Nikolaus Keuth

A local neuro-fuzzy network for high-dimensional models and optimization

Abstract In this paper a new iterative construction algorithm for local model networks is presented. The algorithm is focussed on building models with sparsely distributed data as they occur in engine optimization processes. The validity function of each local model is fitted to the available data using statistical criteria along with regularization and thus allowing an arbitrary orientation and extent in the input space. Local models are consecutively placed into those regions of the input space where the model error is still large thus guaranteeing maximal improvement through each new local model. The orientation and extent of each validity function are also adapted to the available training data such that the determination of the local regression parameters is a well-posed problem. The regularization of the model can be controlled in a distinct manner using only two user-defined parameters. In order to assess the quality of the obtained model, the algorithm also provides accurate model statistics. Different examples illustrate the efficiency of the proposed algorithm. One illustrative example shows how local models are adapted to the shape of the target function in the presence of noise. A second example shows results obtained with measurement databases of IC-engines.

Total least squares in fuzzy system identification: An application to an industrial engine

Takagi-Sugeno fuzzy models have proved to be a powerful tool for the identification of nonlinear dynamic systems. Their generic nonlinear model representation is particularly useful if information about the structure of the nonlinearity is available. In view of a practical applicability in industrial applications two important issues are addressed. First, the problem of unbiased estimation of local model parameters in the presence of input and output noise is considered. For that purpose the concept of total least squares for parameter estimation is reviewed and a related partitioning algorithm based on statistical criteria is presented. Second, the steady-state accuracy of dynamic models is addressed. A concept of constrained TLS parameter optimisation is introduced which enforces the adherence of the model to selected steady-state operating points and thus significantly improves the model accuracy during steady-state phases. Results from a simulation model and from an industrial gas engine power plant demonstrate the capabilities of the proposed concepts.

Optimierte Neuro-Fuzzy-Modelle für Auslegungsprozesse und Simulation im Automotive-Bereich Improved Neuro-Fuzzy Models for Design Processes and Simulation in Automotive Applications

Abstract Dieser Artikel beschreibt einen neuen iterativen Partitionierungsalgorithmus für lokale Neuro-Fuzzy-Modelle. Der Algorithmus zielt darauf ab, datenbasierte Modelle aus dünnbesetzten Eingangsräumen, wie sie im Automotive-Bereich häufig auftreten, zu generieren. Der Gültigkeitsbereich jedes lokalen Neuro-Fuzzy-Modells wird dabei unter Anwendung statistischer Kriterien und Regularisierung an die vorhandenen Daten angepasst. Ausdehnung und Orientierung jedes lokalen Modells orientieren sich auch derart an den vorhandenen Trainingsdaten, dass die Bestimmung der Regressionsparameter ein gut konditioniertes Problem darstellt. Die Regularisierung des Modells kann durch den Benutzer mittels eines Parameters gut konditioniert werden. Um die Qualität und Sicherheit des erstellten Modells beurteilen zu können, wird zusätzlich eine Modellstatistik berechnet. Mehrere Beispiele aus der Praxiserprobung bzw. aus einem Pilotprojekt beschreiben die Effizienz des vorgestellten Algorithmus.

Neue Methoden für die dynamische Identifikation mit Neuro-Fuzzy Netzen (New Methods for Dynamic Identification with Neuro-Fuzzy Networks)

Dieser Artikel behandelt die Identifikation nichtlinearer dynamischer Prozesse mit lokalen Neuro-Fuzzy Netzen. Diese bieten den Vorteil, dass aufgrund ihrer Architektur die Möglichkeit besteht, Prozesswissen bei der Modellbildung mit einzubinden. Es werden Lösungsansätze für zwei wesentliche Problemstellungen der dynamischen Identifikation präsentiert: Einerseits wird die Problematik verrauschter Ein- und Ausgangsdaten behandelt, die bei der Parameterbestimmung mit herkömmlichen Regressionsmethoden zu biasbehafteten Resultaten führt. Als Lösungsweg wird die Total Least Squares Methode vorgestellt, und für die Anwendung in lokalen Neuro-Fuzzy Netzen adaptiert. Andererseits wird eine Methode vorgestellt, mit der die Einhaltung von Stationärpunkten erzwungen werden kann, was wesentlich zur Verbesserung der stationären Genauigkeit des Modells beiträgt. Resultate aus einem Praxisbeispiel illustrieren die Anwendbarkeit beider Konzepte. This article deals with the identification of nonlinear dynamic processes with local Neuro-Fuzzy networks. These networks have the advantage that their architecture offers the possibility to incorporate in-depth process know-how into the modeling procedure. The article presents solution approaches for two major issues in dynamic identification: First, the problem of noisy input- and output data is treated, which causes biased parameters when conventional regression techniques are applied. As a possible solution, the concept of Total Least Squares is presented and adapted for application in local Neuro-Fuzzy Networks. Second, a method for the enforcement of stationary gains is presented that significantly improves the model precision during steady-state phases. Results from a practical example illustrate the applicability of both concepts.

Data based modelling of driveability for internal combustion engines

ABSTRACT Non-convex hull concepts are used to describe shapes or images in data processing. The most widely used non-convex hull methods were introduced to visualize two- or three-dimensional data obtained via 3D scans or GPS signals. Data hulls have also been used as feasibility models in automotive control systems. For these higher-dimensional problems the computational complexity bears an important factor for data processing software. Commonly used methods, such as the convex hull, involve triangulations, which are numerically intractable in practice for dimensions above 10. Simple, spatial models can be computed and evaluated fast, using and floating point operations, respectively, but they perform poorly as feasibility classifiers if the data is close to infeasible regions. This article introduces a class of conic hulls that can be calculated for any input dimension with sub-quadratic, and evaluated with sub-linear, CPU complexity, respectively. These hulls are designed to include screening information of the measurement procedures embedded in the data generating processes.

Successful Integration of a Model Based Calibration Methodology for Non-standard Corrections and Protection Functions

Due to increasing requirements on the performance and emissions at non-standard conditions the amount of correction functions are increasing in order to utilize the maximum performance of the engine within its limitations. To be able to calibrate all those functions more capacity of an altitude test bench would be needed. On one hand those test benches are limited and on the other hand their operation is very expensive. The target of the AVL approach is to minimize the usage of a real test-benches, to maximize performance, keep emissions and increase dataset quality due to frontloading based on a semi-physical engine model. Those semi-physical model are able to simulate the thermodynamic behavior and the emissions under non-standard ambient conditions. Due the physical model parts also a model based system fault simulation can be performed in order to simulate engine failures and validate the component protection calibration. These results are established by the usage of these semi physical models on a hardware in the loop test bed in combination with Cameo automation and optimization (AVL Virtual Test Bed). The main advantage of using a Virtual Test Bed is, that the calibration environment for the application engineer is similar to the test-bench. In this way the calibration engineer does not need to have specific modeling knowledge. By using the Virtual Test Bed as environment the calibration can be done faster and with better engine performance as result. This can mainly be achieved due to the fact that during the calibration process for instance the real hardware limitation do not need to be taken in account. In combination with the automation with AVL Cameo and optimization with AVL Cameo this leads to a faster and better result. The successful integration of this methodology will be described.

Ein erweitertes, adaptives Kraftstoffzuteilungskonzept für den Ottomotor (Extended Nonlinear-Adaptive Engine Fuel Supply Control for Spark Ignition Engines)

Abstract Der Artikel beschreibt zwei neue Methoden zur verbesserten Kraftstoffzuteilung für Verbrennungskraftmaschinen: Die erste eliminiert stationäre Fehler, die z.B. durch ungenaue Kennfelder des volumetrischen Wirkungsgrades oder fehlerhafte Luftmessungen entstehen. Die zweite Methode reduziert instationäre Fehler. Bei Lastsprüngen bildet sich im Ansaugtrakt ein flüssiger Treibstoffwandfilm, was zu einer Abmagerung des Gemisches führt. Um diesen Effekt zu korrigieren, wird eine Zusatzmasse eingespritzt, welche durch ein diskretes Filter berechnet wird. Das Hauptproblem war, einen praxisgerechten und robusten Weg zu finden, um die Filterparameter ausgehend von den verfügbaren λ-Messungen zu adaptieren. Die Leistungsfähigkeit der präsentierten Methoden, gemeinsam mit den entsprechenden Adaptionsalgorithmen, wird anhand von Simulationen mit einem physikalischen Modell eines Ottomotors gezeigt. Die Methoden erreichen in allen Arbeitspunkten des Motors λ-Abweichungen von weniger als einem Prozent.