Data-based processes for the creation and formulation of production data for complex industrial plants.
In complex industrial plants (chemical industry, metallurgy, pulp and paper) a large amount of data is measured and stored during operation periods. In this project this existing production data is used to improve the knowledge of the physics that determine the behaviour of these plants. This is especially valuable for plants where only partial descriptions of the behaviour and only incomplete models exist. The identification and formal definition of unknown influencing variables and their mutual interactions shows technological optimization like reduction of scrap or rework, optimization of resources and energy, production of new qualities, etc.
The explicit goal of the project is the validation of a method for automatic data based modelling, fault identification and -isolation, which consists of the following steps:
- Automatic data pre-processing: The input of this module is raw process or plant data in a format which is typically used for data transmission or storage in industrial production processes (e.g. for quality control). It is important to find a reliable, robust and flexible method which can deal with a large number and variability of data WITHOUT user interaction.
- Variable selection: All measurement channels (of the pre-processed raw data) are checked with linear and non-linear methods for mutual interaction. The results are variable clusters representing groups of measured channels which have significant relationships between each other.
- Databased modelling: For all interactions between variables (which are identified during variable selection) the type of relationship is examined and formally defined. Again, linear and non-linear approaches are used (i.e. linear and polynomial models). Additional modelling methods are tested to enlarge the pool of possible models (i.e. neuronal networks, genetic algorithms, etc.)
- Fault identification is one possible continuation of the previous steps. It is based on the difference between the actually measured and the calculated value of the models for each channel.
The expected results of the method which has to be developed during this project are:
- For each measured channel a statement, which other channels are influencing this channel (result of variable selection)
- Explicitly formulated models (ideally with graphical representation) describing the mutual interaction between selected channels (result of modelling)
- Information, which (groups of) measured channels are having abnormal values with regard to the values of other channels (based on a reference data sample) (result of fault identification)
If the project is successful, the researched method will be the basis for a whole range of further developments leading to final products. These applications differ due to different applications, users and environments:
- General analysis of process data and improvement of process models which are used to control production processes
- Plant-specific analysis of process and plant data to optimize the operations of existing plants (Such plant-specific models can not be realized economically with existing methods)
- Support and acceleration of set-up or ramp-up periods of new plants or after a shutdown
- Plant diagnosis for early detection and prognosis of faults in plants for preventive maintenance
- Monitoring of general plant condition including information, which product qualitities can be produced with the current plant and equipment condition.
DI Dr. Bernhard Huch
VOEST-ALPINE Industrieanlagenbau GmbH & Co KG
Project or cooperation partner
- Linz Center of Mechatronics GmbH
- JKU Linz - Institute for Design and Control of Mechatronical Systems