Virtual sensors that know your process

Polymer melt index and density are only available from infrequent laboratory analysis (hours of delay). Quality control requires continuous real-time feedback across reactor temperature profiles and feed rates.

First-principles kinetic model inside NMPC closes the loop on delayed lab measurements, simultaneously controlling multiple reactors. Linear APC is insufficient due to strong nonlinear coupling between quality attributes and process conditions.

Basis weight (g/m²) cannot be measured in real time in the wire section of a tissue machine — only downstream, too late to correct production deviations.

Hybrid model: first-principles model (FPM) combined with 1D-CNN in parallel; a GRU network dynamically weights their outputs based on the current operating regime.

Permanent downhole gauges (PDGs) are expensive and unreliable in harsh well conditions. Bottom-hole pressure (BHP) is critical for production optimisation and flow assurance.

LSTM soft sensor trained on wellhead / topside measurements. Transfer Learning adapts the model across different well environments and operating conditions with minimal additional data.

Real-time NOx prediction is required for combustion optimisation and emissions compliance. Plant conditions shift continuously with load, fuel quality, and equipment ageing, causing static models to degrade quickly.

Just-In-Time Learning Random Forest (JIT-RF): adapts locally to current operating conditions at each prediction step, handling concept drift from combustion changes without full model retraining.

Lab analysis of effluent quality (COD, TSS, pathogen indicators) is slow and costly. No real-time visibility into treatment performance means delayed response to exceedances.

ML soft sensors trained on low-cost online measurements: turbidity, pH, conductivity. SVR (Support Vector Regression) and Cubist tree-based models predict key quality parameters continuously.

Tablet critical quality attributes (CQAs) — content uniformity, dissolution, hardness — are only measurable at end-of-batch, blocking Real-Time Release Testing (RTRT) and requiring full off-line QA.

Soft sensors infer CQAs from in-line PAT instruments (NIR, Raman, laser diffraction) and process data during manufacturing, following the FDA PAT guidance framework for science- and risk-based process understanding.

Metrology steps (CMP endpoint, film thickness, etch depth) are expensive and sampled at low frequency — creating quality blind spots across thousands of wafers between measurements.

Virtual Metrology (VM) predicts post-process quality from tool telemetry (RF power, chamber pressure, gas flows, process recipe) without physical measurement, enabling higher sampling density at near-zero cost.

Higher SAE automation levels require up to 28 sensors per vehicle at Level 5 vs 8 at Level 1. Physical sensors add cost, weight, power consumption, and failure points.

Virtual sensors (ideal, Hi-Fi, or RSI types) replace or supplement physical units, validated in full vehicle simulation environments. State estimation for chassis, drivetrain, and environment perception uses Kalman-family observers and hybrid ML approaches.