TL;DR
This paper introduces GCMLP, a novel deep learning framework for diagnosing faults in liquid rocket engines (LREs). By leveraging Granger Causality (GC) to extract interpretable features from component-wise analysis, GCMLP achieves high diagnostic accuracy while providing insights into fault propagation mechanisms, addressing the "black-box" problem of traditional deep learning methods.
What Problem Does This Paper Address?
Liquid rocket engines (LREs) are critical for space missions, but their inherent complexity and extreme operating conditions make them prone to failures. Accurate and timely fault diagnosis is paramount for mission success and safety. While current deep learning methods offer powerful diagnostic capabilities, they often act as "black boxes," lacking interpretability. This makes it difficult to understand why a fault is detected or how it propagates through the system, which is crucial for LREs where reliability is paramount. Existing model-driven approaches struggle with the complexity of creating accurate LRE models, and knowledge-driven methods face challenges in acquiring and maintaining expert knowledge. This paper aims to bridge this gap by developing a diagnostic method that is both accurate and interpretable, using Granger Causality to understand the causal relationships between engine variables during fault conditions.
Methods
The core innovation of this paper is the GCMLP (Granger Causality-based Component-wise Multi-Layer Perceptron) framework. It combines an encoder-decoder architecture with Granger Causality (GC) for fault diagnosis.
Here's a breakdown of the key components:
- Component-wise GC Extraction (Encoder): Instead of treating all variables as a single monolithic block, GCMLP uses separate Multi-Layer Perceptrons (MLPs) for each LRE variable. This "component-wise" approach aims to isolate the causal relationships (GC) between variables without being overly influenced by the coupling of all variables simultaneously. These extracted GC relationships are then represented as GC matrices.
- Predictor (Decoder): This module uses the extracted GC information to predict future sensor readings. The accuracy of these predictions serves as a feedback mechanism to refine the GC extraction process, ensuring the GC matrices accurately reflect inter-variable interactions.
- Fault Classifier: The final module takes the GC features extracted by the encoder as input to classify the engine's state (normal or specific fault). The classification errors also help optimize the GC extractor, pushing it to generate GC matrices that are more discriminative for different fault types.
- Sparsity Constraints: To handle the high dimensionality of LRE data and enforce realistic causal structures, various sparsity penalties (like group lasso) are applied to the GC matrix. This encourages a sparse representation, meaning only the most significant causal links are retained.
Datasets
The study utilizes a simulated dataset generated from a Simulink model of a specific LRE type. This model incorporates various subsystems like the oxidizer supply, valve control, and turbopump assembly.
Faults were simulated using fault injection techniques, introducing six critical fault modes:
- Main turbine fault
- Oxidizer pump fault
- Stage-1 fuel pump fault
- Thrust chamber throat ablation
- Oxidizer booster pump fault
- Fuel booster pump fault
Each fault was simulated at two severity levels, resulting in 12 distinct datasets. The data consists of 10 sensor channels sampled at 1 kHz over 5-second durations. For training and evaluation, data from the 0.5-1s interval was used for normal states, and the post-injection transition period (2.5-3s) was used for fault states. The raw data was segmented into windows of length 10 with a step size of 5, yielding a total of 1,274 samples.
Key Insights and Ideas
- Granger Causality for Interpretability: The paper highlights GC as a powerful tool to move beyond "black-box" deep learning. By focusing on predictive relationships between time series, GC provides a framework for understanding how faults propagate through the LRE system, offering a level of interpretability often missing in other data-driven methods.
- Component-wise Analysis is Key: The authors argue that treating LRE variables in a component-wise manner is crucial due to the complex interdependencies. This approach allows for more focused extraction of causal relationships, preventing the overwhelming coupling of all variables from obscuring important fault-specific patterns.
- Predictive Accuracy Validates Causality: The fact that the predictor's accuracy improves when using GC-derived features validates the correctness and relevance of the extracted causal relationships. This creates a feedback loop where predictive performance reinforces the interpretability of the diagnostic features.
- Sparsity Enhances Realism and Performance: The application of sparsity constraints (like group lasso) is not just for computational efficiency; it reflects the inherent sparse nature of causal relationships in complex systems. The study shows that more advanced sparsity techniques (hierarchical and group sparse lasso) lead to better fault localization accuracy, suggesting that capturing multi-dimensional sparsity is important.
Key Results
The GCMLP method demonstrated superior fault localization accuracy compared to several benchmark models, including CNN, GRU, BiGRU, LSTM, BiLSTM, and MLP.
- Overall Accuracy: GCMLP achieved an accuracy of 95.50%, outperforming all other models.
- Precision and Recall: It also achieved high precision (93.03%) and recall (94.71%).
- Specific Fault Performance: Notably, GCMLP showed exceptional performance in identifying healthy states and the fuel booster pump fault (Label 6), achieving 100% accuracy for the latter, where other models struggled significantly.
- Interpretability: The paper provides a case study on thrust chamber throat ablation faults, showing how the extracted GC relationships align with physical understanding of the engine's operation and fault propagation, thus validating the interpretability of the method.
Limitations and Future Work
- Simulation-Based Data: The current study relies on simulated LRE data. While robust, the authors acknowledge that real-world data may present additional complexities and noise.
- Sparsity Penalty Selection: While the paper explores different sparsity penalties and identifies hierarchical and group sparse lasso as more beneficial, the optimal selection and tuning of these penalties for various fault scenarios could be a direction for future work.
- Real-World Validation: The authors suggest that future work should focus on validating the GCMLP framework on real-world LRE operational data to further confirm its efficacy and robustness in practical applications.
- Expanding Fault Scenarios: While six fault types were simulated, exploring a wider range of fault modes and their complex interactions could enhance the generalizability of the method.
Summary updated