OPTIMIZATION OF EFFICIENCY PREDICTION PERFORMANCE FOR SPUN PILE USING HYBRID META-HEURISTIC MACHINE LEARNING ALGORITHMS

Abstract

Kajian ini mencadangkan satu rangka kerja pemodelan hibrid bagi meramalkan kecekapan cerucuk berputar dengan menggunakan algoritma pembelajaran mesin yang dioptimumkan melalui teknik metaheuristik berinspirasi semula jadi. Sebanyak 150 rekod ujian Pile Driving Analyzer (PDA) digunakan untuk membangunkan tiga model ramalan iaitu Adaptive Neuro-Fuzzy Inference System (ANFIS), Support Vector Machine (SVM), dan Regression Tree (RT). Setiap model mempunyai struktur fungsi yang berbeza, di mana ANFIS menggunakan fungsi keahlian berbentuk loceng umum, SVM menggunakan kernel fungsi asas radial Gaussian, dan RT menggunakan pembahagian binari berulang. Untuk meningkatkan prestasi model, hiperparameter telah dioptimumkan menggunakan algoritma Sparrow Search (SSA), Whale Optimization Algorithm (WOA), dan Manta Ray Foraging Optimization (MRFO). Penilaian prestasi menunjukkan bahawa model ANFIS yang dioptimumkan dengan SSA mencapai ketepatan tertinggi dengan RMX (RMSE = 0.123, R² = 1.000) dan FMX (RMSE = 0.560, R² = 1.000). Pengesahan telah dijalankan menggunakan data daripada bangunan lain dalam projek yang sama. Ramalan RMX menunjukkan perbezaan peratusan antara 0.0% hingga 3.7%, manakala FMX berada dalam lingkungan 0.1% hingga 6.4%, yang masih diterima untuk aplikasi praktikal. Keputusan ini membuktikan kebolehsesuaian model dalam persekitaran pembinaan yang serupa. Prototaip berasaskan MATLAB ini menawarkan alternatif yang boleh dipercayai dan menjimatkan kos berbanding kaedah ujian cerucuk konvensional dalam kejuruteraan geoteknik. This study proposes a hybrid modeling framework for predicting the efficiency of spun pile foundations by leveraging machine learning algorithms optimized through nature-inspired metaheuristic techniques. Utilizing a dataset comprising 150 Pile Driving Analyzer (PDA) test records, three predictive models were developed, namely Adaptive Neuro-Fuzzy Inference System (ANFIS), Support Vector Machine (SVM), and Regression Tree (RT). Each model incorporates distinct functional structures, with ANFIS using generalized bell-shaped membership functions, SVM employing a Gaussian radial basis function kernel, and RT utilizing recursive binary splitting. To enhance model performance, hyperparameters were optimized using the Sparrow Search Algorithm (SSA), Whale Optimization Algorithm (WOA), and Manta Ray Foraging Optimization (MRFO). Performance evaluation showed that the ANFIS model optimized by SSA achieved the highest predictive accuracy with RMX (RMSE = 0.123, R² = 1.000) and FMX (RMSE = 0.560, R² = 1.000). Validation was conducted using data from another building within the same project. The RMX predictions showed percentage differences ranging from 0.0% to 3.7%, indicating high reliability. For FMX, percentage differences ranged from 0.1% to 6.4%, within acceptable thresholds for practical application. These results demonstrate the model’s robustness in similar construction environments. The MATLAB-based prototype offers a reliable and cost-effective alternative to conventional pile testing methods in geotechnical engineering.