⚡ electrical engineering

A back propagation neural network with improved sine cosine algorithm approach for multi-response drilling optimization of carbon fiber reinforced polymer (CFRP)

This study proposes an integrated approach combining a Back Propagation Neural Network with an Improved Sine-Cosine Algorithm to predict and optimize multi-response drilling parameters for carbon fiber reinforced polymers, utilizing adaptive operator selection and population management to enhance performance over standard methods.

Original authors: Rachmadi Norcahyo, Achmad Pratama Rifai, Setyo Tri Windras Mara, Bobby Oedy Pramoedyo Soepangkat, Mohammad Khoirul Effendi

Published 2026-09-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rachmadi Norcahyo, Achmad Pratama Rifai, Setyo Tri Windras Mara, Bobby Oedy Pramoedyo Soepangkat, Mohammad Khoirul Effendi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of aerospace, where every ounce of weight matters, engineers rely on a material that is as strong as steel but light as aluminum: carbon fiber reinforced polymer. This composite material, made of incredibly thin carbon strands bound together by plastic resin, forms the backbone of modern aircraft wings and fuselages. However, turning these rigid sheets into usable parts requires drilling thousands of holes to bolt the structure together. This seemingly simple task is fraught with danger for the material. Because the fibers are brittle and the binding resin is soft, the heat and pressure of a standard drill bit can cause the layers to peel apart or crack, a damage known as delamination. Once this damage occurs, the structural integrity of the entire component is compromised, potentially leading to catastrophic failure. The challenge for manufacturers is to find the perfect combination of drill speed, feed rate, and tool shape to make clean holes without ruining the expensive material, a balancing act that has long required trial and error.

To solve this puzzle, a team of researchers from Indonesia and Australia has developed a new digital strategy that combines two powerful tools: a learning computer program and a smart search algorithm. They began by conducting a series of real-world drilling experiments on carbon fiber plates. Using a specialized machine, they drilled holes with two different types of drill bits—a standard twist drill and a brad spur drill—while systematically changing the speed of the drill and how fast it moved into the material. For every single hole, they measured four critical outcomes: the force pushing down on the drill, the twisting torque required to turn it, and the size of the cracks that appeared at both the entry and exit points of the hole. They also tracked how quickly the material was being removed. These physical measurements provided a rich dataset that revealed exactly how different settings affected the quality of the hole.

The researchers then fed this experimental data into a back propagation neural network, a type of computer program designed to learn from examples much like a human student learns from a textbook. By analyzing the relationship between the drilling settings and the resulting damage, the computer built a highly accurate predictive model. This model could look at a new set of settings and instantly predict the likely force, torque, and damage levels without needing to drill a physical hole. However, knowing how to predict the outcome is only half the battle; the real goal was to find the single best set of settings that would minimize damage while maximizing efficiency. To do this, the team employed an improved version of a search method called the sine-cosine algorithm. Imagine a vast landscape of possible settings, where some points are safe and others are dangerous. The algorithm acts as a guide, sending out a swarm of virtual explorers to search this landscape for the lowest point, which represents the perfect drilling conditions.

The team's innovation lay in how they improved this search guide. Standard search algorithms sometimes get stuck in local low points, thinking they have found the best solution when a better one exists nearby. To prevent this, the researchers added two new rules to their algorithm. First, they gave the algorithm the ability to adaptively choose its search strategy based on what worked best in previous steps, essentially learning from its own history to make smarter moves. Second, they introduced a population management technique that ensured the virtual explorers did not all crowd into the same spot too quickly. This kept the search diverse, allowing the algorithm to explore different areas of the solution space thoroughly before settling on a final answer. When they tested this improved system against the standard version, the new approach consistently found better solutions, identifying a specific combination of a twist drill bit, a spindle speed of 3000 revolutions per minute, and a feed rate of 160.1 millimeters per minute as the optimal configuration.

The results of this study offer a clear path forward for manufacturing. The researchers found that using a twist drill bit generally produced less damage at the exit of the hole compared to the brad spur drill, while higher spindle speeds helped soften the material slightly, reducing the force needed and the resulting cracks. Conversely, increasing the speed at which the drill moved into the material increased the force and the damage. By combining their predictive computer model with the improved search algorithm, the team demonstrated that it is possible to calculate the precise drilling parameters needed to produce high-quality holes with minimal waste. This work does not just suggest a theoretical possibility; it provides a concrete, tested method for optimizing a critical industrial process. As the aerospace industry continues to push for lighter and stronger aircraft, such data-driven approaches will be essential for ensuring that the holes holding these advanced structures together are made with the utmost precision and care.

Technical Summary: A Back Propagation Neural Network with Improved Sine Cosine Algorithm Approach for Multi-Response Drilling Optimization of CFRP

Problem Statement
Carbon Fiber Reinforced Polymer (CFRP) is widely utilized in aerospace and aviation due to its high strength-to-weight ratio and durability. However, the drilling process required for assembly often induces severe machining damages, including fiber-matrix debonding, fiber pull-out, and delamination. These defects, particularly hole entry and exit delamination, significantly deteriorate hole quality and component longevity. The study identifies that drilling parameters—specifically drill bit geometry, spindle speed, and feed rate—are the primary contributors to these damages, alongside the resulting thrust force and torque. While previous studies have investigated these factors, there is a need for robust, data-driven approaches to simultaneously model and optimize multiple conflicting responses (thrust force, torque, and delamination) to minimize damage while maintaining material removal rates.

Methodology
The research proposes a hybrid framework integrating a Back Propagation Neural Network (BPNN) with an Improved Sine-Cosine Algorithm (ISCA).

Experimental Data Collection:
Drilling experiments were conducted on 3mm thick CFRP specimens using a CNC machine without cutting fluid. Two drill bit geometries were tested: a two-edge twist drill and a brad spur drill. The experimental design varied spindle speeds (1000, 2000, 3000 rpm) and feed rates (50, 150, 250 mm/min). Responses measured included thrust force ( $F_{Th}$ ), torque ( $M$ ), hole entry delamination ( $F_{Den}$ ), hole exit delamination ( $F_{Dex}$ ), and material removal rate (MRR).
BPNN Modeling:
A BPNN was developed to predict the drilling responses based on the input parameters. The network architecture was optimized using Ant Colony Optimization (ACO) to determine the optimal number of hidden layers and neurons, resulting in a 3-4-7-10-5 configuration (3 inputs, 3 hidden layers with 4, 7, and 10 neurons respectively, and 5 outputs). The model utilized tansig activation functions for hidden layers and purelin for the output layer, trained with the Levenberg-Marquardt algorithm. Data was normalized using a mapminmax function, and the dataset was split into 70% training, 15% validation, and 15% testing.
Improved Sine-Cosine Algorithm (ISCA):
To optimize the BPNN model and find the best parameter combination, the standard Sine-Cosine Algorithm (SCA) was enhanced with two specific mechanisms:
- Adaptive Operator Selection: Unlike the standard SCA which randomly selects between sine and cosine functions with a fixed probability ( $P_y = 0.5$ ), the ISCA dynamically updates the selection probability based on the historical performance of each operator. Weights are adjusted using a decay parameter and fitness-based rewards to favor operators that yield better solutions.
- Population Management Technique: To prevent premature convergence and maintain solution diversity, a population management technique (adapted from Prodhon and Prins, 2008) was implemented. New solutions are accepted into the population only if their distance from existing solutions exceeds a dynamic threshold ( $\Delta$ ). This threshold is adjusted based on the number of successive rejections, ensuring a balance between exploration and exploitation.

Key Contributions
The paper outlines four primary contributions:

Algorithmic Improvement: It introduces the first combination of adaptive operator selection and population management techniques within the Sine-Cosine Algorithm framework.
Application to Machining: It implements an SCA-based heuristic for machining optimization, a domain where such applications are currently rare.
Drilling Parameter Optimization: It presents the first study to utilize an SCA-based algorithm specifically for optimizing CFRP drilling parameters.
Multi-Response Integration: It successfully applies the ISCA to optimize a multi-response BPNN model, simultaneously addressing thrust force, torque, delamination, and material removal rate.

Results

Model Performance: The BPNN demonstrated high predictive accuracy, with Mean Average Percentage Error (MAPE) values for training data ranging from 0% (MRR) to 3.39% (Thrust Force). Correlation coefficients indicated a strong fit between experimental and predicted values.
Optimization Outcomes: The ISCA outperformed both the standard SCA and an Adaptive SCA (ASCA) variant. The optimal solution identified by the BPNN-ISCA approach utilized a twist drill bit, a spindle speed of 3000 rpm, and a feed rate of 160.1 mm/min.
Parameter Effects: The study confirmed that higher spindle speeds generally reduce thrust force and torque due to material softening, while lower feed rates reduce delamination. Regarding drill geometry, the brad spur drill reduced hole entry delamination but increased thrust force compared to the twist drill, whereas the twist drill resulted in less hole exit delamination.

Significance and Claims
The authors claim that the proposed ISCA approach clearly improves the performance of the standard BPNN-SCA framework. By integrating adaptive mechanisms and population management, the algorithm effectively navigates the complex, multi-response optimization landscape of CFRP drilling. The study emphasizes that while previous works have hybridized SCA with other metaheuristics or introduced new equations, this work demonstrates that simple mechanisms—such as updating selection probabilities based on performance and managing population diversity—can significantly enhance the exploitation capabilities of the classical SCA without requiring complex re-derivation of trigonometric update equations.

The paper concludes modestly, suggesting that while the current BPNN architecture was optimized using ACO, future research could compare ACO with other state-of-the-art techniques for network architecture selection. Furthermore, the authors note a gap in the literature regarding the simultaneous optimization of conflicting multiple objectives in machining tasks, positioning this study as a step toward filling that gap.

More like this