the regression vector depends not only on past system inputs but also on model outputs. Thus, the derivative of φ '(1 + 1, θ) with respect to ui (l-j) requires the calculation of the derivative of past model outputs y (l, 0) with respect to ui (l-j). Dynamic system inputs ui (k) have, on the one hand, a direct effect on the model output and, on the other hand, an indirect influence over the past n model outputs. Taking this fact into account and using the chain rule, the derivative of y (l, 0) with respect to ui (lj) is given as follows: SUBSEQUENCY dy (l, 0) θ9 (φ (1, θ), β) dyV- Ι, θ)
du {{l-j) dy (l-1, e) dtn (I-j) nemfidndaeoii δ9 (φ (1, θ), θ) m-n. *). θ9 (φα, θ), θ) dy (l-n, 0) <hn {l-j) aui Q-j) 1 3 (45) j> lk = l-j
In the next time step, the result of (44) is used to calculate the derivative of y (l + I, 0) with respect to ui (l-j). In a recursive process, all model outputs needed for the regression vector in 1-th observation are differentiated. The derivative of g (tp (l, 0) .0) with respect to the elements of the regression vector is given in the nomenclature of the vector matrices as follows: in (44)
The derivative of all system inputs uj (s) used in the regression vector φ (1, θ) with respect to ui (k) is calculated as follows:
Here, 6ij stands for the Kronecker delta, which is 1 for i = j and 0 for i Φ j, and where i, j denote the input index from 1 to nu.
The limited recursive optimization of the excitation signal is based on the calculation of the gradient (30) with respect to all the different inputs ui (k) with i = l ... nu for all observations k = 1 ... N. Compliance with input, input rate and output constraints during the optimization process is ensured by a limited gradient technique. The principle of the method is explained in Fig. 4 for a two-dimensional example. Each iteration (indicated by the index v) becomes the
I RETURNED
AV-3415 AT (quadratic) difference between the gradient 5lv) ViJ (v) and the excitation signal increment Aui minimizes * while approaching the allowable range determined by the constraint vector g &lt; 0 is defined is performed. The constraint vector g = [g1 ... go] ∈ R1xo includes all possible restrictions. A constraint is active if gk = 0, and inactive if gk &lt; 0. Here δ (ν) denotes the variable step length of the gradient method. Mathematically, the problem is expressed as follows: (Au, - (A «i - - + min {49)
ATA
The linearization of the active constraints g (v) act is given by g (v) lin, which must be zero: si * f = i9 {: l) T + (^ f) ΤΔ «<= 0 (50) For the active constraint optimization is a scalar lagrange function L defined with the corresponding active multiplier row vector L (v) act. = Ι (Δ << ί-itr> V <(Am + A «r V) T (51)
The extreme value of the Lagrange function is obtained if the derivative of L with respect to Aui is equal to zero: d £ WdAu, dui eci = 0 (52)
Then the change Aui of the excitation signal is expressed as a function of Zact:
Substituting this result into the constraint condition (50) yields for Xact:
| REPLACED I
* * * · · »· · ·«
AV-3415 AT
K.A. &gt; + r &lt;
you i (54)
Using the result for k (v) act and inserting it into (53) gives the final result for the iterative change Aui of the excitation signal.
In the following, the active constraints in (50) are discussed in detail with regard to input, input rate and output limit overruns. Input Restrictions: For example, assuming that the input in k-th observation exceeds the allowable range defined by [umin, umax], then the following restrictions are active:
Ui (k) &gt; llmar. Auftfc) - umair + = 0 (55) tijffc) &lt; TWa: - A-Uj (fc) + Ufntft - <4 *** (k) = 0 (56)
Rate Restrictions: Adherence to the input rate Auirate (k) = ui (k + l) -ui (k) to [Aumin, Aurnax] is expressed by the following conditions: For Auirate (k) &gt; Aumax: (fc + 1) + Διι &lt; (fc + lw (fc) + Au * (£)] - Δ «βΐΜ, = 0 (57) F» Δίί ("" (λ ·) <Aitmin:
Aui (* + l)} + [u <, * ') (i *) + Aiti (fc)] + Aum <"= 0 (58)
Initial Restrictions: The inclusion of exit constraints must take into account the fact that in autoregressive systems, the input ui (k) is all future model outputs y (k + 1). 1 &gt; 1 influenced. When k-th observation results in an output limit violation where y (k) &gt; ymax, then the system inputs ui (l) must be changed with 1_k so that y (k) is within the allowable range, see Fig. 5. This leads to the following limitations:
| SUBSEQUENT
AV-3415 AT y (k) &gt; ifcno *: yw (k) + ^ = 0 (59) dU (> i Ö (fc) <Jtmfn '- Vy' '(k) ~ *' y Au (+ Vmin = 0
The equivalent of a Model Predictive Control problem requires the calculation of future system inputs, taking into account the development of future system outputs. This requires the calculation of the derivative of the model output with respect to the dynamic excitation signal; see equation (44).
The following example demonstrates the effectiveness of the proposed method for model-based design planning with MLP networks using a non-linear dynamic process. It will be shown how the determinant of the Fisher matrix is iteratively improved while respecting input, input rate and output constraints. As an initial experimental design, an APRB signal is generated, which is subsequently improved by the application of the presented method. The proposed optimization process is applied to a (SISO) Wiener model. The Wiener model is described by a serial arrangement of a linear time invariant transfer function G (z-1) and a static non-linearity NL at the system output. MC: y = juct & ii v (61)
Here, the transfer function describes a second-order oscillating system: _ ,, 0.01867 ^ -1 + 0.0174G * -2V (z "1)} ~ 1 - 1.7826s-1 + -1 / ^ - 1) l62 *
The generation of output data is done on the basis of the illustrated Wiener model, as shown in Fig. 6. Then, using a standard tool, e.g. For example, the NNSYS1D Toolbox for Matlab, which trains a reference model of the underlying process. In this case, an MLP network with five neurons can already approach the used Viennese model relatively well.
As initial experimental design for the optimization process, a prior art APRB signal with 100 samples is used. In Fig. 7, the excitation signal u, the input rate, aurate, and the model output y, together with the associated constraints indicated by dashed | SUBSEQUENT i
AV 3415AT
Lines are shown for the initial, the first valid and the final design plan. An initial design plan that violated the exit constraints was chosen to demonstrate the functionality of the proposed limited optimization process. The first valid design is achieved when all restrictions are met for the first time. Here the optimization algorithm is stopped after 40 iterations and the final design plan is reached. In Fig. 8, the associated iterative magnification of the logarithm of the determinant of the Fisher matrix is shown. The determinant of the Fisher matrix decreases at the first iteration because the initial out-of-bounds of the initial design attempt must be compensated, and at the fourth iteration, the first valid design is achieved. Increasing the determinant of the Fisher matrix causes increased excitation of the system output dynamics, as shown in FIG. This is a useful result, as more information is collected when the system is excited throughout the output range. Increasing the information content of the excitation signal is equivalent to reducing the uncertainty of the estimated model parameters.
The present invention proposes a novel method for experimental design based on a multi-layer perceptron for non-linear dynamic systems. The motivation for this work is the creation of an analytical batch process for the optimization of dynamic excitation signals while respecting input, input rate and output constraints needed for online experiment design. The effectiveness of the proposed model-based design approach is demonstrated on a non-linear dynamic system by showing the iterative magnification of the determinant of the Fisher matrix, resulting in a reduction in the uncertainty of the estimated model parameters. The simulation example shows that the optimization of the information content of the excitation signal leads to an increased system output dynamics. The presented method for optimizing excitation signals also latches input, input rate and output constraints, which is a prerequisite for an online experimental design process.
SUBSEQUENT

权利要求:
Claims (7)
[1]
claims; A machine implemented method for obtaining data from a non-linear dynamic real-time system during a test run, for example, from an internal combustion engine, a power plant or parts thereof, comprising generating a train of dynamic excitation signals for at least one measurement channel according to a previously generated trial design for the test run and measuring the system output of at least one output channel, characterized in that the sequence of dynamic excitation signals has been generated by the method of initiating a trial design with a sequence of excitation signals, wherein output data is obtained by injecting the sequence of excitation signals into a model for the real system in which the model comprises non-linear dynamic models, and wherein a criterion for the information content of excitation signals of the entire experimental planning sequence is determined, and in a following step the entirety of the Sequence of excitation signals is changed, whereby new output data is obtained by feeding the changed sequence of excitation signals into the model for the real system, the criterion for the information content of excitation signals is redetermined, and these steps are repeated until the criterion has reached its optimum value in which the last generated sequence of excitation signals is used as experimental design for the test run of the real system.
[2]
2. A method according to claim 1, characterized in that during each step, compliance with the limits of the excitation signals and / or the model output is checked, whereby the sequence of the excitation signals is changed in the event of non-compliance such that the compliance is restored and the criterion simultaneously approaches its optimum value.
[3]
Method according to claim 1 or 2, characterized in that after each iteration the derivative of the criterion with respect to the dynamic excitation signal is determined, and the iterations are stopped as soon as the derivative falls below a predetermined value or a predetermined number of iterations is reached ,
[4]
4. The method according to any one of claims 1 to 3, characterized in that at each iteration, the simultaneous optimization of the spatial distribution of the experimental points or measurement points and the temporal evolution of the excitation signal is allowed. I REPLACED &lt; «*« «AV-3415AT
[5]
A method according to any one of claims 1 to 4, characterized in that the criterion is determined by the Fisher information matrix, in particular by calculating the track of the inversion of the matrix, by calculating the detection edge or the least eigenvalue.
[6]
6. The method according to any one of claims 1 to 5, characterized in that the output data are determined with a model using multi-layer perceptron networks (MLP) as the non-linear dynamic model architecture.
[7]
Method according to one of claims 1 to 5, characterized in that the output data are determined with a model using a local model network (LMN) or a Takagi Sugeno fuzzy model as the non-linear dynamic model architecture. POSSIBLE 1 AVL List GmbH Fig. 1 offline DoE online DoE existmj; modei Fig. 2 tlfnal rmtifn DoE

FIG. 3, 9 i y {k~1) 3 u (k-d) neural network y (k) u (k-d-m) FIG. 4

NAOHGEREICHT • f II I »M II I II ***** ft * * * * · * ··« »&gt; * * * * • • · I * 1 * 1 · *« I * «t« ft * ft t &gt; ···· 3 AVL List GmbH Initial Design (I) Fig. 7

Samples | SUBSEQUENTLY Fig. 5 • * ·

2 AVL List GmbH

Fig. 6 u

Fig. 8

submitted

类似技术:

---

公开号 | 公开日 | 专利标题

---

AT511577A2|2012-12-15|MACHINE IMPLEMENTED METHOD FOR OBTAINING DATA FROM A NON-LINEAR DYNAMIC ESTATE SYSTEM DURING A TEST RUN

---

DE102017000536B4|2018-10-11|Cell controller for detecting a cause of an abnormality in a manufacturing machine

---

DE102006008482B4|2010-10-07|Input vectoring method in neural networks for auto-identification of a partial discharge source

---

EP2112568B1|2013-05-01|Method for computer-supported control and/or regulation of a technical system

---

DE102013202378B4|2015-07-02|Procedural memory learning and robot control

---

DE102007001025A1|2008-07-03|Method for computer aided control and regulation of technical system, involves carrying out characterization of dynamic behavior of technical systems multiple times by state and action of system

---

EP1052558A1|2000-11-15|Method and device for estimation of condition

---

DE19808197A1|1999-09-09|System and method for diagnosing engine conditions

---

EP1715352A2|2006-10-25|Method and apparatus for diagnosing failures in a mechatronic system

---

EP1327959B1|2006-09-27|Neural network for modelling a physical system and method for building the neural network

---

EP3590011B1|2021-03-31|Method and control device for controlling a technical system

---

EP1021793A2|2000-07-26|Assembly of interconnected computing elements, method for computer-assisted determination of a dynamics which is the base of a dynamic process, and method for computer-assisted training of an assembly of interconnected elements

---

WO2005081076A2|2005-09-01|Method for the prognosis of the state of a combustion chamber using a recurrent, neuronal network

---

DE102018115413B4|2022-02-10|THERMAL SHIFT COMPENSATION DEVICE AND MACHINE LEARNING DEVICE FOR LEARNING FORM MEASUREMENT DATA

---

DE102017006687A1|2019-01-17|Test system for routine testing of test specimens and procedures

---

DE102011081345A1|2013-02-28|Method for creating a model

---

WO2019057489A1|2019-03-28|Method and training data generator for configuring a technical system, and control device for controlling the technical system

---

DE102019123447B4|2021-09-02|Method for operating vehicle components

---

DE102019114378A1|2020-02-27|METHOD AND SYSTEM FOR PREDICTING THE SYSTEM STATUS

---

DE102020204140A1|2021-09-30|Device and automated method for evaluating sensor readings and using the device

---

DE102018109835A1|2019-10-24|Method and device for determining a network configuration of a neural network

---

DE102020201544A1|2021-08-12|Processing of measurement data

---

DE102020205962B3|2021-06-24|Device and method for operating a test bench

---

DE102019201045A1|2020-07-30|Method, device and computer program for determining an action or trajectory of a robot

---

DE102018109851A1|2019-10-24|Method and device for determining a network configuration of a neural network

同族专利:

---

公开号 | 公开日

---

EP2715459A1|2014-04-09|

---

US9404833B2|2016-08-02|

---

US20140067197A1|2014-03-06|

---

EP2715459B1|2015-08-19|

---

JP2014519118A|2014-08-07|

---

JP5885831B2|2016-03-16|

---

WO2012163972A1|2012-12-06|

---

AT511577A3|2014-12-15|

---

AT511577B1|2015-05-15|

引用文献:

---

公开号 | 申请日 | 公开日 | 申请人 | 专利标题

---

DE102016120052A1|2016-10-20|2018-04-26|Technische Universität Darmstadt|Method for determining bases of a test plan|JP2005522778A|2002-04-10|2005-07-28|ケンドロラボラトリープロダクツ，リミテッドパートナーシップ|Data-centric automation|

---

US8065022B2|2005-09-06|2011-11-22|General Electric Company|Methods and systems for neural network modeling of turbine components|

---

JP4764362B2|2007-02-15|2011-08-31|日本放送協会|Event discrimination device and event discrimination program|

---

JP5085175B2|2007-03-30|2012-11-28|公益財団法人鉄道総合技術研究所|Method for estimating dynamic characteristics of suspension system for railway vehicles|

---

US8024610B2|2007-05-24|2011-09-20|Palo Alto Research Center Incorporated|Diagnosing intermittent faults|

---

JP2010086405A|2008-10-01|2010-04-15|Fuji Heavy Ind Ltd|System for adapting control parameter|

---

JP5405252B2|2009-09-18|2014-02-05|本田技研工業株式会社|Learning control system and learning control method|DE102012005197B3|2012-03-16|2013-06-13|Iav Gmbh Ingenieurgesellschaft Auto Und Verkehr|Method for optimizing an internal combustion engine|

---

AT512003A3|2013-01-23|2014-05-15|Avl List Gmbh|Method for determining a control-technical observer for the SoC|

---

AT512251B1|2013-02-28|2014-08-15|Avl List Gmbh|Method of designing a nonlinear controller for non-linear processes|

---

AT512977B1|2013-05-22|2014-12-15|Avl List Gmbh|Method for determining a model of an output of a technical system|

---

DE102015207252A1|2015-04-21|2016-10-27|Avl List Gmbh|Method and device for model-based optimization of a technical device|

---

EP3304218A4|2015-06-05|2019-01-02|Shell International Research Maatschappij B.V.|System and method for background element switching for models in model predictive estimation and control applications|

---

AT517251A2|2015-06-10|2016-12-15|Avl List Gmbh|Method for creating maps|

---

DE102015223974A1|2015-12-02|2017-06-08|Robert Bosch Gmbh|Method and device for influencing vehicle behavior|

---

US10635813B2|2017-10-06|2020-04-28|Sophos Limited|Methods and apparatus for using machine learning on multiple file fragments to identify malware|

---

CN107942662B|2017-11-16|2019-04-05|四川大学|Finite time state feedback controller design method and device|

---

WO2019145912A1|2018-01-26|2019-08-01|Sophos Limited|Methods and apparatus for detection of malicious documents using machine learning|

---

US11270205B2|2018-02-28|2022-03-08|Sophos Limited|Methods and apparatus for identifying the shared importance of multiple nodes within a machine learning model for multiple tasks|

---

US20190302707A1|2018-03-28|2019-10-03|Mitsubishi Electric Research Laboratories, Inc.|Anomaly Detection in Manufacturing Systems Using Structured Neural Networks|

---

CN110554683B|2019-09-09|2020-12-18|北京航天自动控制研究所|Multi-mode self-adaptive dynamic excitation adding method in periodic control process|

法律状态:

优先权:

---

申请号 | 申请日 | 专利标题

---

ATA804/2011A|AT511577B1|2011-05-31|2011-05-31|MACHINE IMPLEMENTED METHOD FOR OBTAINING DATA FROM A NON-LINEAR DYNAMIC ESTATE SYSTEM DURING A TEST RUN|ATA804/2011A| AT511577B1|2011-05-31|2011-05-31|MACHINE IMPLEMENTED METHOD FOR OBTAINING DATA FROM A NON-LINEAR DYNAMIC ESTATE SYSTEM DURING A TEST RUN|

---

EP12725374.8A| EP2715459B1|2011-05-31|2012-05-30|Machine-implemented method for obtaining data from a nonlinear dynamic real system during a test run|

---

JP2014513175A| JP5885831B2|2011-05-31|2012-05-30|Method of acquiring data from a non-linear dynamic real system during a test run implemented on a machine|

---

US14/111,092| US9404833B2|2011-05-31|2012-05-30|Machine-implemented method for obtaining data from a nonlinear dynamic real system during a test run|

---

PCT/EP2012/060156| WO2012163972A1|2011-05-31|2012-05-30|Machine-implemented method for obtaining data from a nonlinear dynamic real system during a test run|