MOTOP - Optimization of Combustion Engines with the Help of Evolutionary Algorithms and Artificial Neural Networks

In co-operation with the BMW Group Munich - as one of the leading German car manufacturers - this project focuses on using different softcomputing techniques. Beyond others, emphasis is placed on artificial neural networks for modelling as well as on evolutionary algorithms for the accomplishment of complex optimization tasks. Therefore it is most important to clearly reduce the expenditure of the engine application at test beds and even to improve the calibration of the engine system. Better results are achieved on the one hand if the relevant engine characteristics like exhaust emissions and the fuel consumption are reduced with regard to legal regulations and a shortage of resources. On the other hand, customers' demands for performance and driving comfort have to be fulfilled.

Tasks

Complex engine functionalities are the answer of engine developers to the legal exhaust guidelines which are getting more and more strict and to the increasing demands for a reduction of the fuel consumption. These engine functions shall optimize the burn process and thus increase the efficiency of the engines. However, with the increasing number of adjustable engine parameters the complexity of the look-up table tuning, the so-called application of electronic control units becomes larger, too. Since using the conventional application procedures is no longer efficient new methodologies are developed and tested for application support in the co-operation. On the one hand it is possible here for the application engineer to gain an insight into the characteristic trait of the engine by means of modelling and visualization techniques On the other hand, a (semi)-automated calibration process of the electronic control unit is provided.

Methods

Genetic algorithms, a special form of evolutionary algorithms have successfully been applied for solving several combinatorial optimization problems in the engine application (e.g. optimization of the test bed schedule, statistical experimental design, look-up table design). Concerning other continuous optimization tasks such as e.g. the compression of look-up tables evolutionary strategies, which are a kind of continuous analogue to genetic algorithms, are used here because they outmatch classical optimization algorithms, e.g. quasi-Newton methods or Sequential Quadratic Programming.

In an earlier stage of the project an off-line process for optimizing look-up tables was developed which permits a separation between the test bed measurements and the optimization of the look-up tables. The objective functions (e.g. fuel consumption and exhaust emission) are measured at the test bed at a list of certain combinations of engine parameters (engine speed, air mass flow, inlet and exhaust valve spreads, ignition angle). The measuring results are used to compute computer models (e.g. neural networks or other regression types) for these objective functions and hence for the engine itself. In the optimization phase which is completely decoupled from the test bed those combinations of engine parameters as inputs for these models are determined where the objective function is limited to a minimum.

At present the project deals with developing an on-line optimization strategy which allows to determine the engine parameters that lead to the desired engine behavior in each driving condition directly at the test bed with as few measuring points as possible.

Statistical Design of Experiments:

Genetic algorithms are used for the computation of D-optimal experimental designs. Classical heuristics that work locally (e.g. the DETMAX or the k-exchange algorithm), sucessively exchange bad candidates for better ones. These heuristics are used in so-called hybrid genetic algorithms as local search operators. By defining a suitable Crossover operator on experimental designs, good subregions of different designs (individuals) are combined to a best design (individual).

Optimization of the Test Bed Schedule as TSP Variant:

The adjustment of some engine parameters at the test beds causes an unintendent and long relaxation time of the total engine system. With the help of an optimized test bed measurement schedule significantly shorter measuring times and also higher robustness degrees can be reached. This problem is a variant of the well-known Traveling Salesman Problem in the parameter space of the engine. Genetic algorithms are very much suited to solve this problem.

Look-up Table Design - Smoothing:

After the off-line or during the on-line optimization of the engine parameters, the following happens quite often: there are several optimum candidates available for the definition of the look-up table at many operating points. They are similar concerning the objective function but may differ significantly in their values. Thus, a selection process is necessary in order to receive a well-defined look-up table which selects the correct candidate at each operating point. A selection criterion is the smoothness of the resulting look-up table in order to achieve fast transitions, particularly for mechanically adjusted engine parameters.

Look-up Table Design - Compression

Due to the limited memory capacity of the controllers within electronic control units control parameters of existing functions may have to be modified in order to share memory with new parameters. By means of evolutionary strategies the storage requirement of existing look-up tables is reduced. Therefore the size of the grid is reduced and the function values at the new grid are adjusted.

Look-up Table Design - Inversion

Some engine functions need information from an inverse look-up table access in order to provide a pre-setting of the corresponding parameters. Instead of accomplishing the access numerically within the original look-up table the inverted look-up table is often also stored within the controller. The techniques of the look-up table compression can also be used for this memory management.

On-line Optimization:

The separation between test bed measurements and the modelling as well as the optimization as it exists in the off-line optimization is lost during the on-line optimization process. Here, both the experimental design and modelling and the optimization are performed directly at the test bed.

Contact

Alexander Sung, Sand 1, Raum 318, Tel. (07071) 29-78979, sung at informatik.uni-tuebingen.de

Prof. Dr. Andreas Zell, Sand 1, Raum 310, Tel. (07071) 29-76455, zell at informatik.uni-tuebingen.de