Analisis Pengaruh Filtering Sinyal Elektrokimia terhadap Akurasi Model Machine Learning dan Deep Learning untuk Klasifikasi Logam Berat

Kamilah, Alya (2026) Analisis Pengaruh Filtering Sinyal Elektrokimia terhadap Akurasi Model Machine Learning dan Deep Learning untuk Klasifikasi Logam Berat. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241066_MASTER_THESIS.pdf Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Penelitian ini mengevaluasi tujuh metode filtering yang terdiri dari kategori konvensional (Savitzky-Golay, Butterworth Lowpass, Moving Median, DWT db4), adaptif (Kalman Filter, Adaptive DWT 6-Level), dan dekomposisi (Adaptive VMD), serta kondisi tanpa filter sebagai baseline. Lima model klasifikasi diuji meliputi Random Forest, XGBoost, SVM, KNN, dan Deep Neural Network (DNN). Eksperimen dilakukan pada dataset 1.500 sampel sinyal CV yang diaugmentasi dengan noise sintetis berbasis model elektroda Ag/AgCl pada SNR 15 dB, menghasilkan 10.500 sampel untuk evaluasi. Sebanyak 60 fitur elektrokimia diekstraksi mencakup fitur statistik, puncak CV, frekuensi, dan wavelet. Hasil penelitian menunjukkan bahwa kombinasi Kalman Filter dengan XGBoost mencapai akurasi tertinggi sebesar 97,52%. Secara rata-rata, Butterworth Lowpass Filter memberikan peningkatan akurasi terbaik (+2,13% dari baseline), diikuti Kalman Filter (+1,76%) dan Moving Median (+1,48%). Filter konvensional secara konsisten mengungguli filter advanced untuk sinyal CV. Sebaliknya, Adaptive VMD justru menurunkan akurasi sebesar 0,73% dari baseline karena ketidaksesuaian asumsi sinyal. Dari sisi model, XGBoost mendominasi dengan rata-rata akurasi 96,56% pada 7 dari 8 konfigurasi filter, sedangkan DNN menunjukkan overfitting gap terendah (0,56%) berkat teknik regularisasi yang efektif. KNN memiliki performa terendah (88,50%) akibat curse of dimensionality pada fitur berdimensi tinggi. Penelitian ini memberikan kontribusi berupa studi ablasi komprehensif pertama yang mengevaluasi 40 kombinasi filter-model untuk klasifikasi logam berat elektrokimia, serta rekomendasi pipeline optimal berdasarkan skenario aplikasi yang dapat langsung diimplementasikan dalam sistem deteksi logam berat berbasis machine learning.
=======================================================================================================================
Heavy metal contamination such as cadmium (Cd) and lead (Pb) poses serious threats to human health and the environment due to their toxic, persistent, and bioaccumulative properties. Heavy metal detection using Cyclic Voltammetry (CV)-based electrochemical methods offers a cheaper, portable, and faster alternative compared to conventional methods such as AAS and ICP-MS. However, electrochemical signals are susceptible to various types of noise that can reduce classification accuracy. This study aims to analyze the influence of various filtering methods on electrochemical signal quality and the accuracy of machine learning (ML) and deep learning (DL) models in classifying cadmium and lead heavy metals. This research evaluated seven filtering methods consisting of conventional categories (Savitzky-Golay, Butterworth Lowpass, Moving Median, DWT db4), adaptive (Kalman Filter, Adaptive DWT 6-Level), and decomposition (Adaptive VMD), along with an unfiltered condition as baseline. Five classification models were tested including Random Forest, XGBoost, SVM, KNN, and Deep Neural Network (DNN). Experiments were conducted on a dataset of 1,500 CV signal samples augmented with synthetic noise based on the Ag/AgCl electrode model at SNR 15 dB, resulting in 10,500 samples for evaluation. A total of 60 electrochemical features were extracted encompassing statistical, CV peak, frequency, and wavelet features. The results demonstrated that the combination of Kalman Filter with XGBoost achieved the highest accuracy of 97.52%. On average, Butterworth Lowpass Filter provided the best accuracy improvement (+2.13% from baseline), followed by Kalman Filter (+1.76%) and Moving Median (+1.48%). Conventional filters consistently outperformed advanced filters for CV signals. Conversely, Adaptive VMD actually decreased accuracy by 0.73% from baseline due to incompatible signal assumptions. In terms of models, XGBoost dominated with an average accuracy of 96.56% across 7 out of 8 filter configurations, while DNN showed the lowest overfitting gap (0.56%) thanks to effective regularization techniques. KNN exhibited the lowest performance (88.50%) due to the curse of dimensionality with high-dimensional features. This research contributes the first comprehensive ablation study evaluating 40 filter-model combinations for electrochemical heavy metal classification, as well as optimal pipeline recommendations based on application scenarios that can be directly implemented in machine learning-based heavy metal detection systems.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Cyclic Voltammetry, Filtering Sinyal, Machine Learning, Deep Learning, Klasifikasi Logam Berat, Kadmium, Timbal, XGBoost, Kalman Filter
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA402.3 Kalman filtering.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Alya Kamilah
Date Deposited:	30 Jan 2026 07:48
Last Modified:	30 Jan 2026 07:48
URI:	http://repository.its.ac.id/id/eprint/131322

Actions (login required)

View Item