Virtual Sensing for Hard to Measure States

56 concepts38 papers1 proposal

In many engineering and scientific domains, the most critical variable you need to track is often the one you cannot physically measure.

Consider a blast furnace processing iron ore, a chemical reactor synthesizing polymers, or a turbine spinning inside a hydroelectric dam. If you want to optimize these systems, you need real-time data on internal temperatures, chemical compositions, or localized mechanical stress. However, placing a physical sensor at the core of these processes is often impossible. Extreme heat, corrosive chemicals, high pressure, and tight spatial constraints create an observability gap where traditional hardware probes melt, fail, or cost too much to maintain.

To bypass these physical limitations, researchers use virtual sensing (often called soft sensing). Instead of deploying physical hardware to measure a target variable directly, virtual sensing uses software models to infer hidden internal states by mathematical cross-referencing with easy-to-measure peripheral data.

[ Physical Environment ] ──(Extreme Conditions)──> [ Hidden State (e.g., Catalyst Activity) ]
         │                                                          ▲
         └─(Accessible Proxies: Temp, Flow)──> [ Virtual Sensor ] ──┘

The Core Intuition: Physical Coupling

Virtual sensing works because physical variables do not exist in isolation. The laws of thermodynamics, fluid dynamics, and electromagnetism deeply couple the internal, unmeasurable states of a system to its external, measurable outputs. When an internal state changes, it sends a predictable "ripple" outward to the periphery of the system.

By deploying machine learning models as universal function approximators, we can invert this relationship—mapping the peripheral ripples back to the core cause.

This strategy allows researchers to turn standard, rugged operational data into high-fidelity diagnostic insights:

  • Indirect Mechanical Sensing: In underground mining, direct sensors like radar or infrared fail due to dust and noise. Pengfei Qi et al. demonstrated in Identification of Rock Properties of Rock Wall Cut by Roadheader Based on PSO-VMD-LSSVM that the mechanical properties of a rock wall can be inferred indirectly by monitoring the electrical current fluctuations of a mining machine's motor. Similarly, Van-Thuan Truong et al. showed in Prediction of Young’s Modulus for Reservoir Rocks using Machine Learning that standard drilling parameters collected in real-time can act as effective virtual sensors for rock stiffness.
  • Hostile Chemical Environments: In blast furnaces, direct physical measurement is obstructed by extreme internal conditions. As surveyed by Yueyang Luo et al. in Data-driven soft sensors in blast furnace ironmaking: a survey, soft sensors bridge this gap by leveraging data from accessible sensors to estimate variables like silicon content and gas utilization.
  • Environmental and Biological Proxies: In freshwater streams, tracking chemical pollutants like nitrate traditionally requires slow, expensive laboratory analysis. Benoît Liquet demonstrated that fusing high-frequency data from robust, low-cost physical sensors (such as temperature and conductivity) provides a reliable proxy for continuous nitrate estimation.

Structuring the Models with Physics and Geometry

Early soft sensors relied on simple linear regressions or unconstrained "black-box" neural networks. However, pure data-driven models often struggle to generalize when operating outside their training distribution, sometimes predicting physically impossible states.

Modern virtual sensing overcomes this by embedding physical constraints and structural geometry directly into the machine learning architecture.

Topology Mimicking Physics

Instead of forcing a generic neural network to learn a physical process from scratch, researchers design network topologies that mirror the underlying physical system. For instance, Samuli Bergman et al. introduced a "grid-structured" deep learning model in A Grid-Structured Model of Tubular Reactors. By structuring the network layers to match the spatial discretization of numerical Partial Differential Equation (PDE) solvers, the model naturally respects physical transport laws, allowing it to accurately track unmeasurable catalyst activity using sparse industrial data.

Relational and Graph Modeling

Industrial plants and environmental networks are inherently relational; sensors are distributed across interconnected pipelines, grids, or spatial zones. Graph Neural Networks (GNNs) allow virtual sensors to exploit these spatial correlations.

Hwa Hui Tew et al. developed a spatio-temporal hypergraph network in ST-HCSS: Deep Spatio-Temporal Hypergraph Convolutional Neural Network for Soft Sensing to model multi-node interactions across complex industrial processes, predicting chemical compositions in real-time.

This relational awareness is so powerful that it can even heal broken hardware networks. In Filling the G_ap_s: Multivariate Time Series Imputation by Graph Neural Networks, Andrea Cini et al. showed that high-fidelity relational modeling enables a system to synthesize data for entirely absent or failed sensors, effectively substituting physical hardware with software-driven estimates during connectivity losses.

[ Physical Sensor A ] ───┐
                         ▼
[ Physical Sensor B ] ───┼─> [ Spatio-Temporal Graph ] ──> [ Virtual Sensor C (Failed Hardware) ]
                         ▲
[ Physical Sensor D ] ───┘

Managing Interference and Multi-Task Dynamics

In large-scale industrial settings, a virtual sensor is rarely tasked with estimating just a single variable. Instead, it must monitor multiple target states simultaneously across different parts of a facility. This introduces a significant machine learning challenge: negative transfer, where optimization for one estimation task inadvertently degrades the accuracy of another.

Yuxin Huang addressed this architectural bottleneck in Modeling Task Relationships in Multivariate Soft Sensor With Balanced Mixture-of-Experts. By implementing a Balanced Mixture-of-Experts (BMoE) framework with task-specific gating and dynamic gradient balancing, the model isolates distinct physical phenomena within shared neural pathways. This ensures that predicting one chemical concentration does not interfere with the simultaneous estimation of another.

When these multi-variable models are combined with real-world physical boundaries, they form Hybrid Digital Twins. As outlined by Huangbin Liang et al. in Harnessing hybrid digital twinning for decision-support in smart infrastructures, integrating sparse physical sensor data with physics-grounded models allows operators to reconstruct a comprehensive view of structural health and operational responses in entirely unmonitored "dead zones."

Why Virtual Sensing Works

Virtual sensing transforms a hardware limitation into a software optimization problem. It succeeds because:

  1. Information Redundancy: Physical systems distribute information. A change in internal temperature changes pressure, fluid velocity, and casing strain simultaneously.
  2. Architectural Alignment: Modern models do not act as blind black boxes; they are structured as grids, graphs, or digital twins that match the spatial and physical reality of the machines they monitor.
  3. Cost and Resilience: Software does not melt. By moving the sensing burden from hostile physical environments into computational models, systems become cheaper to maintain and resilient against hardware failures.

Key Foundational Work

Nonlinear Robust PLS Modeling of Wastewater Effluent Quality Indices (2011) — contributed 2 concept elements to this direction.
THEORY, ALGORITHMS AND TECHNOLOGY IN THE DESIGN OF CONTROL SYSTEMS (2005) — contributed 1 concept elements to this direction.
Machine Learning and Sensor Fusion for Estimating Continuous Energy Expenditure (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_virtual_sensing_a475117f,
  title = {virtual sensing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/virtual-sensing/rdr_a475117f618949689dc0a4ea32082cb2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_virtual_sensing_a475117f_v1,
  title = {virtual sensing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/virtual-sensing/rdr_a475117f618949689dc0a4ea32082cb2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake