Motivation

In the process industry, models of a large variety are used for diverse applications. For most purposes, a trade-off between computing time and accuracy is made. For example, the use of physical process models in operating assistant systems leaves limited scope for accurate calculations and will fail frequently if complex mechanisms on different time scales need to be considered. At this point, the large amounts of data available in the process industry and the recent progress in the area of digital technologies offer great potential in the design of innovative data-driven models. These data-driven models offer the ability to find unknown correlations in process signals and map the behavior to some dependent output signal, provided that enough data is available. Additionally, they give the option of a quick adjustment of the model to changing operating conditions. As tempting as this may sound there are always the drawbacks of limited insight into a data-driven model as well as its poor extrapolation behavior. This project, therefore, aims to develop and investigate different concepts of combination for data-driven and physical models – so-called hybrid models – to overcome the limitations of the individual model types.

System description

The process under consideration is a vertical roller mill (VRM) for comminution of cement and coal particles. The main mechanisms of the mill are the comminution of the particles on the grinding table, the particle classification during the transport to the sifter as well as the sifting process itself. The available process data comprise gas air flow, sifter torque speed, acoustic signals, etc. The desired outputs of the model are the particle size distribution and the estimation for the plant status, e.g. an indicator of the clogging of the sifter.

Methodology

The model development process consists of three different stages:

1. Development of a physical model for the vertical roller mill.

   This model will be implemented using the software DYSSOL, a flowsheet platform that is developed within the DFG Priority Program SPP 1679 “Dynamic simulation of interconnected solids processes”. DYSSOL is especially suited for simulating dynamic models of different solids processes as it gives the possibility to describe the multidimensional nature of solids properties, e.g. size, shape and porosity distributions, in multiphase systems.

2. Development of data-driven model and signal processing.

   Process signals usually contain a lot of disturbances as well as outliers and other unusable or plainly wrong information. Thus, without a proper assessment, the process data should not be used for the training of data-driven models. The signal processing consists of clustering techniques and statistical methods like principal component analysis. These methods allow for example the detection of outliers and the increase of the information content in the data, respectively. The data-driven model will be implemented as an artificial neural network (ANN) . Such an ANN consists of interconnected neurons that change incoming signals according to some predefined functionality. The connections of the neurons possess individual weights which are iteratively changed during the training process and ultimately store the ANN’s “knowledge”.

3. Combination of the two model types.

   There are a lot of concepts for coupling a physical model with a data-driven one. Some intend to directly combine the results of both model types in different manners, whereas others use the data-driven models as a parameter-estimator for the physical models. At first the most intuitive concept will be used: a parallel coupling of the models that results in a superposition of the single outputs. In this formation the data-driven model eventually describes the residuals between the measured output and the results of the physical model, i.e. it tries to compensate the inherent structural deviations of the physical model from reality.