Data Driven Sensors Use Your Process Data to Improve Process Prediction and Monitoring

Ivana Lukec
Data Driven Sensors

The necessity to operate industrial units at minimum operating costs and still meet all requirements of product quality and tight environmental laws requires continuous process monitoring, optimization, and efficient quality control. Nowadays soft sensors are used as the supplement to online instrument measurements for process monitoring and control improvement. 

Given that the industrial applications must be simple and easy to use but still reliable, the objective of this article is to illustrate the application of the soft sensors (data-driven sensors) on examples of commercial plants.

Soft sensors have proven their predictability of quality properties or content as a supplement to hardware process analyzers because their measurement may often become impossible due to instrument failure, maintenance or repair. Moreover, soft sensors are more frequently capable of predicting the properties then are the hardware analyzers, which is especially important in real time optimization.

How to use soft sensors

A mathematical model is the core of soft sensor application. Those mathematical models of processes designed to estimate relevant process variables can help to reduce the need for measuring devices, improve system reliability and develop tight control policies. Plant models devoted to the estimation of plant variables are known either as inferential models, virtual sensors or soft sensors. Soft Sensors offer a number of attractive properties:



  • they represent a low-cost alternative to expensive hardware devices, allowing the realization of more comprehensive monitoring networks;
  • they can work in parallel with hardware sensors, giving useful information for fault detection tasks, thus allowing the realization of more reliable processes;
  • they can easily be implemented on existing hardware (e.g. microcontrollers) and retuned when system parameters change;
  • they allow real-time estimation of data, overcoming the time delays introduced by slow hardware sensors (e.g. gas chromatographs), thus improving the performance of the control strategies.

They are used as the bridge between product quality and control, helping achieving product quality is controlled before it is disturbed.

What are components for development of soft sensor

The important components of a good soft sensor are process data, process knowledge, and adequate modeling tool.

The first step is to understand the process and have a clear vision of the goals that the soft sensor application should achieve: Is it a certain property that needs to be estimated? Is there existent hardware analyzer or a lab analysis? Will the soft sensor be applied for process monitoring only or also as a feedback control? What measured variables such as temperatures, pressures or flows could be used? What is a suggested time span?
When all the answers related to understanding of the process are cleared, adequate process data should be selected and processed. Adequacy of the data is dependent both on their quantity and their quality. Because, as good the data are – that good the application will be. Operating area for which the soft sensors should be applied also depends on the selected and processed data and one should be careful about the use of the sensor for the operating area outside of the designed operating area. The data taken for the development must include the whole operating area of the soft sensor. 

After the definition of the goals, the modeling tool is needed to analyze the data, test them and identify the representable model. The proper tool is selected based on answers to these questions: Do we know the mathematical technique to be used prior to developing the model?

Mostly, the answer to this question is – no. Because, even if the process is well known in advance, process data should be tested for linearity. If linearity can be assumed from the data then mathematical techniques such as linear regression, principal component analysis can be applied. If it is not the case, then artificial neural networks, partial least squares or non-linear regression techniques should be applied.

Some of the commercial products allow that data is first processed so that their behavior could be analyzed, whether they show linear or nonlinear dependency while others are specialized for a certain mathematical technique: e.g. for application of artificial neural networks.
Process behavior most often is non-linear. However, whether the developed model should also be non-linear depends on the operating range in which the model will be used. If the process is controlled and the operating range is small, a linear process model may be an adequate approximation of reality.
In cases where it is possible to apply linear regression – the results of a soft sensor may be applied with a simple algorithm in MS Excel. Examples of that will be shown in SimulateLive.com articles.
Dynamic behavior of the data also needs to be known in order to define time lag or time constant of the soft sensor property compared to data taken from the plant.


Installation of soft sensor on control system

Once established, a soft sensor has to be applied on the unit control system. Dependent on the control system, its user interface is defined. Some control systems already have the possibility to use such application integrated with the system while for other – it should be defined. Mostly, the soft sensor developer can provide the appropriate algorithm and instructions to the DCS engineer who is able to define it at the control system.
A user interface is part of the control system to which a mathematical algorithm is defined. The soft sensor algorithm can be defined directly on the control system, independently of the mathematical tool used for developing the model. But, it can also be connected to another application. 

The ability to use this kind of applications is immense. From defining individual soft sensors to improved monitoring and quality control, to developing the whole systems of them to control and predict the behavior of the whole process sections. Advanced process control applications would be significantly less efficient if soft sensors wouldn't be applied. Due to the integration of knowledge that is stored with their application, due to their low cost and predictability – they can and should be used all around the process industry.