Machine Learning based Ensemble Flame Regime Classification for Mesoscale Combustors based on Insights from Linear and Nonlinear Dynamic Analysis

This study employs Recurrence Quantification Analysis and Statistical-Spectral analysis of OH* chemiluminescence and acoustic pressure signals to extract dynamical features from mesoscale combustor flames, which are then utilized in a stacking ensemble machine learning framework to accurately classify distinct flame regimes such as stable, extinction-ignition, and propagating flames.

The Gist

🌟 The Big Picture: Teaching a Computer to "Listen" to Fire

Imagine you have a tiny, high-tech campfire inside a glass tube. This isn't just any fire; it's a mesoscale combustor—a miniature engine the size of a pencil lead. These tiny engines are the future of portable power (think super-efficient batteries for drones or phones), but they are tricky. Sometimes they burn steadily, sometimes they sputter and die, and sometimes they race back and forth like a runaway train.

The researchers wanted to solve a puzzle: How can we tell exactly what kind of fire is happening just by listening to it and watching it flicker?

They didn't just look at the fire; they used Machine Learning (smart computer programs) to act like a super-sensei, teaching the computer to recognize the unique "personality" of three different types of flames.

---

🔥 The Three "Personalities" of the Flame

The team identified three distinct ways the fire behaves, like three different characters in a play:

1. The Steady Camper (Stable Flame):

   • What it does: It sits in one spot, burning calmly and consistently.
   • The Sound: It sounds like white noise (like a radio tuned between stations) or a gentle, chaotic hiss.
   • The Vibe: It's boring but reliable.

2. The Sneezy Fire (FREI - Flames with Repetitive Extinction and Ignition):

   • What it does: It lights up, runs a short distance, dies out, waits a moment, and then sneezes (ignites) again. It's a cycle of "Light! Go! Die! Wait! Light!"
   • The Sound: It sounds like a rhythmic drumbeat. Boom... pause... Boom... pause.
   • The Vibe: It's a stop-and-go traffic jam of fire.

3. The Rocket Runner (Propagating Flame):

   • What it does: Once it lights, it doesn't stop. It races all the way to the other end of the tube before dying, then restarts.
   • The Sound: This is the loudest. As it races, it creates a powerful, high-pitched hum (like a jet engine) because the fire is shaking the air inside the tube.
   • The Vibe: It's a high-speed chase with a lot of noise.

---

🕵️‍♂️ The Detective Work: How They Analyzed the Fire

The researchers didn't just guess; they used two different "detective kits" to analyze the data from high-speed cameras and microphones.

Kit #1: The "Time Travel" Map (Nonlinear Analysis)

Imagine taking a photo of the fire's behavior every millisecond and stacking them to create a 3D map.

• The Analogy: Think of a dance floor.
  • If the fire is Steady, the dancers are moving randomly everywhere (chaos).
  • If the fire is Sneezy, the dancers are doing a specific routine over and over (a perfect circle).
  • If the fire is a Runner, the dancers sprint in a line, then stop, then sprint again (a long, straight path with breaks).
• The Tool: They used something called Recurrence Quantification Analysis (RQA). It's like looking at a "repetition map" to see if the fire is doing the same dance moves again and again.

Kit #2: The "Music Producer" (Linear/Spectral Analysis)

This kit looks at the fire's sound and light as if it were a song.

• The Analogy: Think of an equalizer on a stereo.
  • They checked: Is the sound mostly high-pitched (noise)? Is there a deep bass beat (rhythm)? Is the song chaotic or organized?
• The Tool: They measured things like "how loud the bass is" (Dominant Frequency) and "how messy the sound is" (Entropy).

---

🤖 The "Brain" That Learned the Difference

Once they had all this data (the dance maps and the music stats), they fed it into a Machine Learning "Stacking Ensemble."

• The Analogy: Imagine a panel of four expert judges (a math wizard, a pattern spotter, a logic bot, and a probability guru).
  • Each judge looks at the fire data and votes: "Is this the Steady Camper, the Sneezy Fire, or the Rocket Runner?"
  • Then, a Head Coach (the Meta-Learner) listens to all four judges. The Coach doesn't just take a majority vote; it learns how to combine their opinions to make the perfect final decision.

🏆 The Results: Did It Work?

Yes, perfectly.

• The computer got the classification right almost 100% of the time.
• The Surprise: The researchers found that you didn't actually need the complex "Time Travel" maps (the nonlinear stuff) to get the right answer. The simpler "Music Producer" stats (linear stuff) were enough to tell the difference between the flames perfectly.
• Why this matters: It means we can build smaller, cheaper, and faster sensors for these tiny engines. We don't need super-complex math to know if the engine is running safely or about to fail.

💡 The Takeaway

This paper is like teaching a computer to recognize the difference between a humming refrigerator, a ticking clock, and a screaming siren just by listening to the sound waves.

By understanding these "personalities" of fire in tiny engines, engineers can design better, safer, and more efficient micro-power systems for the future. The fire isn't just burning; it's speaking, and now, thanks to this study, we finally know how to understand its language.

Technical Summary

1. Problem Statement

Micro- and mesoscale combustors are critical for compact power generation (MEMS) and fire safety technologies. However, combustion in these confined channels is highly sensitive to flame-wall interactions, leading to complex, unsteady flame behaviors that differ significantly from large-scale systems.

• The Challenge: Identifying and classifying distinct flame regimes (Stable, Repetitive Extinction/Ignition, and Propagating) is difficult due to the complex interplay of thermo-acoustic coupling, heat losses, and radical quenching.
• The Gap: While linear and nonlinear time-series analysis tools (like Recurrence Quantification Analysis) have been used for large-scale thermo-acoustic instability detection, they have not been explicitly applied to characterize the unique topologies of mesoscale flame regimes. Furthermore, there is a need for robust, automated classification frameworks to distinguish these regimes based on experimental signals.

2. Methodology

Experimental Setup

• Apparatus: An optically accessible mesoscale combustor consisting of a 5 mm inner diameter quartz tube (320 mm long) with an external heater to simulate upstream heat recirculation.
• Fuel: Premixed methane-air.
• Parameters: Equivalence ratios ($\Phi$) of 0.8, 1.0, and 1.2; Mixture velocities ($\bar{u}$) from 0.2 to 0.7 m/s (Reynolds numbers 64–224).
• Data Acquisition:
  • OH Chemiluminescence:* Captured via a Photomultiplier Tube (PMT) to represent heat release rate.
  • Acoustic Pressure: Captured via a microphone downstream of the exit.
  • High-Speed Imaging: 4000 fps to visualize flame spatial dynamics.

Data Analysis Framework

The study employed a two-pronged approach: Physical Characterization followed by Machine Learning Classification.

A. Signal Preprocessing & Feature Extraction

1. Sliding Window & Normalization: Signals were segmented with overlap and Z-score normalized to preserve underlying dynamics while standardizing variance.
2. Nonlinear Analysis (Recurrence Quantification Analysis - RQA):
   • Phase space reconstruction using Takens' embedding theorem (optimized time lag $\tau$ and dimension $d$).
   • Construction of Recurrence Plots (RPs) using a Fixed Amount of Neighbors (FAN) thresholding scheme (2% recurrence rate).
   • Features: 20 RQA metrics extracted (e.g., Recurrence Rate, Determinism, Laminarity, Trapping Time, Entropy) to quantify stochasticity, periodicity, and intermittency.
3. Linear/Spectral Analysis:
   • Statistical-Spectral Features: 22 features extracted including Skewness, Kurtosis, Lag-1 Autocorrelation, Dominant Frequency, Spectral Centroid, Spectral Entropy, and Spectral Slope.
   • Time-Frequency Analysis: Continuous Wavelet Transform (CWT) scalograms were generated to visualize the evolution of frequency content over time.

B. Machine Learning Classification

• Framework: A Stacking Ensemble approach.
• Base Learners: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Naive Bayes, and Logistic Regression.
• Meta-Learner: Multi-Layer Perceptron (MLP).
• Training Strategy: Stratified 5-fold cross-validation, SMOTE for class imbalance, and rigorous hyperparameter tuning via GridSearchCV.
• Validation: A held-out test set comprising unseen operating conditions was used to prevent data leakage and ensure generalizability.
• Dimensionality Reduction: PCA and Isomap were used to visualize feature space clustering.

3. Key Contributions

1. Physical Insight into Mesoscale Dynamics: The study elucidates the distinct dynamical signatures of three flame regimes (Stable, FREI, Propagating) using both linear and nonlinear tools, linking signal topologies to physical phenomena (e.g., thermo-acoustic coupling vs. relaxation oscillation).
2. Novel Application of RQA: First-time application of Recurrence Quantification Analysis specifically to mesoscale flame regimes to decode their topological structures.
3. Robust Classification Framework: Development of a Stacking Ensemble model that achieves near-perfect classification accuracy using features derived from both RQA and Statistical-Spectral analyses.
4. Comparative Analysis: Demonstrated that while RQA provides deep physical insight into nonlinear dynamics, simpler linear Statistical-Spectral features are sufficient for high-accuracy regime classification, offering a computationally efficient alternative.

4. Key Results

A. Dynamical Characterization of Flame Regimes

• Stable Flame (SF):
  • Dynamics: Stationary flame at a fixed location.
  • Signatures: Signals resemble high-frequency stochastic noise. Recurrence plots show homogeneous point distributions; phase plots show stochastic trajectories.
• Flames with Repetitive Extinction and Ignition (FREI):
  • Dynamics: Cyclic ignition, upstream propagation, and extinction within the downstream half of the tube.
  • Signatures: Relaxation oscillator behavior. Heat release shows periodic rise/decay; acoustic pressure shows damped oscillator response. Recurrence plots show parallel diagonal lines (periodicity) with vertical structures (intermittency). No sustained thermo-acoustic coupling due to short flame travel distance.
• Propagating Flame (PF):
  • Dynamics: Flame traverses the entire tube length, extinguishing only at the upstream mesh.
  • Signatures: Strong thermo-acoustic coupling during propagation. Scalograms show a single-frequency response (natural harmonic) coupled with heat release. Recurrence plots show large "box-like" structures (intermittency) where high-frequency oscillations are overshadowed by the cycle time scale. Phase plots show disc-like or "mace" structures.

B. Classification Performance

• Accuracy: The Stacking Ensemble framework achieved high accuracy (near-perfect) in distinguishing the three regimes using both feature sets independently.
• Feature Efficacy:
  • RQA Features: Successfully captured the nonlinear structural differences, confirmed by clear clustering in Isomap and PCA maps.
  • Statistical-Spectral Features: Also achieved flawless classification. This suggests that for the specific task of regime identification, computationally cheaper linear features are sufficient, though RQA offers superior physical interpretability.
• Visual Validation: Dimensionality reduction maps (PCA/Isomap) showed distinct, non-overlapping clusters for SF, FREI, and PF, confirming the physical distinctness of the regimes.

5. Significance and Conclusion

This work bridges the gap between complex nonlinear dynamical systems theory and practical combustion engineering.

• Scientific Impact: It provides a rigorous framework for understanding how flame topology evolves in mesoscale channels, specifically highlighting the transition from stochastic behavior (Stable) to relaxation oscillations (FREI) and intermittent thermo-acoustic coupling (Propagating).
• Engineering Impact: The proposed ML framework offers a robust, real-time capable method for monitoring and controlling mesoscale combustors. By identifying the regime based on OH* or acoustic signals, systems can be actively managed to prevent unwanted instabilities or extinction.
• Methodological Value: The study validates that while deep nonlinear analysis (RQA) is essential for understanding the physics, standard statistical-spectral features are highly effective for classification, allowing for a trade-off between computational cost and physical insight depending on the application needs.