TL;DR
Optimizing finger seals for aero-engines is a balancing act between preventing leaks (low hysteresis) and reducing wear (low contact pressure). This paper introduces a high-accuracy surrogate model using BP neural networks and a novel Hybrid Teaching-Learning-Based Optimization (HTLBO) algorithm to provide engineers with a suite of optimal design configurations that satisfy both requirements.
What Problem Does This Paper Address?
Finger seals are advanced flexible seals used in high-performance machinery like aircraft engines. While they are more cost-effective than traditional brush seals, they suffer from two conflicting performance issues: hysteresis and wear.
Hysteresis occurs when the "fingers" of the seal fail to spring back quickly after a rotor moves, leading to temporary gaps and leakage. Engineers can increase the stiffness of the fingers to reduce this lag, but doing so increases the contact pressure against the rotor, which accelerates wear and shortens the seal's lifespan. Because the relationship between the seal's complex geometry (like the curvature of the finger beams) and its performance is highly non-linear, traditional mathematical models struggle to find the "sweet spot" for design parameters.
Methods
The authors propose a two-stage approach to solve this multi-objective optimization problem:
- Surrogate Modeling with BP Neural Networks: Instead of running slow, computationally expensive Finite Element Analysis (FEA) for every possible design, the researchers trained a Back-Propagation (BP) neural network. This network acts as a fast "proxy" that maps structural parameters—such as the base circle radius ($r_b$), number of beams ($n$), and finger shoe height ($x_g$)—directly to performance metrics (hysteresis rate and contact pressure).
- Hybrid Teaching-Learning-Based Optimization (HTLBO): The authors modified the standard TLBO algorithm—which mimics a classroom environment where a "teacher" improves the average grade of "students"—by integrating Pareto dominance. This allows the algorithm to handle multiple conflicting objectives simultaneously. Rather than producing a single "best" answer, it generates a Pareto front: a set of optimal solutions where you cannot improve one metric (like leakage) without worsening another (like wear).
Datasets
The study utilized a dataset generated through ANSYS finite element simulation based on a uniform experimental design.
- Training/Testing Data: 50 sets of structural parameters were used to train and validate the BP neural network.
- Validation: The results of the optimization model (MHACP) were compared against standard ANSYS simulations to ensure physical accuracy.
- Benchmarking: The HTLBO algorithm's performance was compared against the Nash-GA (Genetic Algorithm) to prove its superior stability and optimization capability.
Key Insights and Ideas
- The Power of Surrogates: The BP neural network achieved a linear regression fit exceeding 0.99, proving that machine learning can effectively replace complex physical simulations in the design loop without losing accuracy.
- Pareto over Weighting: Traditional optimization often "weights" different goals (e.g., 60% importance to leakage, 40% to wear) to get one answer. By using Pareto dominance, this method provides a range of solutions, allowing engineers to choose a specific design based on whether a particular engine needs to prioritize longevity or airtightness.
- Geometric Sensitivity: The study highlights that parameters like the base circle radius ($r_b$) have a profound impact on the radial stiffness of the beams. A larger $r_b$ reduces the beam's inclination, increasing stiffness and helping with hysteresis but potentially increasing wear.
Key Results
- Accuracy: The proposed MHACP model's simulation results were found to be equivalent to high-fidelity ANSYS results within a 95% confidence interval.
- Optimization Efficiency: The HTLBO algorithm demonstrated better stability and found more "non-dominated" solutions (better trade-offs) than the Nash-GA method.
- Performance Gains: In specific test cases, the optimized designs significantly lowered the hysteresis rate compared to baseline models while maintaining or reducing average contact pressure, directly translating to more efficient and durable seals.
Limitations and Future Work
While the paper provides a robust framework for structural optimization, it primarily focuses on the geometric design of the finger beams. The authors note that:
- Operating Conditions: The current model assumes specific pressure differences and rotor displacements; future work could incorporate a wider range of dynamic operating environments.
- Material Factors: Integrating material fatigue and thermal expansion into the multi-objective model could further refine the accuracy for real-world aero-engine applications.
Summary updated