ML Model for Combustion Instability Detection
Train a neural network to hear when a combustor is about to go unstable
Last reviewed: March 2026Overview
Thermoacoustic combustion instability is one of the most dangerous phenomena in propulsion engineering. When pressure oscillations in a combustion chamber couple with the unsteady heat release from the flame, they can amplify into violent oscillations that destroy hardware in seconds. This failure mode has plagued rocket engine development since the F-1 engines of the Saturn V and remains a critical design challenge for modern engines at SpaceX, Blue Origin, and in gas turbine power plants.
The instability manifests as a characteristic pattern in high-frequency pressure data: the transition from stable combustion (broadband noise) to instability (sharp spectral peaks and growing oscillation amplitude) follows a pattern that is difficult to define with simple threshold rules but is well-suited to deep learning classifiers. In this project you will train either an LSTM (Long Short-Term Memory) network or a 1D convolutional neural network to classify short windows of pressure data as stable, transitioning, or unstable.
You will work with either experimental combustion pressure data from published research or synthetic data generated from a thermoacoustic model (a Rijke tube simulator). The trained classifier could serve as a real-time monitoring system — flagging the onset of instability fast enough for a control system to intervene before damage occurs. This is an active area of research with direct industry relevance.
What You'll Learn
- ✓ Understand the physics of thermoacoustic instability and its coupling mechanisms
- ✓ Preprocess high-frequency pressure signals: windowing, normalization, and spectral feature extraction
- ✓ Implement LSTM and 1D-CNN architectures for time-series classification in PyTorch
- ✓ Evaluate classifier performance with precision, recall, and ROC curves on imbalanced temporal data
- ✓ Address the practical challenge of detecting instability onset early enough for control intervention
Step-by-Step Guide
Study Thermoacoustic Instability Physics
Review the Rayleigh criterion: instability occurs when pressure oscillations and heat release fluctuations are in phase, causing net energy addition to the acoustic field. Read Lieuwen and Yang's Combustion Instabilities in Gas Turbine Engines (Chapter 1–3) or the freely available review papers on the topic.
Understand the three regimes you will classify: stable (low-amplitude broadband pressure fluctuations), transitioning (intermittent bursts of oscillation as the system approaches the stability boundary), and unstable (sustained high-amplitude oscillations at one or more acoustic modes). The transitioning regime is the most important to detect — it represents the early warning window.
Acquire or Generate Pressure Data
Option A: Use published experimental datasets. The University of Cambridge combustion dynamics group and TU Munich have published time-series pressure data from laboratory combustors. Search for datasets associated with papers on thermoacoustic intermittency or early warning indicators.
Option B: Generate synthetic data from a Rijke tube model — a simple 1D thermoacoustic system governed by a damped wave equation with a nonlinear heat source. Implement the model in Python (van der Pol oscillator approximation) and vary the heat release parameter to sweep from stable through transitional to unstable. Sample pressure at 10+ kHz to capture the acoustic modes. Generate at least 1,000 windows per class.
Preprocess and Window the Data
Segment the continuous pressure signal into fixed-length windows (e.g., 1024 or 2048 samples). Label each window as stable, transitioning, or unstable based on the RMS pressure amplitude or a manual annotation. Apply z-score normalization within each window.
Extract optional spectral features: compute the short-time Fourier transform (STFT) or the power spectral density for each window and note the dominant frequency and its amplitude. These can serve as additional input channels or as features for a baseline model. Create a train/validation/test split by time — never split randomly, since adjacent windows are highly correlated.
Build the LSTM Classifier
Implement a sequence classifier in PyTorch: an LSTM with 2 layers, 128 hidden units, followed by a fully connected output layer with 3 classes (stable/transitioning/unstable). The input is the raw pressure window as a sequence of length N with 1 feature (pressure amplitude).
Use cross-entropy loss with class weights to handle the class imbalance (stable windows typically outnumber transitional ones by 5:1 or more). Train with Adam optimizer, learning rate 1e-3, and batch size 64. Monitor validation accuracy and loss for overfitting. Alternatively, implement a 1D-CNN with 3 convolutional layers (kernel size 7, stride 2) followed by global average pooling — this architecture is often faster to train and can match LSTM accuracy on this type of data.
Evaluate Detection Performance
The key metric is not just accuracy but how early the model can detect the transition to instability. Compute precision and recall for each class, and plot the ROC curve for the "transitioning" class. A high recall for the transitioning class means the model catches most instability onset events — even at the cost of some false alarms.
Calculate the detection lead time: for each instability event in the test set, find the first window that the model labels as "transitioning" and measure the time gap before full instability develops. A lead time of 50–200 ms is typically needed for a control system to respond (e.g., by adjusting fuel flow or air staging).
Compare LSTM vs. 1D-CNN vs. Baseline
Train a baseline classifier using hand-crafted features: RMS pressure, dominant spectral peak frequency, spectral peak amplitude, and spectral entropy — fed into a Random Forest. Compare this traditional approach against the LSTM and 1D-CNN end-to-end models on the same test set.
Analyze where each approach fails. The hand-crafted features often struggle with the transitional regime where the signal is intermittent. The deep learning models should capture the temporal patterns of intermittent bursting more effectively, but they require more data to train. Document these trade-offs clearly.
Discuss Real-Time Deployment
Consider the practical requirements for deploying this model in a real combustor monitoring system. Estimate the inference latency of your model on a single window (measure it with PyTorch). Can it run faster than real-time at the required sample rate? What hardware (CPU, GPU, FPGA) would be needed?
Discuss domain shift: a model trained on one combustor geometry will not directly transfer to another without retraining or adaptation. What strategies (transfer learning, physics-informed features, domain-adaptive training) could improve generalization? Write up your conclusions as a research-style report suitable for a conference poster or short paper submission.
Career Connection
See how this project connects to real aerospace careers.
Aerospace Engineer →
Combustion instability remains one of the hardest problems in propulsion engineering — ML-based early detection is an active research investment at NASA, AFRL, and every major engine manufacturer.
Aerospace Manufacturing →
Engine test stand monitoring uses similar time-series anomaly detection to protect expensive prototype hardware during development testing.
Avionics Technician →
Engine vibration and pressure monitoring systems on aircraft use analogous signal processing and anomaly detection — the same sensor-to-decision pipeline.
Space Operations →
Real-time engine health monitoring during launch is critical for mission safety — instability detection models are directly applicable to launch vehicle engine telemetry.
Go Further
- Implement a Transformer-based classifier (using self-attention on the pressure sequence) and compare against the LSTM and CNN — Transformers often excel at capturing long-range temporal dependencies.
- Add a control loop: connect your classifier to a simple thermoacoustic model and implement closed-loop instability suppression by adjusting the heat release parameter when the model detects transitioning behavior.
- Apply your model to gas turbine data from a different combustor geometry and measure the domain shift — quantify how much retraining is needed to achieve acceptable accuracy.
- Explore physics-informed features: compute the Rayleigh index (time-average of pressure × heat release correlation) and add it as an input feature to improve detection accuracy.