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 can reduce gas-turbine efficiency and longevity, yet detecting it before full-blown instability is difficult. This paper applies an end-to-end convolutional selective autoencoder to high-speed flame video; it suppresses stable-region frames, reveals unstable-region frames, and partially reveals intermittency before full-blown instability.
Problem
Combustion instability reduces aircraft gas-turbine efficiency and longevity, and detecting it before full-blown instability is extremely difficult.
Source: §1 · p. 2, §2.2 · p. 3Method
Results
Stable-region frames are suppressed, unstable-region frames are visible, and intermittency is partially revealed and localized before full-blown instability, with greater prominence than other state-of-the-art approaches.
Source: §5 · pp. 7–8Takeaways & Limitations
The framework is presented as an early-detection diagnostic for combustion instability, with validation on laboratory-scale swirl-stabilized combustor data.
Source: §6 · p. 8
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 introduces a deep convolutional selective autoencoder for detecting combustion instability early from high-speed flame videos. It learns implicit labels for intermediate transition states from explicitly labeled stable and unstable frames while extracting flame structures for interpretation.
1. INTRODUCTION
- The proposed selective autoencoder detects instability onset from high-speed flame videos labeled only as stable or unstable, addressing the difficulty of recognizing a fast transition early.Extreme-state labels are used to derive soft labels for intermediate frames.Source: §1 · pp. 1–2
- Selective masking lets the model derive implicit soft labels from explicitly labeled extreme states when intermediate transition-state labels are unavailable.The approach is intended for tracking the continuous transition from combustion stability to instability.Source: §1 · pp. 1–2
- Prior work explored Deep Belief Networks for visualizing intermediate combustion-instability features and later applied a modular neural-symbolic approach.The authors describe these studies as antecedents to exploring machine-learning techniques for early detection.Source: §1 · p. 2
- The framework targets domain-interpretable coherent flame structures without extensive expert-guided feature handcrafting or reliance on a pure black-box model.Full-blown instability is associated with high-amplitude oscillations and coherent, mushroom-shaped flame structures.Source: §1 · p. 2
- A configurable-granularity metric tracks instability onset and pre-transition intermittence, including millisecond-scale bursts with small, partially observable coherent structures.The framework also aims to determine the degree of relationship between temporally close observations in high-dimensional data such as high-speed video.Source: §1 · p. 2
- The paper presents extensive validation using laboratory-scale combustion data collected under realistic operating conditions.Source: §1 · p. 2
2. BACKGROUND
The background introduces convolutional networks, combustion instability, and implicit labeling as context for the paper’s approach.
2.1 Convolutional networks
- Convolutional networks use local nonlinear mappings and shared kernels to extract features; invariance-rich codes and fully connected layers support dimensionality reduction and classification.Feature maps preserve local patterns, while distributed computing and GPUs can support efficient learning.Source: §2.1 · pp. 2–3
2.2 The problem of combustion instability
- Combustion instability couples heat-release and pressure oscillations, producing high-amplitude flame oscillations and coherent structures; POD and DMD are common detection methods.Instability reduces aircraft gas-turbine engine efficiency and longevity; coherent structures can drive large-scale velocity and flame-shape oscillations.Source: §2.2 · p. 3
2.3 Implicit labeling
- Implicit labeling derives soft labels from explicitly labeled extreme classes, combining prior knowledge with sequence labeling to identify transition phases and degrees of class membership.The paper fuses this labeling approach with a convolutional autoencoder; its dog–wolf example illustrates intermediate class membership.Source: §2.3 · p. 3
3. CONVOLUTIONAL SELECTIVE AUTOENCODER
The method trains a convolutional autoencoder on selectively masked image targets to learn flame features through reconstruction. It evaluates instability using the correlation ratio between input and inferred output frames.
3. CONVOLUTIONAL SELECTIVE AUTOENCODER
- Frames are converted into input-target pairs by masking the undesired class with black pixels, then normalized to zero mean and unit standard deviation.The masked frames serve as explicit ground-truth targets.Source: §3 · p. 3
- The encoder uses convolution, pooling, and a bottleneck to extract features, while the decoder reconstructs output frames through deconvolution and upsampling.The encoding and decoding layers use ReLU nonlinearities.Source: §3 · pp. 3–4, Figure 3 · p. 4
- Training minimizes reconstruction error with regularization to avoid overfitting, using Nesterov momentum-based stochastic gradient descent for updates.The authors report improved results compared with ADAGRAD and ADADELTA for reconstruction-error updates.Source: §3 · p. 4
3.1 Instability measure
- The correlation ratio measures instability by comparing input and inferred-output frames, with low computational requirements and tolerance for differing image intensities.The measure does not directly include the actual images in its computation.Source: §3.1 · p. 5
4. DATASET AND IMPLEMENTATION
The study collected flame videos from a laboratory combustor and trained a convolutional model on labeled grayscale frames resized for computation. Five transition videos tested the trained architecture as combustion moved from stable to unstable conditions.
Dataset collection and Experimental setup
- The combustor dataset captured 9,000 frames over 3 seconds across four instability conditions at two fuel-air flow settings and both premixing levels.The reported settings were 45 lpm fuel and 900 lpm air, and 28 lpm fuel and 600 lpm air; Figure 5 shows the apparatus and an unstable-flame sequence.Source: §4 · p. 5, Figure 5 · p. 6
- Five 7-second transition videos tested stable-to-unstable progression through intermittence, induced by reducing fuel flow or increasing air flow.The five conditions were AFR 500 with FFR 40 to 28 or 30; FFR 40 with AFR 500 to 600; AFR 600 with FFR 50 to 35; and FFR 50 with AFR 700 to 800 lpm.Source: §4 · p. 5
Training process
- Training used 63,000 grayscale frames resized from 100 × 237 to 64 × 64, with 35,000 labeled stable and 28,000 unstable.The frames combined two premixing lengths and a range of air and fuel flow rates.Source: §4 · p. 5
- Training used 3 × 3 filters, non-overlapping 2 × 2 max pooling, batches of 128, and Theano’s LeNet 5 early-stopping algorithm.The model used Nesterov stochastic gradient descent with learning rate 0.0001 and momentum 0.975, and was trained for 100 epochs with ℓ2 regularization of 0.0001 for each parameter group and ℓ1 sparsity regularization.Source: §4 · pp. 5–6
- 5,090,249 learnable parameters formed the trained network, and generating results for a 21,841-frame transition sequence took approximately 35.5 seconds.The training-loss and validation-loss profiles are shown in Figure 4; training loss was raised by lowering the regularization parameter to permit more epochs and further reduce validation loss.Source: Figure 4 · p. 5, §4 · p. 6
5. RESULTS AND DISCUSSIONS
The results show that the selective autoencoder can expose intermittency in flame-video transitions before full combustion instability, while its average instability measure can support soft frame labeling. Its feature maps reveal different responses to stable and unstable flame patterns.
Early Detection of combustion instability
- The trained model reproduces explicit labels by assigning stable-region frames label 0 and retaining unstable-region frames during training.Figure 8 illustrates the pipeline's ability to reproduce those labels.Source: §5 · p. 6, Figure 8 · p. 7
- Unstable-region feature maps activate more strongly on characteristic mushroom structures, while stable-region information rapidly diffuses through hidden layers.Fully connected layers reduce image dimensions to explanatory features and reshape them to restore feature maps and outputs toward input dimensions.Source: §5 · p. 6, Figure 9 · p. 7
- Transition-protocol outputs are evaluated with a similarity measure and locally regression-smoothed weighted moving averages to visualize trends.The test examines masking and revealing flame-image features across different transition conditions.Source: §5 · p. 7
- The framework partially reveals unstable features during intermittency, localizing it before full-blown instability more prominently than other state-of-the-art approaches.The authors report that better intermittency tracking leads to early detection with fewer false alarms.Source: §5 · pp. 7–8, Figure 7 · p. 7
Frame labeling
- A smoothed average instability measure can simplify frames around the abrupt stable-to-unstable transition into a linear plot for inferring likely neighboring frames.The authors note that smoothing removes transients but that combustor dynamics may limit this labeling approach for generic applications; implicit labels could support later HMM or STSA learning.Source: §5 · p. 8
- In static applications, soft labels can fix frames back to their original positions; frame-to-frame growth of two “mushroom-like” lobes and increasing average instability measure indicate the framework’s graduated transitioning ability.Source: §5 · p. 8
6. CONCLUSIONS AND FUTURE WORKS
The convolutional selective autoencoder generates fuzzy labels from hard-labeled examples and is applied to early detection of combustion instabilities in high-speed flame video. Laboratory-scale validation is reported, while broader validation and extensions remain future work.
6. CONCLUSIONS AND FUTURE WORKS
- The framework detects combustion instabilities from high-speed flame video, with validation on data from a laboratory-scale swirl-stabilized combustor assessed using a high-fidelity similarity metric against ground-truth unstable flames.It generates fuzzy labels from prior knowledge of hard-labeled examples.Source: §6 · p. 8
- Using the same similarity measure, the architecture is extended to neighborhood implicit graph labeling and can generalize to high-dimensional data by interpolating explicit labels into soft labels.Source: §6 · p. 8
- Large-scale validation is underway to demonstrate wider applicability; future work includes multi-class implicit labeling, while the framework enables domain experts to study coherent structures in combustion instabilities.Source: §6 · p. 8