Introduction to soft sensors
Soft sensors combine mathematical models with process data, allowing real-time inference of hard-to-measure variables such as laboratory measurements. Imagine a precise estimate of your pH levels, whenever, wherever, without having to take a single sample.
This blog post explores advancements in soft sensor research and application, highlighting how developments in artificial intelligence and machine learning are making these tools more reliable and versatile. This is our third blog post on soft sensors, so if you feel that you may need a longer reminder than the one above on what soft sensors are or how they are developed, check out our previous two blog posts before diving into this one.
Advances in soft sensors
Soft sensors have been around for a longer period of time, and while their promised value potential has always been great, they never seemed to fly off the shelves. A well-crafted traditional soft sensor by an expert in the field can produce an excellent end-product but will require an expert with a substantial amount of time with limited guarantees of success. Recently the tools have experienced a sudden increase in popularity. Advances in soft sensor research and the surge of artificial intelligence have made soft sensors significantly more accurate and adaptable, ready to address challenges faced across the manufacturing industry:
- Better handling of complex and dynamic processes:
Traditional models struggle with processes that change over time or have highly complex relationships between variables. While even artificial intelligence requires some level of consistency—and capturing new process phenomena is still difficult—recent modelling methods now incorporate ways to dynamically track changes and identify patterns in data that were previously hard to capture (Yang et al., 2023). - Customizing to diverse operating conditions:
Industrial systems often vary significantly across facilities and regions. Adapting soft sensors to these differences in operating conditions was previously limited by static models, insufficient data, and a lack of advanced algorithms for dynamic, real-time learning. Advances in machine learning, data collection, and computational power now enable soft sensors to better generalize across environments, update in real time, and handle complex, nonlinear industrial processes effectively (Loo et al., 2023). - Adapting to missing or incomplete data:
In industrial environments, sensor malfunctions or data loss are common. Early soft sensors struggled to compensate for such disruptions, leading to inaccuracies and gaps in process monitoring. New advancements enable soft sensors to handle missing data by using historical trends, statistical inference, and machine learning methods (Xie et al., 2020).
While the observations in this blog post come from research papers, these developments are far from confined to labs. The techniques described are already part of SimAnalytics’ FH Soft sensors used in real-world industrial settings. By addressing challenges like managing complex processes, adapting to changing conditions, and handling incomplete data, FH Soft sensors provide reliable predictions for hard-to-measure variables, helping industry professionals make faster decisions to improve process efficiency, quality, and consistency.
In our next blog post on soft sensors, we explore how these advancements translate into practical use cases across various industries. Dive into the blog post here.
References
Loo, J. Y., Ding, Z. Y., Nurzaman, S. G., Ting, C.-M., & Tan, C. P. (2023). Unsupervised cross-domain soft sensor modeling via deep physics-inspired particle flow Bayes. Neural Networks and Learning Systems Journal.
Xie, R., Jan, N. M., Hao, K., Chen, L., & Huang, B. (2020). Supervised variational autoencoders for soft sensor modeling with missing data. IEEE Transactions on Industrial Informatics, 16(4), 2820–2830. https://doi.org/10.1109/TII.2019.2951622
Yang, Z., Jia, R., Wang, P., Yao, L., & Shen, B. (2023). Supervised attention-based bidirectional long short-term memory network for nonlinear dynamic soft sensor application. ACS Omega, 8(4), 4196–4208. https://doi.org/10.1021/acsomega.2c07400
