Parameters optimization of a walnut shell-kernel separation device using response surface methodology and ANN coupled genetic algorithm

Liu Hualong, Mao Biqi, Hong Zhang, Peiyu Chen, Yurong Tang, Fan Xiuwen, Yongcheng Zhang, Jianping Wang, Yong ZengInternational Journal of Agricultural and Biological Engineering 2025 · DOI

TL;DR

Researchers have developed a high-efficiency, two-stage tandem device for separating walnut kernels from shells, utilizing a hybrid approach of Computational Fluid Dynamics (CFD), Artificial Neural Networks (ANN), and Genetic Algorithms. The optimized system slashes impurity rates by up to 77% compared to traditional single-stage devices while maintaining a high processing capacity of 720 kg/h.

What Problem Does This Paper Address?

Walnuts are a high-value crop, but the "shelling" process remains a significant industrial bottleneck. After walnuts are cracked, the resulting mixture of shell fragments and kernels must be separated. Because shells and kernels often have similar shapes, textures, and densities, traditional pneumatic (air-based) separation frequently suffers from two issues: high impurity rates (shells left in the kernel batch) and high loss rates (kernels accidentally blown away with the shells).

While single-stage separation is common, it struggles with the multi-size and multi-component nature of cracked walnuts. This paper tackles the challenge by designing a more complex two-stage tandem device and finding the "Goldilocks" zone of operating parameters—wind speeds and deflector plate angles—where separation efficiency is maximized.

Methods

The researchers employed a multi-disciplinary "stack" to optimize the device:

  1. Hardware Design: A two-stage tandem duct system featuring adjustable deflector plates. These plates are crucial; they prevent materials from clustering along the duct walls due to centrifugal force.
  2. CFD Simulation: Using ANSYS Fluent, the team simulated airflow through various duct curvatures to find the optimal bend radius (determined to be 550 mm) for uniform pressure and velocity distribution.
  3. Modeling (RSM vs. ANN):
    • Response Surface Methodology (RSM): A standard statistical method used to find the relationship between variables (angles and wind speeds) and outcomes (impurity and loss).
  4. Artificial Neural Networks (ANN): A more sophisticated machine learning approach. The team used a 3-layer feed-forward network to capture complex, non-linear interactions that simple polynomials might miss.
  5. Optimization (NSGA-II & TOPSIS): Since you can't minimize impurity without potentially increasing loss (a classic multi-objective trade-off), they used the Non-dominated Sorting Genetic Algorithm-II (NSGA-II) to find a "Pareto front" of optimal solutions. Finally, TOPSIS (Technique for Order Preference by Similarity to an Ideal Solution) was used to select the single best configuration from that front.

Datasets

The study used a primary experimental dataset generated from 'Wen 185' walnuts harvested in Xinjiang, China.

  • Sample Size: 30 experimental runs were conducted based on a Central Composite Design (CCD).
  • Variables: The dataset tracked four inputs: first-stage deflector angle ($x_1$), second-stage deflector angle ($x_2$), first-stage wind speed ($x_3$), and second-stage wind speed ($x_4$).
  • Validation: The models were validated against physical experiments using hot-wire anemometers to ensure the CFD simulations matched reality.

Key Insights and Ideas

  • ANN Outperforms RSM: While both models were accurate, the ANN showed superior predictive ability ($R^2$ of 0.988 vs 0.971 for impurity). This suggests that the physics of turbulent airflow and particle collision in a separation duct are too complex for traditional second-order regression models to fully capture.
  • The "Deflector" Secret Sauce: The study highlights that the deflector plate isn't just a barrier; it’s an airflow redirector. Without it, centrifugal forces cause particles to cluster, leading to uneven separation. The optimal angles (39° and 36°) were found to be the "sweet spot" for dispersing the material into the air stream.
  • Wind Speed Thresholds: The optimal first-stage wind speed was found to be quite high (21 m/s). The authors noted this is equivalent to a "force eight gale," necessary because of the small frontal area and high density of walnut components.

Key Results

The optimization yielded specific parameters: $x_1=39^\circ$, $x_2=36^\circ$, $x_3=21$ m/s, and $x_4=13.5$ m/s. Under these conditions, the system achieved:

  • Impurity Rate: 4.51%
  • Loss Rate: 6.62%
  • Efficiency Gains: Compared to single-stage devices, the impurity rate was reduced by 73.98% to 77.55%, and the loss rate dropped by 9.44% to 53.96%.
  • Throughput: The device handled 720 kg/h, a 33% increase over previous pneumatic designs.

Limitations and Future Work

While the results are impressive, the authors acknowledge a few areas for improvement:

  • Kernel Damage: High wind speeds (21 m/s) might cause physical damage to the kernels upon impact with the duct walls, which wasn't explicitly measured in this study.
  • Energy Constraints: The optimization focused on separation quality but did not include energy consumption as a constraint. Future models could aim for "green" separation by minimizing fan power.
  • Complex Dynamics: The study primarily used a continuous phase model for air. Future work could incorporate more detailed discrete element method (DEM) simulations to model individual particle-to-particle collisions.

Summary updated

Abstract

Walnut shell-kernel separation after cracking is crucial for providing raw materials for further processing. However, impurities and losses during separation limit complete separation. To address this, a two-stage tandem separation device was designed and optimized. Computational fluid dynamics (CFD) was used to analyze the effects of four bent duct structures on the flow field. Response surface methodology (RSM) and artificial neural networks (ANN) were employed to predict separation performance under various conditions. Both models accurately predicted performance, with ANN showing superior predictive ability. The optimal design was determined using non-dominated sorting genetic algorithm-II (NSGA-II) and technique for order preference by similarity to an ideal solution (TOPSIS): the inclination of the first stage deflector plate (x1) was 39°,the inclination of the second stage deflector plate (x2) was 36°, the wind speed of the first stage fan (x3) was 21 m/s, and the wind speed of the second stage fan (x4) was 13.5 m/s; impurity rate (y1) was 4.51%, and loss rate (y2) was 6.62%. Compared with traditional single-stage devices, the optimized device reduced impurity rate by 73.98% to 77.55% and loss rate by 9.44% to 53.96%, significantly improving separation efficiency and quality. This study provides theoretical guidance for designing and optimizing shell-kernel separation devices. Keywords: walnut, two-stage separation device, multi-objective optimization, response surface methodology, artificial neural network DOI: 10.25165/j.ijabe.20251805.9375 Citation: Liu H L, Mao B Q, Zhang H, Chen P Y, Tang Y R, Fan X W, et al. Parameters optimization of a walnut shell-kernel separation device using response surface methodology and ANN coupled genetic algorithm. Int J Agric & Biol Eng, 2025; 18(5): 306–316.
BibTeX
@article{hualong2025parameters,
  author = {Liu Hualong and Mao Biqi and Hong Zhang and Peiyu Chen and Yurong Tang and Fan Xiuwen and Yongcheng Zhang and Jianping Wang and Yong Zeng},
  title = {{Parameters optimization of a walnut shell-kernel separation device using response surface methodology and ANN coupled genetic algorithm}},
  journal = {International Journal of Agricultural and Biological Engineering},
  year = {2025},
  doi = {10.25165/j.ijabe.20251805.9375},
}
Figure