TL;DR
This paper introduces a Data-Driven Evolutionary Optimisation (DDEO) method, leveraging a neural network surrogate model, to significantly reduce the low-frequency discrete-spectrum (DS) force of marine propellers. By drastically cutting down computational time compared to traditional methods, DDEO achieves comparable or better noise reduction results, making it a promising approach for complex engineering design problems.
What Problem Does This Paper Address?
Marine propellers generate significant noise, a critical factor for ship detection and identification via sonar. A key component of this noise is the low-frequency discrete-spectrum (DS) force, which arises from blade rotation in non-uniform flow. Reducing this force is highly desirable but computationally expensive, as traditional methods like Computational Fluid Dynamics (CFD) require extensive simulations for each design iteration. Existing optimization techniques, while effective, are often hampered by the sheer computational cost associated with re-evaluating propeller performance at every step. This paper aims to bridge this gap by developing a more efficient optimization framework.
Methods
The core of this work is the development and application of a Data-Driven Evolutionary Optimisation (DDEO) method. This approach combines a traditional evolutionary optimization algorithm (specifically, NSGA-II-DE, an enhanced version of NSGA-II) with a data-driven surrogate model.
- Surrogate Model: A Back-Propagation (BP) fully connected neural network is trained to approximate the complex simulation model that predicts the propeller's DS force and thrust coefficient. The network's architecture (depth and width of hidden layers) is optimized through empirical analysis to achieve high predictive accuracy.
- Data-Driven Evolutionary Optimisation (DDEO): Instead of directly feeding every design iteration into the computationally expensive simulation model, DDEO uses the trained surrogate model for preliminary evaluation. Only promising candidate designs, identified by the surrogate model as likely to yield non-dominated results, are then passed to the actual simulation model. This "preliminary selection" drastically reduces the number of full simulations required.
- Online Mode Training: The surrogate model is trained incrementally ("online mode") as new simulation data becomes available during the optimization process, allowing it to continuously improve its predictive capabilities.
Datasets
The paper describes the generation of a large dataset for training and evaluating the surrogate model.
- Generated Dataset: A dataset of 336,000 records was generated using the simulation model. Each record consists of four input coefficients defining the propeller's skew distribution and three output values: the first-order DS thrust (DST1), the second-order DS thrust (DST2), and the thrust coefficient (KT).
- Dataset Split: This dataset was split into training (202,000 records), validation (67,000 records), and testing (67,000 records) sets.
- Benchmark: The study uses a standard five-blade high-skew propeller (HSP) from the ship Seiun-Maru as a case study for evaluation.
Key Insights and Ideas
- Surrogate Models as Efficiency Boosters: The paper strongly demonstrates that well-trained neural network surrogate models can dramatically accelerate computationally intensive engineering optimizations. The DDEO framework effectively leverages this by using the surrogate to filter out unpromising designs early on.
- Neural Network Architecture Matters: The precise structure of the neural network (number of layers and nodes) significantly impacts its predictive accuracy. The authors found a specific four-layer architecture (64-128-256-64 nodes) to be optimal for their task.
- Online Training for Better Performance: Employing an "online" training strategy for the surrogate model, where it's incrementally updated, leads to better predictive accuracy and ultimately a superior optimization solution compared to a purely "offline" trained model.
- Skew Distribution is Key: The study confirms that optimizing the propeller's skew distribution is a primary method for reducing DS forces, and the proposed DDEO method is effective in finding optimal skew configurations.
Key Results
- Computational Speedup: The DDEO method achieved a remarkable time reduction, with the total optimization time being only 0.68% of that required by a Traditional Evolutionary Optimisation Method (TEOM, specifically NSGA-II-DE). The DDEO process took approximately 62.14 minutes, while the TEOM took 202.80 minutes.
- Predictive Accuracy: The trained neural network surrogate model achieved high accuracy, with prediction errors of 0.21% for DST1, 5.71% for DST2, and 0.01% for KT on the test dataset.
- Noise Reduction: Both DDEO and TEOM achieved similar optimization results in terms of noise reduction. However, DDEO yielded a slightly smaller optimized DST1 value.
- Pareto Solutions: The Pareto solutions generated by DDEO and TEOM showed comparable distributions, indicating that DDEO effectively explores the solution space.
Limitations and Future Work
- Limitations:
- The paper acknowledges that while the surrogate model significantly reduces computation, it's still an approximation. Excessive errors in the surrogate could potentially mislead the evolutionary process.
- The study relies on empirical design of the neural network structure; a more systematic or automated approach could be explored.
- Future Work:
- The authors plan to conduct physical model tests in a water tunnel to validate the optimization results obtained through their DDEO method.
- They also intend to explore several other model tests in water tunnels to further verify the optimization outcomes.
Summary updated