Data-Driven Soft Sensing in Industrial Processes

Data-driven soft sensing refers to the use of statistical and machine-learning techniques to infer critical process variables that are difficult, costly or time-consuming to measure directly. By exploiting correlations between readily available sensor data and target variables, soft sensors deliver real-time estimates of product quality, reaction conversion, energy consumption or environmental emissions. This paradigm has gained traction in petrochemical, pharmaceutical, food and energy industries as it reduces reliance on hardware instrumentation, lowers operational costs and enables faster process control. Typical approaches range from multivariate linear regression and partial least squares to black-box neural networks and grey-box hybrids that integrate first-principle knowledge. Recent efforts focus on addressing data non-stationarity, nonlinearity and noise through advanced optimisation, dynamic modelling and transfer-learning frameworks. When embedded within Industry 4.0 architectures, data-driven soft sensors support continuous monitoring, predictive maintenance and closed-loop control, promoting energy efficiency, consistent product quality and reduced environmental impact on a global scale.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent contributions include a mixed-integer optimisation framework that performs simultaneous input selection and outlier filtering in the training of soft sensors. By formulating variable selection and data cleansing as a single global problem, this approach delivers robust prediction performance and computational efficiency across multiple industrial datasets. In the field of building services, data-driven soft sensors for heat pump systems have been developed using artificial neural networks, multivariate polynomial regression and empirical models to reconstruct missing measurements, thus enabling improved control strategies, fault detection and grid interaction. Progress in dynamic modelling has been demonstrated via supervised bidirectional long short-term memory networks, which extract nonlinear temporal features and integrate past and future context to enhance real-time quality prediction under varying operating conditions.

Data-Driven Soft Sensing in Industrial Processes publication trend

The graph below shows the total number of articles in data-driven soft sensing in industrial processes across all publications each year (not limited to Nature Index journals).

Technical terms

Soft sensor: A virtual sensor that uses mathematical or machine-learning models to estimate hard-to-measure process variables from readily available data.

Data-driven modelling: An approach that relies on historical process data and statistical or machine-learning algorithms rather than mechanistic equations to predict system behaviour.

Mixed-integer optimisation: A mathematical programming technique in which some decision variables are constrained to be integer-valued, used here for simultaneous input selection and outlier handling.

Artificial neural network (ANN): A computational model inspired by biological neural networks, capable of capturing complex nonlinear relationships in data.

Bidirectional long short-term memory network (BiLSTM): A recurrent neural network architecture that processes sequence data in both forward and backward directions to capture time-dependent features.

References

1. A Mixed-Integer Formulation for the Simultaneous Input Selection and Outlier Filtering in Soft Sensor Training. Information Systems Frontiers (2024).
2. Data-driven soft sensors targeting heat pump systems. Energy Conversion and Management (2023).
3. A Supervised Bidirectional Long Short-Term Memory Network for Data-Driven Dynamic Soft Sensor Modeling. IEEE Transactions on Instrumentation and Measurement (2022).