Early Detection of Combustion Instabilities using Deep Convolutional Selective Autoencoders on Hi-speed Flame Video (1603.07839v1)

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.

• The paper introduces a novel deep convolutional selective autoencoder that analyzes high-speed flame video for early detection of combustion instabilities.
• Empirical results from a laboratory combustor show the network detects instability indications earlier than existing methods, correlating with intermittent detection.
• This approach has strong implications for integration into industrial monitoring systems for preemptive alerts and potential application in other high-dimensional anomaly detection problems.

Early Detection of Combustion Instabilities using Deep Convolutional Selective Autoencoders on Hi-speed Flame Video

The paper "Early Detection of Combustion Instabilities using Deep Convolutional Selective Autoencoders on Hi-speed Flame Video" presents a sophisticated methodology aimed at tackling the formidable challenge of early combustion instability detection. The authors propose a novel convolutional selective autoencoder architecture that leverages deep learning techniques to analyze rapidly arriving flame image frames. This approach is particularly relevant due to the critical nature of combustion instabilities, which manifest as high-amplitude flame oscillations, leading to potential inefficiencies and safety concerns in gas turbine engines. The ability to detect such instabilities early is paramount for the operational safety of these systems.

In this paper, the authors utilize high-speed video footage of combustors to train an autoencoder designed to differentiate stable flame images from unstable ones. This differentiation is essential as combustor instability possesses a sudden, bifurcation-type nature, making early detection challenging using conventional methods. The convolutional selective autoencoder capitalizes on selectively masking stable flame images, allowing the model to focus on detecting subtle features indicative of impending instability.

The methodological framework consists of several distinctive phases: application of preprocessing, convolutional, and pooling layers; adoption of fully connected layers for feature abstraction; and implementation of a novel error minimization technique using the Nesterov momentum-based stochastic gradient descent. The network effectively mitigates the interpretability limitations associated with traditional black-box models by providing explicit insights into the features extracted during combustion instaibility detection.

The authors present empirical results based on data gathered from a laboratory-scale swirl-stabilized combustor. This validation demonstrates the network's capacity to discern instability indications earlier than existing techniques, signifying its potential as a transformative tool for maintaining combustion safety and efficiency. Specifically, the autoencoder outputs correlate well with intermittent detection—a precursor to full-blown instability. This capability is further underscored by investigating the relationship between consecutive frames, delineating a fine-grained understanding of flame dynamics throughout the transition from stability to instability.

The research emphasizes the strong implications of this approach in real-world applications, suggesting that such a network could be integrated into regular monitoring in industrial settings, providing preemptive alerts of combustion anomalies. Furthermore, this framework extends beyond combustion instability, presenting a robust technique applicable to a range of anomaly detection problems across high-dimensional domains.

There is considerable room for future exploration following this research. Extending this model to real-world, full-scale engine conditions is a logical next step for validating its wide applicability and robustness. Additionally, the authors envision adapting the framework for multi-class problems, thereby broadening its utility in diverse implicit labeling scenarios that require labeling beyond binary stable-unstable classifications.

In conclusion, this work introduces a significant advancement in automated diagnostics for combustion systems. The application of deep convolutional selective autoencoders leverages the power of deep learning to ensure proactive measures can be implemented in scenarios of combustion instability, providing a deeper understanding of complex fluid dynamics phenomena and safeguarding turbine operations. Through continued validation and expansion, this research could fundamentally enhance predictive maintenance practices within the aerospace and energy industries.