Source-linked AI summary
AI-Powered Flare Combustion Efficiency Estimation
Afeefa Azam, Iyyakutti Iyappan Ganapathi, Fares Ossama Abdelhafez, Divya Velayudhan, Maregu Assefa Habtie, Hamad Karki, Khalid Yousef Al Awadhi, Naoufel Werghi
TL;DR
Flare operators need reliable, real-time combustion-efficiency measurements, but conventional instruments are costly, fragile, and calibration-intensive for remote deployment. The paper proposes a vision-language encoder with a regression model that predicts efficiency from thermal video, achieving R2 = 0.822 while supporting deployment through a graphical interface.
Problem
Conventional gas analyzers and hyperspectral cameras are expensive, fragile, and require frequent calibration, limiting practical combustion-efficiency monitoring at remote or budget-constrained sites.
Method
The paper combines thermal video, synchronized sensor-derived labels, a fine-tuned vision-language encoder, and a regression head to estimate combustion efficiency.
Results
R2 = 0.822, with the proposed regressor achieving the lowest errors across every reported metric.
Takeaways & Limitations
The framework provides a vision-centric route to real-time flare-efficiency estimation that can reduce reliance on specialised instrumentation.
Abstract
from arXiv · showhide
Achieving high combustion efficiency in flare stacks is crucial for adhering to regulatory standards and controlling the release of hydrocarbons into the environment. Traditional instruments like gas analyzers and hyperspectral cameras are expensive, fragile, and require frequent calibration, which makes them impractical for remote or budget constrained industrial sites. We propose an innovative solution that combines a lightweight vision-language encoder with a compact multi-layer perceptron to predict combustion efficiency directly from low-cost thermal video footage. The fully trained model is integrated into an easy-to-deploy graphical user interface. This interface overlays predicted combustion efficiency values on each video frame, displays real-time trends in combustion efficiency, shows the distribution of combustion efficiency across all frames in the video, and allows users to export CSV reports. Over a six-month period, the system achieved 99% uptime and required less than 15 minutes of maintenance per week.
I. INTRODUCTION
The paper motivates low-cost, scalable flare-efficiency estimation because conventional instruments are expensive, calibration-intensive, and unreliable at remote sites. It proposes a vision-centric framework using synchronized thermal imagery and sensor-derived labels, with vision-language representations outperforming traditional backbones.
- I. INTRODUCTION: Conventional gas analyzers and hyperspectral cameras are expensive, require frequent calibration, and can fail under harsh flare-site conditions.These constraints limit large-scale and remote deployment and prevent extracting efficiency from standard video feeds.
- I. INTRODUCTION: The proposed framework estimates combustion efficiency from synchronized thermal video and sensor data using a vision-centric learning approach.The framework combines thermal streams with low-cost gas-analyzer and weather-sensor measurements.
- I. INTRODUCTION: A custom platform captures RGB, infrared, and thermal imagery alongside CO, CO2, and NO2 measurements, which are converted into scalar efficiency labels.The dataset addresses the absence of a public multimodal flare-analysis dataset.
- I. INTRODUCTION: The preprocessing pipeline synchronizes frame rates, aligns camera views, and pairs each frame with corresponding sensor readings before model training.This produces frame-level multimodal examples for efficiency estimation.
- I. INTRODUCTION: Vision-language representations outperform traditional backbones for combustion-efficiency prediction, with the proposed regressor achieving R2 = 0.822 and the lowest errors across every metric.The authors attribute this performance to image-text pre-training that provides combustion-aware semantics transferable to regression.
- I. INTRODUCTION: Prior work includes sensor-based, infrared, RGB, and multimodal vision approaches, while control strategies remain dependent on trustworthy sensors and manual tuning.These dependencies are described as brittle under shifting process conditions.
III. IMAGE–TEXT ALIGNMENT
The image–text alignment stage converts scalar combustion-efficiency values into three natural-language prompts. The prompts assign each image to low, medium, or high efficiency according to explicit thresholds.
- III. IMAGE–TEXT ALIGNMENT: Each flare image is paired with an efficiency-derived natural-language prompt representing its combustion-efficiency band.The method maps the scalar value η into one of three semantically distinct descriptions.
- III. IMAGE–TEXT ALIGNMENT: η > 95% is labeled high combustion efficiency.The corresponding prompt describes a flare exhibiting high combustion efficiency.
- III. IMAGE–TEXT ALIGNMENT: 80% < η ≤95% is labeled medium combustion efficiency.The corresponding prompt describes a flare exhibiting medium combustion efficiency.
- III. IMAGE–TEXT ALIGNMENT: η ≤80% is labeled low combustion efficiency.The corresponding prompt describes a flare exhibiting low combustion efficiency.
B. Visual Pre-processing
The system preprocesses frames for CLIP-based visual encoding and trains the representation with image-text and regression objectives. A regression head converts the learned visual representation into quantitative combustion-efficiency predictions.
- B. Visual Pre-processing: Frames are resized to 224 × 224 pixels and normalized with CLIP ViT-B/32 statistics, without color-space augmentation.The preprocessing preserves domain-specific cues and avoids artificial chromatic variations.
- B. Visual Pre-processing: CLIP ViT-B/32 uses 12-layer vision and text Transformers that project inputs into a shared 512-dimensional embedding space.The branches are optimized with cosine similarity under a contrastive learning objective.
- B. Visual Pre-processing: The training setup retains the CLIP backbone while the regression head predicts efficiency and receives supervision through an L1 loss.The supplied training passages also specify λ = 0.1, AdamW optimization, a cosine one-cycle schedule, and gradient clipping at norm 1.0.
- B. Visual Pre-processing: Mini-batches contain 64 image-text pairs, and hard-negative mining is disabled to preserve the ordinal regression signal.Validation is performed after each epoch using MAE and Pearson correlation among the reported metrics.
- B. Visual Pre-processing: Classification accuracy is also evaluated across low, medium, and high efficiency bands.
- B. Visual Pre-processing: The model augments CLIP representations with a regression head that maps flame patterns directly to quantitative combustion efficiency.This design is intended to eliminate the need for specialised instrumentation during prediction.
E. Curating High-Confidence Image–Text Pairs with a Pretrained CLIP Scorer
The method uses CLIP similarity to curate high-confidence image–text pairs after fine-tuning. Pairs below a conservative similarity threshold are discarded to retain examples whose captions closely match their visual representations.
- E. Curating High-Confidence Image–Text Pairs with a Pretrained CLIP Scorer: Image–caption pairs with cosine similarity below 0.9 are discarded during high-confidence data curation.Similarity is computed between L2-normalized image and text vectors; scores near 1 indicate stronger visual-textual alignment.
IV. PROPOSED REGRESSOR
The proposed regressor freezes a fine-tuned CLIP image encoder and uses a compact MLP to map visual embeddings to combustion-efficiency estimates. Training uses standardized 224 × 224 inputs, squared-error optimization, and validation-based checkpoint selection.
- A frozen ViT-B/32 encoder produces a 512-dimensional, L2-normalised image embedding for regression.The encoder processes 224 × 224 × 3 input images.
- A single-hidden-layer MLP maps the embedding to a scalar combustion-efficiency estimate, using dropout probability 0.2 after ReLU.
- Images are resized to 224 × 224 pixels, standardized with ImageNet statistics, and split into stratified 80% training and 20% validation folds without augmentation.Stratification preserves the combustion-efficiency distribution across folds.
- Parameters are optimized with Adam at a constant learning rate of 1×10^-4 using a squared-error objective.
- The model trains for up to 200 epochs with patience-5 early stopping, retaining the checkpoint with the lowest validation MSE.
V. EXPERIMENTAL SETUP
The experiment synchronizes and annotates multimodal recordings, then compares nine vision backbones for single-frame combustion-efficiency regression. The proposed regressor achieves R2 = 0.822, the lowest errors across metrics, and outperforms the reported alternatives.
- Recordings are synchronized to 25 fps, thermally registered to a reference, aligned with sensor timelines, and tagged with efficiency and wind measurements.
- The comparison reports MAE, MSE, RMSE, and R2, where lower error values and higher R2 indicate better predictive fidelity.
- R2 = 0.822, with the proposed regressor achieving the lowest errors across every reported metric and outperforming all competitors by a wide margin.Table I evaluates MAE, MSE, RMSE, and R2 on the held-out set.
- DenseNet121 is the strongest conventional CNN, but its R2 lags CLIP by 35 percentage points.The passage attributes the gap to insufficient capture of subtle flame-related cues by densely connected features alone.
- ViT-B/16, SWIN-Tiny, and EfficientNet-B7 underperform relative to their parameter counts in this single-frame regression setting.The passage suggests temporal context or stronger inductive biases may be needed for larger gains.
VI. GUI DESIGN AND FUNCTIONALITY
The GUI supports real-time combustion-efficiency prediction from thermal video by combining frame processing, visual overlays, temporal trends, and distribution analysis.
- A real-time plot displays framewise CE trends using both raw and smoothed curves, with controls for video navigation, analysis, and report export.
- The interface overlays CE values from 0-100% on each frame and provides adjustable opacity for visual feedback.
- The processing pipeline extracts and preprocesses frames, detects flame regions, calculates CE with the proposed regressor, and validates and smooths results.
- 15-30 fps processing speed enables real-time combustion-efficiency analysis on standard industrial hardware.
- Distribution analysis includes configurable rolling-average windows, adjustable histogram bins, dynamic range adaptation, and interactive element highlighting.
C. Statistical Analysis and Real-Time Graphs
The system provides statistical summaries and graphing tools to examine combustion-efficiency trends, variability, distribution shape, and relationships with flame characteristics.
- Statistical analysis reports central tendency, dispersion, and distribution-shape measures for combustion efficiency.
- Available graphs include temporal CE-versus-time lines, CE-versus-flame-characteristics scatter plots, and process-monitoring control charts.
- GUI auxiliary controls are organized around their primary operational use cases.
D. Report Generation
The reporting subsystem combines statistical summaries, visualizations, file-management controls, and automated export options to support monitoring and documentation.
- Export formats include CSV raw frame data, PDF formatted reports, and Excel tables and charts.
- Reports can include executive summaries, detailed statistical analyses, and visualizations such as graphs and distributions.
- Automated reporting supports scheduled generation, email distribution, and cloud-storage integration.
- Common tasks were completed in 3-5 clicks, while new users achieved basic proficiency in least time.
- 4.8/5 average user-survey rating indicates favorable reported satisfaction with the GUI.
- The prototype is intended for rapid deployment in oil and gas facilities, with installation typically completed in minimum hours and reporting supporting operational decisions and regulatory documentation.
VII. CONCLUSION
The paper presents an end-to-end vision-centric framework for real-time flare combustion-efficiency estimation using thermal cameras, gas analysis, temporal alignment, and vision-language representations. Single-frame inference achieved MAE 0.039 and R2 0.822, while future work targets richer environmental and multimodal inputs and improved accuracy.
- MAE 0.039 and R2 0.822 were achieved for single-frame combustion-efficiency inference using vision-language representations.
- The framework combines thermal cameras with a gas analyzer and uses frame-level temporal alignment and spatial registration to create a labeled dataset.
- Future work includes adding synchronized environmental variables such as wind speed and humidity.
- Future work also proposes multimodal video embeddings and aims to improve predictive accuracy while retaining low-cost deployability.