Abstract
This paper proposes an end-to-end convolutional selective autoencoder approach for early detection of combustion instabilities using rapidly arriving flame image frames. The instabilities arising in combustion processes cause significant deterioration and safety issues in various human-engineered systems such as land and air based gas turbine engines. These properties are described as self-sustaining, large amplitude pressure oscillations and show varying spatial scales periodic coherent vortex structure shedding. However, such instability is extremely difficult to detect before a combustion process becomes completely unstable due to its sudden (bifurcation-type) nature. In this context, an autoencoder is trained to selectively mask stable flame and allow unstable flame image frames. In that process, the model learns to identify and extract rich descriptive and explanatory flame shape features. With such a training scheme, the selective autoencoder is shown to be able to detect subtle instability features as a combustion process makes transition from stable to unstable region. As a consequence, the deep learning tool-chain can perform as an early detection framework for combustion instabilities that will have a transformative impact on the safety and performance of modern engines.
Introduction
In high-performance engineering systems like jet engines and power plants, "combustion instability" represents a critical failure mode. This phenomenon occurs when heat release from a flame couples with the engine's acoustic pressure, resulting in massive, self-sustaining pressure oscillations. Because the transition from stable to unstable combustion is often a sudden "bifurcation," traditional sensors like pressure transducers frequently fail to provide sufficient warning. By the time a significant pressure spike is detected, catastrophic damage may already be underway.
To address this, researchers are turning to high-speed visual data. The onset of instability is characterized by specific geometric features, such as coherent vortex shedding (often appearing as "mushroom-shaped" structures). Detecting these precursors—short, millisecond-long bursts of instability known as intermittency—is the key to early warning systems.
Methodology: The Selective Autoencoder
The core of the proposed solution is a Deep Convolutional Selective Autoencoder (S-CAE). While a standard autoencoder aims for perfect reconstruction of all inputs, the S-CAE introduces a "Selectivity" constraint during training.
Implicit Labeling and Selective Reconstruction
The model is trained using a dataset of 63,000 grayscale frames (35,000 stable, 28,000 unstable) captured from a laboratory-scale swirl-stabilized combustor at 3,000 to 9,000 frames per second. The training process follows two rules:
- Unstable Reconstruction: The model must reconstruct unstable flame images as accurately as possible.
- Stable Suppression: For stable images, the model is trained to output a blank (black) frame.
This approach forces the network to ignore the "background noise" of normal combustion and become hyper-specialized in identifying the specific visual signatures of instability. To quantify the level of instability in real-time, the system calculates a Correlation Ratio between the input frame and the reconstructed output. A high correlation indicates the presence of unstable features.
Results and Discussion
The S-CAE was tested on transition videos where instability was manually induced by varying fuel or air flow rates. The model demonstrated a remarkable ability to localize intermittency periods that were invisible to traditional pressure-based analysis or standard Proper Orthogonal Decomposition (POD).
Key Performance Metrics
- Early Detection: The model identifies the very first "flickers" of instability, providing a continuous "soft label" (0 to 1) that indicates how close the system is to a tipping point.
- Computational Efficiency: The system is highly optimized for speed. It processed a sequence of 21,841 frames in approximately 35.5 seconds on a single NVIDIA Titan Black GPU, demonstrating potential for real-time monitoring in industrial settings.
Conclusion and Future Work
The Selective Autoencoder framework represents a significant step forward in engine safety. By turning reconstruction "failure" into a sensitive detection metric, the system can spot precursors to engine failure that were previously undetectable.
Future research will focus on transitioning from laboratory-scale combustors to field-ready systems, which involve harsher lighting and more complex geometries. Additionally, researchers plan to expand the model to "multi-class implicit labeling" to distinguish between different types of combustion anomalies, eventually integrating these visual alerts directly into automated engine control systems.
Summary updated