Introduction of Soft Sensors
Industry 4.0 is one of the important topics for manufacturing which its main goal is to implement smart technologies such as IoT, smart sensors, etc. in the industry to make manufacturing faster and more efficient. In fact, it is a digital transformation to optimize all aspects of the manufacturing process.
For this purpose, many sensors are used to monitor processes. They take measurements, collect data, and send data to a centralized platform, and then this data is used to analyze the process. These sensors can be used to control and measure temperature, pressure, etc. However, implementing many sensors is also costly. Also, it is essential to make sure that all the sensors work correctly. In case that a sensor fails, it takes time for replacement, and it can affect the process of manufacturing. Of course, it increases downtime and maintenance costs.
One of the tools, which can support manufacturers to solve this issue in their process, is soft sensors. This article is a brief introduction of soft sensors and their applications in the industry, which have a high potential for implementation in manufacturing to increase the quality of the controlling the process and also reduce the maintenance cost as well as downtime.
What is a soft sensor?
For several years, numerous sensors are used in industrial processing plants to deliver data for process monitoring and controlling. Therefore, a huge amount of data is collected, which gives essential information about the process. Almost two decades ago, researchers tried to use this data to build predictive mathematical models that can deliver similar information as the hardware sensors, and they called them soft sensors [2].
A soft sensor, virtual sensor, or inferential model is a mathematical model of processes designed based on experimental data and/or phenomenological knowledge about the process on system identification procedures, which can significantly reduce the need for measurement equipment and develop streamlined control strategies [1].
The name of the soft sensor has two terms; soft, which refers to software-based, and sensor, which refers to providing information similar to a hardware sensor [1].
Generally, there are three models for soft sensors; model-driven, data-driven, and grey-box model. The model-driven (or white-box) soft sensors are based on full phenomenological knowledge about the process background. This kind of soft sensor is implemented for the planning and development of the process plants. Moreover, this model often describes a simplified theoretical background of the process, and therefore, it can not present the real situation, which can be influenced by many unpredictable factors [2].
The data-driven (or black-box) soft sensors are based on historical data. This historical data is measured within the processing plants, hence this kind of soft sensors can describe the real process condition better than the model-driven soft sensors. This feature of data-driven soft sensors increases the popularity of this model more in the industry [2].
The last model of the soft sensors is the grey-box model. It is a combination of model-driven and data-driven soft sensors. For instance, sometimes the full phenomenological knowledge of the process background is not available, or it is impossible to model. Hence, it is tried to predict the output using available parts, and then compare this predicted output with actual output. This comparison gives us a difference. A data-driven model can be developed to model the difference between the predicted and actual outputs. Then the outputs of both these models are combined to generate the final output. One of the advantages of this model is that it partially depends on data, therefore, we have higher quality estimates of results and variables in the system, and of course, it decreases the cost of the process [3].
The general developing methodology of the soft sensor consists of several processes such as data inspection, historical data selection, data preprocessing, a mathematical model selection, training and validation, and soft sensor maintenance. For selecting a mathematical model, there are many possible models for soft sensors, for instance, model based on Principal Component Analysis (PCA), Recurrent Neural Network (RNN), Support Vector Machines (SVM), Adaptive Neuro-Fuzzy Inference System (ANFIS), or even a combination of these methods. The following figure shows an overview of the soft sensors in the process for industry [2].
Soft Sensor Application and Use Case
Generally, we can categorize the application of the soft sensor into four general modes:
Process monitoring and fault detection: One of the applications of soft sensors is to monitor the process and to detect any fault. Both unsupervised and supervised learning methods can be used.
Online prediction: Soft sensors can be used to predict a value that cannot be measured online automatically. This is the most common application of soft sensors in the industry.
Sensor fault detection: For any reason, sometimes the hardware sensors can fail. In this scenario, a soft sensor can detect this fault and work as a backup solution for the measuring device until a new sensor will be replaced, or even work instead of the missing hardware sensor [2].
What-if analysis: Another application of soft sensors is to use them as a simulation of the system dynamics. Hence, we can obtain a deeper understanding of system behavior, and design a better control system [1].
Soft sensors have been implemented in different industries. For example, different models based on the Fuzzy Inference System (FIS), Partial least Squares (PLS), and Artificial Neural Network (ANN) have been developed for the iron and steelmaking to control the temperature or estimate the exit concentration of hydrochloric acid of the pickling process.
Oil and gas processing is another field that soft sensors are used so often. Different models have been implemented for different goals. For instance, to control and estimate the rejected flow rate of a hydro cyclone, or to predict CO2 emission in the plant-wide facility of oil and gas rig [4].
In addition to the two above examples, there are many other use cases for soft sensors such as implementing them for food processing, chemical plants, water treatment, material processing and energy materials, soft robotics, programmable logic controller (PLC), etc.
Conclusion and outlook
As I mentioned, there are different possibilities to use soft sensors in the industry. We can use them as a backup solution for the hardware sensor, monitoring and fault detection in the process, simulation of the hardware detector, and online prediction. Moreover, there are many use cases for soft sensors in the industry. Based on these applications for soft sensors, these are practical tools for industry 4.0 and bring many benefits for companies such as reduces maintenance costs and downtime as well as increase better quality control.
Despite the advantages of soft sensors, there are still few challenges for them. The main challenge is the maintenance of the model developed for soft sensors. However, there are some efforts to overcome this issue by implementing an adaptive model for the soft sensors.
If you have any questions or want to know how you can implement soft sensors in your business or company, you are most welcome to contact me.
[1] Luigi Fortuna, Salvatore Graziani, Alessandro Rizzo and Maria G. Xibilia “Soft Sensors for Monitoring and Control of Industrial Processes”, London, UK, Springer, 2007.
[2] Petr Kadlec, Bogdan Gabrys, Sibylle Strandt, Data-driven Soft Sensors in the process industry, Computers & Chemical Engineering,
Volume 33, Issue 4, 2009, Pages 795–814, ISSN 0098–1354,
https://doi.org/10.1016/j.compchemeng.2008.12.012.
[3] Sandip Kumar Lahiri, Multivariable Predictive Control
Applications in Industry, John Wiley & Sons Ltd, USA, 2017.
[4] Iftikhar Ahmad, Ahsan Ayub, Manabu Kano, Izzat Iqbal Cheema, Gray-box Soft Sensors in Process Industry: Current Practice, and Future Prospects in Era of Big Data, Processes, Volume 8, 2020, Number 2, https://doi.org/10.3390/pr8020243
