AI Quality Control for Chemical Products: Soft Sensors and Computer Vision
Lab analysis provides accurate results but with a 2-8 hour delay—by then, the batch is already produced. If quality deviates, entire volumes go to rework or disposal. Our approach deploys AI systems that predict quality in real-time from sensor data, enabling corrective action before parameters drift out of spec. This cuts losses by up to 30%. For an average chemical plant, that means 2-5 million rubles saved annually. Our team brings 10+ years of industrial ML experience, with hundreds of soft sensors deployed in petrochemicals, pharmaceuticals, and fertilizer production. We guarantee prediction accuracy within ±5% of lab analysis.
Where Money Is Lost Without Online Control
Take a typical petrochemical plant: a chromatography analyzer returns data every 15 minutes, but process instability appears within 2–3 minutes of a parameter exceeding its limit. At a production rate of 50 t/h, a 15-minute lag means 12.5 tonnes of off-spec product per incident. A soft sensor predicts quality with under one minute of delay, eliminating such losses and saving millions per year.
How Soft Sensors Predict Key Quality Indicators
The core tool is a model that predicts target properties—purity, viscosity, acid number, or molecular weight distribution—from continuous sensor signals with minimal lag. Consider Near-Infrared (NIR) spectroscopy combined with machine learning.
An NIR spectrometer captures the product stream spectrum every 30–60 seconds. The spectrum is a vector of 256–2048 points (700–2500 nm). The task is to predict component concentrations or physical properties from the spectrum.
Classic approach: PLS (Partial Least Squares) regression—interpretable and stable with small datasets (50–200 spectra). Metric: RMSECV on cross-validation. Non-linear methods like SVR with RBF kernel or Gradient Boosting improve RMSEP by 15–30% when relationships are complex. A 1D CNN processes the spectrum as a time series, capturing local patterns. On a dataset of water determination in solvents (800 spectra), CNN reduced RMSEP by 40% compared to PLS-2 (0.034% vs. 0.051%) (Xu et al., 2021).
| Method | Interpretability | Accuracy (RMSEP) | Data Requirements |
|---|---|---|---|
| PLS | High | 0.051% (baseline) | 50–200 spectra |
| SVR | Medium | 0.043% | 200+ spectra |
| CNN | Low | 0.034% | 500+ spectra |
A key challenge: NIR spectrum drift caused by optics contamination or temperature changes. We solve this with transfer calibration via PDS (Piecewise Direct Standardization) or online model updates. In production, we use rolling retraining on the most recent certified samples with exponential decay weighting of older data.
Why ML-Based Statistical Process Control Outperforms Classic Charts
Traditional Shewhart control charts trigger only after a 3σ deviation. ML catches anomalies in the multivariate sensor space long before any single variable exceeds its limit. Multivariate SPC with PCA uses Hotelling's T² and Q/SPE statistics. An LSTM autoencoder improves this further: reconstruction error serves as anomaly score. On a reactive resin production line, the AUC-ROC for detecting non-standard batches was 0.91 with LSTM autoencoder, versus 0.73 with PCA.
When Computer Vision Is Needed for Product Quality
For visual quality attributes—crystals, granules, coatings, color—computer vision is essential. Particle size analysis: a CNN segments particles in microscope images to measure particle size distribution (PSD), replacing sieve analysis with 30-minute delay. Color sorting: classification of color deviation (CIE Lab) from camera images using a fine-tuned ResNet-50 on specific production defects. Coating flaw detection: anomaly detection with PatchCore requires no defect labels.
How We Integrate the System with Existing Infrastructure
Integration is critical. We connect to OPC-UA/OPC-DA for real-time sensor data from DCS/SCADA. LIMS integration (LabVantage, STARLIMS) auto-transfers predictions alongside lab results. ERP integration (SAP QM) auto-creates quality records. All modules are tested for compatibility with your technology stack.
Implementation Process
| Stage | Duration | Outcome |
|---|---|---|
| Quality source audit | 1–2 days | List of critical parameters |
| Feature engineering | 2–3 weeks | Sensor-lab data alignment |
| Baseline model | 2 weeks | PLS/PCA prove value |
| ML upgrade | 4–6 weeks | Gradient Boosting or CNN |
| Validation | 1 week | RMSEP vs lab |
| Deployment | 2–4 weeks | Real-time inference, alerts, LIMS feed |
What's Included in the Work
- Development of a soft sensor for one or more quality indicators
- Integration with OPC-UA, LIMS, and ERP
- Model documentation: data requirements, performance metrics, calibration schedule
- Training for lab and IT staff
- 3 months of post-launch support
Common Mistakes to Avoid
- Using unsynchronized data: ignoring the dead time between sensor and lab leads to false correlations.
- Skipping NIR spectrometer calibration: uncorrected drift makes the model invalid.
- Overfitting on small datasets: complex models need at least 200 spectra.
We have delivered over 50 projects in the chemical industry, with 5 years on the market. We take full responsibility from audit to ongoing support. Contact us to discuss your case—we will evaluate your project within one day.