Soft sensor

from Wikipedia, the free encyclopedia

A soft sensor (made up of the words “software” and “ sensor ”), also known as a virtual sensor or sensor fusion , is not a real sensor, but rather a dependency simulation of representative measured variables on a target variable. Thus, the target variable is not measured directly, but is calculated on the basis of correlated measured variables and a model of the correlation.

The determination of the dependency can take place in different ways, for example with the help of artificial neural networks or multivariate methods .

Soft sensors are mainly used wherever the ambient conditions prevent real sensors or their use would be too expensive.

definition

The soft sensor uses hardware sensors to determine correlating process variables (x, y) and transmits them as input variables (m) to the stored model, which calculates the target variable in the form of an evaluation algorithm. The process to be monitored is influenced by controllable (u) and non-controllable (d) influencing variables.
Model validation with measured target value

Soft sensors map the dependence of correlating measured variables on a target variable in order to be able to calculate them. This means that the target variable is not determined with real measurement sensors in the classic sense, but can be determined based on the relationships to other measured variables. As in a simulation, the soft sensor reflects the state of the environment in order to calculate the corresponding actual value of the target variable for each state of the hardware measurement sensors. The target variable does not have to be a physical variable, but can also be a characteristic value, a tendency or an abstract variable . A simple example of this are resistance thermometers . These do not measure the temperature directly, but a temperature-dependent change in the electrical resistance, which is then converted into the temperature using a simple correlation. This corresponds to the concept of the soft sensor. Usually, however, only images with more than two participating input variables on one output variable are referred to as soft sensors.

Creating soft sensors

The function of a soft sensor is defined by a model , which reflects the relationships between target and measured variables. The main effort is therefore to generate the model. This can be done with the help of various methods.

If all relationships are known and can be expressed using a chemical or physical formula, one speaks of rigorous modeling. The advantage here is that all states are already known, which is why we also speak of white box models. The disadvantage, however, is that in most technical applications the process to be modeled is not completely known, since a large number of complex influences overlap that can only be described approximately or with the help of simplifications and assumptions. The multivariate methods include various analysis and regression methods . Here z. B. all correlated measured variables are combined into main components and these are transferred in a new value range with reduced dimensions. Thus, one part of the process, analogous to the rigorous modeling, is already known, while the other has to be determined, hence also called the gray box . The disadvantage of the method is that many processes can only be described with a large number of main components, which means that there is hardly any simplification.

Another approach are the artificial neural networks . With this black box method, the mathematical relationships are unknown. Since it is a purely data-based self-learning modeling, relationships that cannot be analytically resolved can also be described with it, provided that they are represented in the database. The danger of artificial neural networks is what is known as overfitting , in which the network learns its training data sets by heart without mapping the actual process.

Advantages and disadvantages

For the training phase in the modeling phase, most methods require a large database of the measured variables as well as target variables, which requires complex data acquisition in advance. Another problem with soft sensors is their individuality. This means that they are not very robust against changes in the environmental conditions. If one of the measured variables is outside the model range due to process changes, a great inaccuracy of the model prediction must be expected.

Soft sensors offer advantages primarily due to their real-time adaptation options. Among other things, this also offers the option of developing process monitoring into a closed control loop and thus recognizing process deviations at an early stage and being able to take countermeasures in good time. Soft sensors can also be used to monitor hardware sensors. Since the measured actual value of the hardware sensor can be compared with the calculated target value of the soft sensor at any time. It is thus possible to detect incorrect measurement results due to hardware sensor defects and to compensate for their failure temporarily if necessary. Furthermore, soft sensors offer the possibility of quantifying the influence of the correlating measured variables on the target variable and thus determining their dependencies in order to gain a greater understanding of the process. It would be B. possible in a process to find the optimal settings of the process parameters of the correlating measured variables for the target variable.

application areas

The fields of application of soft sensors are very diverse. The greatest spread is found in the chemical industry. They are also used in the system control of combustion processes in power plants. In more recent research work, the use in plastics processing is also being promoted, where this has already been successfully implemented. As part of the development of soft sensors, precise process models are created that also allow the use of soft sensors for process analysis and optimization. This enables the process parameters to be adjusted to improve energy efficiency, cost efficiency and quality. Applications can be found e.g. B. already in the plastics sector.

Individual evidence

  1. ↑ L. Fortuna, S. Graziani, A. Rizzo, MG Xibilia: Soft Sensor for Monitoring and Control of Industrial Processes. Springer-Verlag, London, 2006, ISBN 1-84628-479-1 .
  2. ↑ SKZ - The plastics center according to Luttmann et al .: Soft sensors in bioprocessing: A status report and recommendations. In: Biotechnology Journal. 7, 2012, pp. 1040-1048.
  3. ↑ SKZ - The plastics center according to Yiagopoulos, among others: Development of a Softsensor for On-line MFI Monitoring in Reactive Polypropylene Extrusion. In: ECHEMA monographs. 138, 305 (2004).
  4. ^ T. Hochrein, I. Alig: Process measurement technology in plastics processing. Vogel Business Media, Würzburg 2011, ISBN 978-3-8343-3117-5 .
  5. ↑ C. Kugler, T. Hochrein, M. Bastian, T. Froese: Hidden treasures in data graves. In: QZ. Year 59, Carl Hanser Verlag, Munich 2014.
  6. ↑ C. Kugler, K. Dietl, T. Hochrein, P. Heidemeyer, M. Bastian: Robust soft sensor based on an artificial neural network for real-time determination of the melt viscosity of polymers. In: PPS-29. Nuremberg 2013.
  7. ↑ C. Kugler, T. Froese, T. Hochrein, M. Bastian: Real tasks for virtual sensors. In: plastics. Carl Hanser Verlag, Munich, issue 2/2012.