Source-linked AI summary
Early Detection of Combustion Instabilities using Deep Convolutional Selective Autoencoders on Hi-speed Flame Video
Adedotun Akintayo, Kin Gwn Lore, Soumalya Sarkar, Soumik Sarkar
TL;DR
Combustion instability harms engine efficiency and longevity, yet its onset is difficult to detect because flame-video changes are fast and transient. The paper develops a convolutional selective autoencoder that learns from explicitly labeled stable and unstable flames to generate soft transition labels. In laboratory-scale combustor videos, the framework suppresses stable frames, reveals unstable features, and localizes intermittency before full-blown instability.
Problem
Early instability onset is difficult to detect because combustion transitions are fast and spatio-temporally transient, despite instability reducing aircraft gas-turbine efficiency and longevity.
Method
An end-to-end convolutional selective autoencoder learns soft labels from explicitly labeled stable and unstable flame frames in high-speed video.
Results
The model suppresses stable-region frames, reveals unstable features, and partially reveals intermittent unstable features before full-blown instability.
Takeaways & Limitations
The framework provides an interpretable early-detection and soft-labeling approach for high-dimensional flame-video data.
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
The paper proposes a convolutional selective autoencoder for implicit labeling and early detection of combustion instability from high-dimensional flame videos. It targets subtle transition phenomena while avoiding extensive expert-guided feature engineering and is validated under realistic laboratory conditions.
- The approach derives soft labels from explicitly labeled stable and unstable flame classes to track continuous transitions.
- A convolutional selective autoencoder is proposed for early detection of combustion instability.
- The framework learns coherent image structures from high-dimensional, rapidly arriving video without extensive expert-guided feature handcrafting.
- A granularity-controlled metric is constructed to detect pre-transition intermittence, defined as millisecond-scale, partially observable instability bursts.
- Validation uses laboratory-scale combustion data collected under various realistic operating conditions.
2. BACKGROUND
The background reviews labeling methods and convolutional networks before framing combustion instability as a high-impact, transient visual-structure problem. Implicit labeling supplies soft transition information from hard labels and is fused with a convolutional autoencoder.
- Convolutional networks reduce dimensionality through local neighborhood matching and learn hierarchical features using shared kernels.
- 2.2 The problem of combustion instability: Combustion instability reduces aircraft gas-turbine efficiency and longevity and involves high-amplitude flame oscillations at discrete acoustic frequencies.
- 2.2 The problem of combustion instability: Coherent vortical structures drive large-scale velocity and flame-shape oscillations, motivating image-based analysis alongside POD and DMD.
- 2.3 Implicit labeling: Implicit labeling derives soft labels from explicitly labeled extreme positive and negative classes to represent intermediate transition states.
AUTOENCODER
The paper uses an end-to-end convolutional autoencoder that encodes flame-image features and reconstructs outputs while selectively masking undesired classes. Its architecture combines convolution, pooling, bottleneck encoding, deconvolution, unpooling, regularized training, and correlation-ratio similarity measurement.
- Labeling and preprocessing: Explicit labels selectively mask undesired-class flame frames with black pixels to create input-output training pairs.Stable and unstable frames are assigned ground-truth classes before masking and normalization.
- Feature extraction: Convolutional and deconvolutional layers learn feature maps from local image patches, while max pooling selects highly activated local features.The architecture uses learned filters to enforce local correlation and pooling for representative feature selection.
- Architecture: The convolutional autoencoder learns image features through convolution and subsampling, then reconstructs original dimensions through deconvolution and upsampling.The encoder extracts meaningful features, while the decoder reconstructs the output into the original image dimensions.
- Architecture: The bottleneck encodes the most important features after feature maps are flattened into row vectors and passed through fully connected layers.This stage reduces the representation before decoding.
- Optimization: The model is trained with regularization and Nesterov momentum-based stochastic gradient descent, with weights updated iteratively using a learning rate.The regularization function is used to avoid overfitting, and the learning rate is the optimization step size.
- Similarity measurement: The correlation ratio measures similarity between image frames with differing intensities, ranging from 0 for uncorrelated images to 1 for fully correlated images.The metric was selected for low computational requirements and because it does not directly include the actual images in the computation.
4. DATASET AND IMPLEMENTATION
The study collects high-speed flame videos from a laboratory-scale swirl-stabilized combustor and trains the network on labeled stable and unstable frames. Transition videos then support implementation and evaluation of the trained model, with GPU execution enabling rapid processing.
- Dataset collection: Thermoacoustic instability was induced in a laboratory-scale combustor with a 30 mm swirler to collect high-speed training videos under four conditions.Videos captured 9000 frames over 3 seconds at specified fuel and air flow rates and premixing settings.
- Transition testing: Five 7-second transition videos captured combustion progressing from stable to unstable through intermittence while fuel flow was reduced or air flow increased.Intermittence is described as fast switching between stability and instability preceding persistent instability.
- Training implementation: 63,000 grayscale frames were resized from 100 × 237 to 64 × 64, with 35,000 labeled stable and 28,000 labeled unstable.The training images combined datasets spanning two premixing lengths and varied air and fuel flow rates.
- Training implementation: The implementation used 3 × 3 filters, non-overlapping 2 × 2 max pooling, batches of 128 examples, and early stopping selected through cross validation.These settings were reported as experimentally less costly or suitable for training.
- Results and execution: 21,841 transition-sequence frames were processed in approximately 35.5 seconds using the trained model and GPU acceleration.The network contained 5,090,249 learnable parameters, and training progress was tracked through training and validation losses.
5. RESULTS AND DISCUSSIONS
The selective autoencoder filters flame-video features to expose instability during transitions, including subtle intermittent behavior before full instability. Its feature maps and soft labels support interpretation of stable, unstable, and intermediate flame states.
- Early Detection of combustion instability: The model is trained with stable frames masked as zero and unstable frames retained, reproducing explicit labels while exposing learned feature maps.The fully connected layers compress images into explanatory features and reshape them to reconstruct outputs with input-like dimensions.
- Early Detection of combustion instability: Unstable frames activate more feature-map units responding to mushroom structures, whereas stable-frame information rapidly diffuses through hidden layers.Joint parameters balance discarded and retained information from stable and unstable training sets.
- Transition protocols: The model suppresses stable-region outputs and reveals unstable-region outputs, functioning as a filter for selected flame features.A similarity measure evaluates masking strength, and local regression smoothing visualizes transitional trends.
- Transition protocols: Intermittent regions partially reveal unstable features, and the proposed metric localizes intermittencies before full-blown instability more prominently than other approaches.The paper links better intermittency tracking with earlier detection and fewer false alarms.
- Transition protocols: Three consecutive transition frames show gradual mushroom-structure growth and increasing average instability measures, indicating graduated movement between stable and unstable regions.The framework’s soft labels can help restore frame positions in static applications, while hidden-layer marginalization captures neighboring flame patterns.
6. CONCLUSIONS AND FUTURE WORKS
The paper concludes that an end-to-end convolutional selective autoencoder can generate fuzzy labels from hard labels for early combustion-instability detection. Laboratory-scale validation supports its diagnostic use and interpretation of coherent structures, while broader validation and multiclass extensions remain future work.
- Conclusions and Future Works: An end-to-end convolutional selective autoencoder generates fuzzy labels from hard-labeled examples for early detection using hi-speed flame video.The framework derives soft labels by interpolating between explicitly labeled extreme classes.
- Conclusions and Future Works: Laboratory-scale swirl-stabilized combustor data validate the framework using a high-fidelity similarity metric for closeness to ground-truth unstable flames.The same measure is also used for the neighborhood implicit graph-labeling problem.
- Conclusions and Future Works: The framework is presented as an efficient laboratory diagnostic and helps domain experts learn about coherent structures appearing during combustion instabilities.The authors state that large-scale validation is underway to assess wider applicability.
- Conclusions and Future Works: Future research will extend the framework to multiclass implicit labeling, while large-scale validation remains necessary for wider applicability.The current evidence is limited to laboratory experiments.