Dimensionality Reduction and Dynamical Mode Recognition of Circular Arrays of Flame Oscillators Using Deep Neural Network

This study proposes a novel Bi-LSTM-VAE-WDC framework that effectively reduces high-dimensional spatiotemporal combustion data to a low-dimensional phase space and utilizes Wasserstein distance-based classification to accurately recognize and distinguish dynamical oscillation modes in circular flame arrays, outperforming traditional PCA and VAE methods.

The Gist

The Big Picture: Taming the Fire Dance

Imagine a room full of eight candles arranged in a circle. If you light them all, they don't just burn steadily; they flicker, sway, and dance in complex patterns. Sometimes they move in perfect unison, other times they break into groups, and sometimes one or two go out, causing the whole circle to wobble differently.

In real-world engines (like those in airplanes), this "dancing" is dangerous. If the flames get out of sync, it can cause massive vibrations that break the engine apart. Engineers need to know exactly how the flames are dancing so they can stop the bad dances.

The Problem: The data describing these flames is overwhelming. It's like trying to describe a dance by tracking every single pixel of a video for every single flame, every millisecond. It's too much information for a human (or a standard computer) to make sense of quickly.

The Solution: The authors of this paper built a "smart translator" using Artificial Intelligence. This translator takes the chaotic, high-definition video of the flames and compresses it down into a simple, easy-to-read map.

---

The Three-Step Magic Trick

The researchers created a system called Bi-LSTM-VAE-WDC. That sounds scary, but think of it as a three-step process:

1. The Compression Machine (The Bi-LSTM-VAE)

Imagine you have a giant, messy library of books (the raw flame data). You want to summarize the whole library into just two numbers that tell you the "mood" of the story.

• VAE (Variational Autoencoder): This is the librarian who reads the books and writes a summary.
• LSTM (Long Short-Term Memory): This is the librarian's superpower. Regular librarians might forget what happened five pages ago. But this librarian has a "memory cell" that remembers the past and predicts the future. Since flames dance in a loop (time moves forward, but the pattern repeats), remembering the past helps predict the next step.
• Bi- (Bidirectional): This librarian reads the story both forwards and backwards to get the full context.

The Result: Instead of thousands of data points, this machine squeezes the entire flame dance into a tiny, 2D map (like a flat piece of paper). On this map, every different type of flame dance gets its own specific spot.

2. Drawing the Territory (Gaussian Kernel Density)

Once the flames are mapped onto this 2D paper, the researchers draw "clouds" around the dots.

• If the flames are doing a "Happy Dance," the dots cluster in the top-left corner.
• If they are doing a "Sad Dance," the dots cluster in the bottom-right.

The researchers use a special technique to draw smooth, fuzzy clouds around these dots. This shows us exactly where a specific type of dance lives on the map.

3. The Distance Ruler (Wasserstein Distance)

Now, imagine you have a new, unknown flame dance. You want to know what kind of dance it is.

• You drop the new dance onto your 2D map.
• You use a special ruler called the Wasserstein Distance.
• The Analogy: Imagine you have a pile of sand (the new dance) and a pile of sand in a specific shape (the known "Happy Dance"). The Wasserstein distance measures exactly how much effort it would take to move the new pile of sand to match the shape of the known pile.
• If the effort is low, the dances are the same. If the effort is high, they are different.

---

Why This Paper is a Big Deal

1. It's Better at Sorting than the Old Methods
The researchers compared their new "Smart Librarian" (Bi-LSTM-VAE) against older methods like PCA (a simple linear sorter) and standard VAEs.

• The Old Way: When they tried to map the dances, the "Happy" dots and "Sad" dots would overlap and mix together. It was like trying to sort red and blue marbles that were stuck in a ball of mud. You couldn't tell them apart.
• The New Way: The new method kept the dots perfectly separated. The "Happy" zone and the "Sad" zone were distinct islands with no bridges between them. This makes it incredibly easy to identify what the flames are doing.

2. It Handles Broken Symmetry
Real engines aren't perfect. Sometimes a flame goes out, or a nozzle gets clogged. This breaks the perfect circle.

• The researchers tested their system with 8 flames, then 7, then 6, and even with flames missing in different spots.
• Even when the circle was broken, their AI could still recognize the unique "fingerprint" of the new, broken pattern. It's like recognizing a song even if the singer is missing a few notes.

The Bottom Line

This paper presents a new, super-smart way to watch complex fire systems. By using a deep-learning AI that remembers time and a special mathematical ruler, they can take a chaotic, high-speed video of flickering flames and instantly say: "Ah, I know this dance! It's the 'Missing Flame' wobble, not the 'Perfect Circle' spin."

This is a huge step forward for making jet engines and gas turbines safer, because if we can recognize the dangerous dances early, we can stop the engine from breaking before it happens.

Technical Summary

1. Problem Statement

Combustion instabilities in aero-engines and gas turbines, particularly in annular combustion systems, pose significant operational risks. These instabilities arise from complex, multi-scale space-time couplings in thermal-fluid systems.

• The Challenge: Accurately recognizing various oscillation modes is a prerequisite for controlling these instabilities. However, the high-dimensional spatial-temporal data generated by complex combustion systems makes traditional dynamical mode recognition difficult.
• Limitations of Existing Methods:
  • Knowledge-based models: Struggle with increasing system complexity.
  • Linear Data-driven methods (PCA, POD, KPCA): Assume data can be represented by linear combinations of orthogonal bases, failing to capture complex non-linear dynamics.
  • Standard Non-linear methods (VAE): While effective, they may not fully exploit temporal correlations in cyclic flame oscillation data where past and future states are strongly correlated.
• Specific Context: The study focuses on circular arrays of flame oscillators (specifically octuple flickering buoyant diffusion flames) where symmetry breaking (e.g., missing flames) creates complex, asymmetric dynamical patterns that are difficult to monitor.

2. Methodology

The authors propose a novel hybrid framework called Bi-LSTM-VAE-WDC, which integrates dimensionality reduction with a distance-based classifier.

A. Dimensionality Reduction: Bi-LSTM-VAE

Instead of using standard fully connected layers in a Variational Autoencoder (VAE), the authors replaced them with a two-layer Bidirectional Long Short-Term Memory (Bi-LSTM) network.

• Architecture:
  • Encoder: Maps high-dimensional input data ($X$, containing velocity components $U, V, W$, temperature $T$, and time $t$) into a low-dimensional 2D latent space ($Z$). It utilizes a reparameterization trick ($\mu, \sigma^2$) to introduce stochasticity.
  • Decoder: Reconstructs the original data from the latent space.
  • Orthogonal Decomposition: An additional layer transforms the decoder output via Singular Value Decomposition (SVD) to aid pattern analysis.
• Why Bi-LSTM? Standard LSTMs process data sequentially (past to future). However, flame oscillations are cyclic. Bi-LSTM processes data in both forward and backward directions, capturing temporal correlations from both preceding and succeeding moments, which is crucial for periodic flame dynamics.
• Training: The model is trained to minimize Mean Squared Error (MSE) and Kullback-Leibler (KL) divergence using the Adam optimizer.

B. Mode Classification: Wasserstein Distance Classifier (WDC)

Once data is projected into the low-dimensional latent space, the authors use a 2D Wasserstein Distance (WD) based classifier to distinguish modes.

• Process:
  1. Density Estimation: Gaussian Kernel Density Estimates (GKDE) are computed for the latent points of different flame configurations.
  2. Grid Discretization: The 2D latent space is divided into a $100 \times 100$ grid to create discrete probability distributions ($P$ and $Q$).
  3. Distance Calculation: The 2D Earth Mover's Distance (EMD) is calculated between distributions. A smaller WD value indicates higher similarity between two dynamical modes.
• Classification: Test data is classified by comparing its WD value against benchmark distributions of known modes.

C. Numerical Simulation Setup

• System: Eight identical flickering buoyant diffusion flames arranged in a circle ($Re=100$).
• Configurations: Six scenarios were simulated, including a symmetric octuple array and five asymmetric cases with one or two missing flames (breaking rotational symmetry).
• Data: Time-series data collected from 15 sensors per flame (4 physical variables + time) at 1000 Hz for 5 seconds (approx. 50 periods). Total features: 481 per time step.

3. Key Contributions

1. Novel Architecture: First application of a Bi-LSTM-VAE specifically for reducing the dimensionality of circular flame oscillator data. The bidirectional approach effectively captures the cyclic temporal nature of flame flickering.
2. Unsupervised Mode Recognition: The integration of Wasserstein Distance in the latent space allows for the unsupervised classification of dynamical modes without requiring labeled training data for the classifier itself.
3. Superior Non-Linear Mapping: Demonstrates that Bi-LSTM-VAE creates a non-overlapping distribution of phase points in the 2D latent space, whereas standard VAE and PCA result in overlapping clusters, making mode distinction impossible.
4. Symmetry Breaking Analysis: Successfully identifies and distinguishes complex dynamical modes arising from symmetry breaking (missing flames) in circular arrays, a scenario relevant to real-world engine failures (e.g., engine-out scenarios).

4. Results

• Reconstruction Accuracy: The Bi-LSTM-VAE achieved significantly lower reconstruction errors (MSE) compared to both standard VAE and PCA.
  • Example (7.0D case): Bi-LSTM-VAE MSE = 0.150 vs. VAE (0.242) and PCA (0.563).
• Latent Space Separation:
  • PCA & VAE: Produced overlapping phase trajectories for different flame modes, making them indistinguishable in the 2D space.
  • Bi-LSTM-VAE: Produced distinct, non-intersecting loops (limit cycles) for each of the six configurations. This clear separation enables precise mode identification.
• Classification Performance: The WD-based classifier successfully matched test data to benchmark modes with 100% accuracy. The diagonal elements in the WD matrix (comparing a mode to itself) showed the lowest distance values (darkest colors in heatmaps), confirming the method's ability to quantify similarity.

5. Significance and Future Outlook

• Impact on Combustion Control: This method provides a robust tool for real-time monitoring and control of combustion instabilities in annular combustors, particularly in extreme scenarios where symmetry is broken (e.g., engine failure).
• Scalability: The approach shows potential for studying turbulent combustion and large-scale flame arrays where physical modeling is too complex and linear reduction methods fail.
• Future Work: The authors acknowledge that future research must address large-scale turbulent flames, turbulence-chemistry interactions, radiative heat loss, and thermo-acoustic instabilities in real turbine engines.

In summary, this paper presents a state-of-the-art deep learning framework that successfully overcomes the "curse of dimensionality" in combustion dynamics, offering a precise, unsupervised method for recognizing complex oscillation modes in symmetric and asymmetric flame systems.