Source-linked AI summary
On Accurate and Reliable Anomaly Detection for Gas Turbine Combustors: A Deep Learning Approach
Weizhong Yan, Lijie Yu
TL;DR
Gas turbine combustor anomaly detection requires earlier and more reliable identification of faults, but existing approaches provide inadequate performance and handcrafted features are difficult to scale. The paper uses unsupervised deep feature learning followed by neural-network classification, and reports improved detection performance on real-world combustion-system data. The authors identify broader real-world validation and alternative deep-learning methods as future work.
Problem
Existing combustor anomaly-detection solutions have inadequate performance, while handcrafted feature generation is manual, problem-specific, and unscalable.
Method
The paper uses unsupervised deep learning to hierarchically learn features from exhaust-temperature measurements, then feeds them to a neural-network classifier.
Results
Deep feature learning significantly improved combustor anomaly-detection performance over handcrafted features; at a 1% false-positive rate, sensitivity was approximately 0.99 ± 0.01 versus 0.96 ± 0.02.
Takeaways & Limitations
The study demonstrates that deep feature learning can generate effective features from raw thermocouple measurements for combustor anomaly detection.
Takeaways & Limitations
The authors call for more thorough studies using additional real-world data and for exploring deep-learning methods beyond SDAE.
Abstract
from arXiv · showhide
Monitoring gas turbine combustors health, in particular, early detecting abnormal behaviors and incipient faults, is critical in ensuring gas turbines operating efficiently and in preventing costly unplanned maintenance. One popular means of detecting combustor abnormalities is through continuously monitoring exhaust gas temperature profiles. Over the years many anomaly detection technologies have been explored for detecting combustor faults, however, the performance (detection rate) of anomaly detection solutions fielded is still inadequate. Advanced technologies that can improve detection performance are in great need. Aiming for improving anomaly detection performance, in this paper we introduce recently-developed deep learning (DL) in machine learning into the combustors anomaly detection application. Specifically, we use deep learning to hierarchically learn features from the sensor measurements of exhaust gas temperatures. And we then use the learned features as the input to a neural network classifier for performing combustor anomaly detection. Since such deep learned features potentially better capture complex relations among all sensor measurements and the underlying combustor behavior than handcrafted features do, we expect the learned features can lead to a more accurate and robust anomaly detection. Using the data collected from a real-world gas turbine combustion system, we demonstrated that the proposed deep learning based anomaly detection significantly indeed improved combustor anomaly detection performance.
1. INTRODUCTION
Gas turbine combustor anomaly detection is difficult because operating conditions vary widely, while existing rule-based and machine-learning approaches have inadequate or insufficiently improved detection performance. The paper therefore explores unsupervised deep feature learning to automate feature generation from exhaust-temperature measurements.
- Early detection of combustor abnormalities matters because missed faults can cause turbine trips, component-life damage, and higher NOx and CO emissions.
- Detection is challenging because combustor behavior depends on machine type, fuel, ambient conditions, and equipment aging.
- Exhaust-temperature profiles are widely monitored because they provide information about gas-turbine and combustor thermal performance.
- Knowledge-based rules have inadequate detection performance and require laborious design, while conventional machine-learning models have delivered only limited improvement.
- Handcrafted features can outperform raw exhaust temperatures but are manual, problem-specific, and unscalable, motivating automated feature generation through deep learning.
- The paper investigates unsupervised feature learning because industrial PHM provides abundant unlabeled data, whereas accurate labels are costly and often impossible to obtain.
2. RELATED WORK
Prior work spans anomaly detection across PHM applications and combustor-specific studies, but feature engineering remains a central, difficult, application-dependent task. Representation learning offers an alternative by transforming raw inputs into task-useful features through learned abstractions.
- 2.1. Anomaly detection: Anomaly detection has been applied broadly in PHM, including aircraft engines, wind turbines, locomotive engines, marine gas turbines, and combined-cycle power plants.
- 2.1. Anomaly detection: Combustor anomaly studies have used symbolic time-series analysis, failure precursors, traditional neural networks, and logistic regression, with some work directly classifying exhaust-temperature profiles.
- 2.2. Feature engineering: Feature engineering transforms raw data into representations that better capture the underlying problem and can improve predictive accuracy on unseen data.
- 2.2. Feature engineering: Feature engineering is important but challenging because it is manual, problem-specific, and typically requires collaboration between machine-learning and domain experts.
- 2.3. Feature (representation) learning: Feature learning, or representation learning, learns a transformation of raw inputs into representations that can be exploited by machine-learning tasks.
- 2.3. Feature (representation) learning: Unsupervised representation learning uses unlabeled data and includes methods such as PCA, ICA, k-means, autoencoders, RBMs, and sparse coding.
3. METHODOLOGY
The methodology learns robust features without class labels using stacked denoising autoencoders, then supplies those features to an extreme learning machine classifier. The pipeline combines reconstruction-based representation learning with fast supervised classification.
- 3. METHODOLOGY: The proposed pipeline uses unsupervised representation learning to derive features from raw data and feeds those features into a separate supervised classifier.
- 3.1. SDAE for unsupervised feature learning: Stacked denoising autoencoders are selected because their denoising training criterion is intended to produce features more robust to input noise and useful for classification.
- 3.1. SDAE for unsupervised feature learning: The method stacks denoising autoencoders into an SDAE structure and uses the resulting learned features as inputs to an ELM classifier.
- 3.1. SDAE for unsupervised feature learning: An autoencoder learns parameters by minimizing reconstruction error over training examples, with regularization used to avoid learning a noninformative identity mapping.
- 3.1. SDAE for unsupervised feature learning: Denoising autoencoders corrupt inputs during encoding but reconstruct the clean inputs, using corruption as a training criterion for higher-level representations.
- 3.2. ELM for classification: ELM fixes randomly generated input-to-hidden connections, so training only solves for hidden-to-output connections through a linear least-squares problem.
4. CASE STUDY AND RESULTS
The case study evaluates unsupervised deep feature learning for combustor anomaly detection using real turbine data and compares it with handcrafted features. Across repeated cross-validation, learned features provide higher and less variable ROC performance.
- 4.1. The business application: The study uses several months of data from one heavy-duty industrial turbine, including 13,791 event-free samples for unsupervised feature learning.The remaining event and event-free data are used to train and test the classifier after filtering bad and part-load observations.
- 4.2. Deep feature learning: A two-layer stacked denoising autoencoder learns 12 features, which are then supplied to an extreme learning machine classifier.The autoencoder configuration includes 30 hidden neurons in the first layer and 12 in the second; the classifier uses a large hidden layer and sample weighting for class imbalance.
- 4.3. Handcrafted features: The comparison holds the ELM classifier setting fixed and changes only the feature set, using ROC curves and repeated 5-fold cross-validation.Handcrafted features are primarily simple statistics computed from thermocouple profiles.
- 4.4. Results: Deep learned features produce significantly better classification performance and smaller ROC variation than handcrafted features.The learned representations encode patterns in TC profiles rather than individual numerical statistics.
5. CONCLUSION
The conclusion reports that unsupervised deep feature learning can generate useful features from raw exhaust-temperature measurements for combustor anomaly detection. The authors identify broader validation and exploration of other deep learning methods as future work.
- Applying SDAE to raw time-series TC measurements effectively generates features and improves combustor anomaly detection.
- The authors propose more thorough studies with additional real-world data and investigation of deep learning methods beyond SDAE.
BIOGRAPHIES
The biographies profile two GE researchers whose work spans machine learning, signal processing, information fusion, and equipment health management.
- Weizhong Yan is a GE Principal Scientist whose research includes neural networks, big data analytics, feature engineering, feature learning, ensemble learning, and time series forecasting.
- Yan obtained his PhD from Rensselaer Polytechnic Institute and has been with GE since 1998.
- Lijie Yu is a GE senior analytics engineer specialized in signal processing, machine learning, information fusion, and equipment health management.