ETAS ASCMO-DYNAMIC enables users to create data-based models that model the dynamic/transient behavior of complex systems. ASCMO-DYNAMIC provides a wealth of functions and options for visualizing and analyzing the system behavior. To proceed optimizations an export to ASCMO-MOCA is possible. It can also be used for creating experimental designs based on the DoE methodology (design of experiments).
AI methods from the field of machine learning are used that allow the user to accurately model complex relationships without requiring precise knowledge of the algorithms in the background. For this reason, the software is used by a wide range of users. They include less experienced users, who appreciate parameter-free, automated model creation, as well as experts, who are given all the freedom they need thanks to the extensive configuration options available.
A typical application of ASCMO-DYNAMIC is the modeling of transient processes in internal combustion engines. Particularly in the context of Real Driving Emissions (RDE), a purely stationary consideration of the emissions and fuel consumption is often not sufficient. Dynamic effects (e.g. peaks) can have a significant influence on the overall outcome. When modeling relevant variables like fuel consumption and pollutant emissions, not only are the currently set values (e.g. engine speed, load, and engine control variables) taken into consideration, but their historical values and the rate of change are taken into account too. This enables a detailed, time-resolved analysis of the influence of dynamic driving maneuvers on the relevant output variables.
The methodology used in ASCMO-DYNAMIC is not tied to the internal combustion engine, which means the tool is also used in areas like electric mobility (e.g. battery modeling).
- Easy to use with no specialized knowledge required
- Powerful AI methods from the field of machine learning
- Interactive graphical representation of data and results
- Ability to share models and data using standardized formats
- MATLAB® and COM interfaces for remote control and the integration of customer-specific functions
Features and functions
Since data-based modeling is used in ASCMO-DYNAMIC, this data must typically be obtained through measurements on a real system. An essential element for this is the design of experiments (DoE), which enables a maximum model accuracy to be achieved with a minimum of effort spent on measuring.
The DoE experimental design module (ExpeDesDynamic) makes it easy to plan the underlying measurements. Using design of experiments methods, a time curve of the adjustment variables is proposed in which different final values with different gradients are approached methodically and in a space-filling way.
The variation range as well as the maximum gradients of the parameters to be adjusted can be limited graphically or numerically in up to four dimensions in each case. Using a formula editor, this is possible in any number of dimensions. In addition, stationary points and excerpts from real driving cycles can be specifically added and a consolidation of the points in certain regions of the experimental space can be set. Special aspects and peculiarities concerning dynamic measurements performed on engine test benches are taken into account in a highly practical way. Graphical representations of the curves (trajectories) in all dimensions facilitate straightforward validation and evaluation of the DoE plan.
ASCMO-DYNAMIC uses time-based scope views to display the system behavior based on the model. The upper section shows the input variables and the lower section shows the output variables over time. Besides this main view in ASCMO-DYNAMIC, there are also various other options available for visualizing the dependencies and influences. This includes, for instance, the scatter plot, which can present the data as a point cloud in any number of different combinations.
Extensive analysis options are available for model quality and prediction quality. For example, a variety of statistical information is displayed, such as mean and extreme values and the root-mean-square model error (RMSE). In addition, the model prediction can be compared with a real measurement in order to compare the curves over time.
An important criterion for successful modeling is a suitable sampling rate during measurement. Since there is typically a tendency to sample at a rate that is too high, ASCMO-DYNAMIC includes the option to reduce the amount of data through downsampling. This increases the model quality and at the same time shortens the time needed for model training
The main function of modeling in ASCMO-DYNAMIC is to enable accurate, straightforward, and robust prediction of the dynamic/transient system behavior. The tool makes it easier for less experienced users to use the advanced modeling methods required for this, but at the same time it gives experts the freedom to adjust the model parameters with a high degree of flexibility. In ASCMO-DYNAMIC, the modeling is based on Gaussian process models. This consists of a statistical learning method using measurement data based on Gaussian processes that has already proven its worth many times. With this, the specific mathematical function that best represents the real system behavior is determined automatically.
Further advantages of Gaussian process models:
- Measurement noise is taken into account, so overfitting can be avoided
- Robustness against measurement outliers
- Locally resolved model variance gives the user a measure of the reliability of the model prediction
When modeling dynamic system behavior, it is necessary to consider not only the current point in time but also the past time steps. Various modeling methods are available for this purpose. To use the Gaussian process models, a structure is employed in which time steps from the past are fed into the model as additional inputs. The past time steps required for this are determined automatically by ASCMO-DYNAMIC.
For special tasks, additional model types from the class of recurrent neural networks (RNN) are offered alongside the Gaussian process models. The RNN cell type LSTM (long short-term memory) is, for example, well suited for problems with influences that extend far back into the past.
Experienced users can activate expert mode, which gives them access to all the model options and parameters and thus allows them to make specific adjustments when needed that have an influence on the model training. The model quality is indicated by easy-to-understand charts and key figures.
All models can also be exported in different formats and used freely outside ASCMO as a so-called plant model. Common examples of possible formats are Simulink, C‑Code, and FMU/FMI.
A special format is available for creating an export for use on the Bosch AMU (Advanced Modeling Unit hardware accelerator). The AMU is a special module on the Bosch MDG1 ECU and is used for evaluating the ASCMO model stored on it in volume-production vehicles.
In ASCMO-DYNAMIC, the models can be used as a virtual measuring instrument. An example of this is the prediction of emissions over a complete driving cycle based on time series for the input variables. For all the desired parameter combinations of input variables, the model will then calculate the corresponding output variables. Just like real measurements (e.g. from the engine test bench), this artificial measurement data can be used, analyzed graphically, and exported as a CSV file for further processing in Excel.
ASCMO-DYNAMIC is open and flexible. The tool supports all relevant data formats that are used, for example, on the test bench and in calibration. ASCMO-DYNAMIC can export the created models in various formats for use in other environments. Using the MATLAB® interface, the user can easily integrate customer-specific functions and models, automate operations with scripting, or incorporate a test bench automation.